Then ENIAC arrived. It calculated a single trajectory in thirty seconds. Margaret and her colleagues took twelve hours.

Six women were selected to learn how to operate the machine. Margaret was one. Eleanor was not.

The official reason Eleanor gave was practical: her mother was ill and needed care. This was true. What was also true is that Margaret happened to be in a section where the officer coordinating the selection could observe her work. Eleanor was on a different shift. Margaret mentioned she was curious about machines. Eleanor never got the chance to be curious because nobody asked.

Margaret helped invent programming. She developed subroutines. She became foundational to an entire industry.

Eleanor remained a human computer for three more years before taking a job at an insurance company. She died in a state she’d never planned to move to.

This is the part they leave out of the “technology creates new jobs” story.

The space between macro truth and individual reality

I spent years researching what happened to skilled workers during technological disruptions. Nineteen stories, from ancient Egypt to 1990s Atlanta. And the part I kept finding, the part no one talks about honestly, is the middle.

The macro narrative is almost always true. Economies adapt. New industries emerge. Productivity goes up. If you zoom out far enough, everything works out.

But if you’re Margaret or Eleanor, you don’t live at the zoom-out level. You live at the level of which shift you’re on, whether the right person can see your work, and whether your mother is sick.

The stories that get told about technological transitions skip the middle because the middle is messy, contingent, and uncomfortable. It doesn’t reduce to advice. It doesn’t make a good TED talk.

What the middle actually looks like

Across nineteen cases, the middle follows a consistent pattern.

First, there’s a bifurcation. The market splits. People who want the new, cheaper version and people who still want the old, premium version. For a while, skilled workers retreat to the premium segment and do fine. Janet, the travel agent in Atlanta, watched Expedia launch and thought it was absurd. Her luxury clients didn’t care about Expedia. Her phone kept ringing.

Then the premium segment shrinks. Not because the cheaper version gets better (though it does), but because market expectations shift. What used to be “basic” becomes “good enough.” Lisa, the travel agent at the next desk, watched her walk-in traffic thin. Families started saying: “Let me check Expedia and get back to you.”

Then there’s a scramble. The workers who waited for the premium market to sustain them discover it can’t. Lisa tried to specialize in 2004. But the niches were already staffed by agents who’d pivoted five years earlier. The luxury market was not waiting for her.

Finally, a shakeout. Janet survived, but lost 500 clients and ended up with 150. That’s not a success story. That’s what survival looks like.

The timeline problem

The part that struck me most was the timeline. The disruption was visible for years before it became urgent. The signs were there. But the window for effective action was much shorter than the window of visibility.

Rachel, a junior accountant in Chicago, taught herself Lotus 1-2-3 at her kitchen table in April 1983 while her firm slept. She had about five years before everyone caught up. The accountants who learned Lotus in 1988 found the advantage had collapsed to zero.

You could see the change coming for a decade. You could act on it effectively for three to five years. After that, the good positions were taken.

What this means

I’m not going to pretend I know exactly how AI will reshape knowledge work. But I can tell you what the pattern looks like, because it’s been the same for 4,000 years.

The people who came through weren’t the ones who correctly predicted the future. They were the ones who moved before they had complete information. Margaret moved toward ENIAC while it was still unreliable. Rachel learned spreadsheets while her colleagues waited. Thomas the blacksmith started selling reaper parts before he fully understood the machine.

None of them had a strategy document. They had a bias toward action when the ground shifted.

And none of them came through unchanged. That’s the honest cost. Adaptation works. But it costs something. The people in my book who navigated displacement successfully still lost the professional identity they’d planned to have.

The stories don’t promise everyone comes through. They show you what the people who did come through actually did. That’s the best I can offer. It’s also more than most of the AI conversation is giving you right now.

This Has Happened Before*: US | UK

Anne Booth sits in her cottage on Chapel Lane, spinning wool into thread. Her left foot works the treadle in a rhythm her body knows better than her mind. Through the thin wall, she can hear Margaret’s wheel running, and Margaret hears hers. The merchant pays six pence a day. The weavers want more thread than they’ve ever wanted before. Anne has no reason to think anything will change.

Then news of the Spinning Jenny arrives. A man can spin eight threads at once. In Lancashire, spinners break into the inventor’s house and destroy his machines. In Calverley, the reaction is quieter: skepticism, then cautious calculation.

Here’s the part that keeps me up at night.

For thirty years, old and new systems ran in parallel. Merchants bought from both. Some spinners operated jennies. Others kept hand wheels for fine work. Competition intensified, but work remained. The demand for thread was so great that surely there was room for both.

But wages started falling. Six pence a day in 1765. Three pence by 1790. One and a half pence by 1810.

The parish records tell the rest of the story. By 1800, spinners were increasingly rare. By 1820, they’re nearly gone. New notations appear: “spinner and laundress,” suggesting spinning alone could no longer sustain a household. Then: “pauper,” “able-bodied poor,” “seeking work.”

The window felt open the entire time. The demand for thread never stopped. What stopped was the demand for thread at a price that justified paying a person to produce it.

I call this the Yorkshire Question: when your premium is rising because the cheap alternative isn’t good enough *yet*, how do you know when “yet” is about to expire?

The spinners who thought quality would always protect them spent the premium years feeling safe. The ones who used those years to reposition, learn new machines, move into quality assurance, and find adjacent work came through.

For anyone whose AI-generated competition is currently at 70% quality: that percentage is a moving target. The market’s definition of “good enough” is about to shift underneath you.

The question isn’t whether the window exists. It’s what you’re doing while it’s open.

This Has Happened Before: US | UK

He has worked this estate on the Nile’s west bank since he was twelve. He is past thirty now, which on a Nile estate in the early Middle Kingdom makes him one of the older carriers. His shoulders are uneven. His knees ache before midday. But he knows which part of the canal runs cleanest after the flood recedes. He knows the date palms on the upper terrace need water three days before the onions because their roots sit in drier soil. No one told him this. He learned it by watching what wilted and what didn’t.

Then one day, a stranger arrives with a wooden pole on a pivot, a stone lashed to one end, a bucket on the other. A shaduf. Pull the long end down, and the bucket dips into the canal. Release, and the counterweight lifts the water. A child could work it.

Within a season, the morning shift drops from twenty carriers to eight. The eight who remain are not the strongest or most experienced. They’re the ones who happen to work sections the shaduf can’t reach.

This is the opening story of This Has Happened Before, and I started with it for a reason. The pattern it establishes is the pattern that repeats across four thousand years.

The gap between the headline and the reality

The shaduf arrived in Egypt around 2000 BCE, probably from Mesopotamia. A single device could move five to ten times as much water as a carrier with a bucket. The productivity gain was not subtle.

But here’s the part the economic histories skip: the transition took two to three hundred years. Shadufs spread first through Upper Egypt, along the major canal systems. Manual carriers persisted for generations in smaller holdings and remote fields. The workforce didn’t vanish. It shrank. And the roles that replaced it were different: water-scheduling officials, canal-maintenance crews, readers who measured the Nile’s height for tax and planting decisions.

The people who mattered most were no longer the ones moving water. They were the ones deciding where it should go.

This is the gap between the headline and the reality. The headline says “technology replaces workers.” The reality says the value shifts, visibly and permanently, toward the people who coordinate, schedule, and decide how the tool is used.

What Kha knew but couldn’t name

Kha knew which terrace to water first every morning. He’d been making that decision for eighteen years. But Senbi, the estate overseer, never asked him why. The gap between making good decisions and having someone notice you make them is why most carriers disappeared.

That gap is the cruelest part of every displacement story I found. The knowledge doesn’t vanish. The visibility does.

If your work involves assembling information that an AI can now assemble faster, you are Kha. If you write first drafts that a model can produce in seconds, you are Kha. This doesn’t mean your job disappears tomorrow. It means the part of your job that felt most like yours is no longer the reason anyone needs you.

What remains valuable is the knowledge that lives in experience. The small adjustments made without thinking. The decisions no one writes down because no one has ever had to. The oversight that keeps the system functioning, not just operating, but thriving.

But having that knowledge isn’t enough. You have to make it visible. Write down the parts of your job that require timing or context that isn’t captured in any system your organization uses. Not your tasks. Your decisions. The things you do that nobody asked you to do, but that make the work better.

That is where your value is moving, whether you move with it or not.

*This Has Happened Before: What Four Thousand Years of Displacement Can Teach You About AI* is available on Amazon: US | UK

The optimists say: Don’t worry, we always adapt, technology creates more jobs than it destroys. They’re not wrong, exactly. But “the economy eventually adjusts” is not advice that pays your mortgage next year while your company decides whether to staff or automate your function.

The pessimists say: the end is coming, this time is different, you’re going to be obsolete. There’s real ground under that concern. But “you’re doomed” isn’t advice either.

What I couldn’t find was whether something was honest. Something that said, "Yes, displacement is real.” People lose jobs, markets shrink, and valuable skills become worthless. And also: some people navigate it. Some sectors contract instead of disappearing. Some companies choose differently. Some skills prove more durable than expected. And yes, there are things you can actually do about it, but you have to understand what’s really happening first.

So I went looking at the historical record.

If you’ve read this Substack before, you’ll know I’m fascinated by what happens when technology makes skilled work cheap. Not just modern history, but far enough back that the outcomes were settled, the myth-making had mostly stopped, and we can see the real patterns clearly.

So I’ve written a book about water carriers in Egypt. Spinners in Yorkshire. Travel agents in Atlanta. Craftspeople. Clerks. Skilled workers whose expertise was suddenly available for a fraction of the cost. What actually happened to them? Did they retrain? Did they get crushed? Did entire occupations vanish, or did they shrink and adapt?

I found 19 stories spanning 4,000 years. Egypt 2000 BC to Chicago 1983. From the Bronze Age collapse to the early personal computer era.

The patterns repeat with uncomfortable consistency.

What surprised me most wasn’t how often displacement happens. It’s that the outcomes follow predictable lines. Whether we’re talking about farmers with iron tools in 1200 BC or accountants with spreadsheet software in 1985, the same three or four things determine what happens to the people in the middle. And nobody talks about them. The popular narrative is either “everything changes” or “people adapt”. Both are useless. The actual story is much stranger and much more useful than either.

This book is not about AI. It’s about what happens when skilled work becomes cheap. That’s happened before. Many times. And we have a record of what works and what doesn’t, what gets people through and what leaves them stranded, what to do with a 3-year window, and what to do when the window closes without warning.

I wrote it for people like the ones I talk to every week at MX8. People doing real work that requires judgment, expertise, and time—and watching AI tools do parts of it faster. People are wondering whether they’re being paranoid or should take action. People who want straight talk, not reassurance, and not dread.

If you’ve been thinking about AI and your career the way most people have, as a future abstract problem, this book asks you to think about it the way historical people had to: as something happening right now, in real time, with real decisions to make today.

The core lesson isn’t comforting. But it’s actionable. And it’s based on what actually happened to actual people in situations very much like yours.

This Has Happened Before is available now. on Amazon.

If you read it and find it useful, I have one ask: leave a review on Amazon. The book is new, reviews matter more than you’d think at this stage, and I’d genuinely appreciate it.

Then, William Seward Burroughs produces a machine that adds with perfect consistency. It does not lose focus. It does not transpose sevens after lunch. It simply works. And thirty men who defined their professional worth by the steadiness of their hands are suddenly in trouble.

Market researchers should find this story uncomfortably familiar.

The Quiet Infrastructure

Most research teams still run on the same operational backbone that has existed for decades. Someone writes a questionnaire. Someone codes the data. Someone produces the tables. Someone reads the tables and turns them into a narrative. None of this is glamorous. All of it takes time. And nearly all of it is now automatable.

In recent posts, I have argued that AI is already capable of removing most of that operational drag. Not in theory. In practice. It can write surveys from briefs, summarise datasets fluently, code open-ended responses without supervision, generate synthetic data to explore hypotheses, and produce first drafts of reports that are clearer than many human ones. It can answer questions about a dataset conversationally, without an analyst as intermediary.

AI has become a machine for analysis in the way Burroughs built a machine for arithmetic.

And just like nineteenth-century clerks, researchers are responding with familiar arguments. It is not accurate enough. It misses nuance. Clients will not trust it. The human touch matters. All of these points contain some truth. None of them changes the direction of travel.

What Burroughs Actually Changed

The adding machine did not eliminate accounting. It eliminated the tedious part of accounting and forced the people doing it to move up the value chain. Clerks who embraced the machine became accountants. Clerks who clung to the ledger were laid off.

The same transformation is waiting for research. AI does not replace humans. It replaces tasks that were previously done only by humans because there was no alternative. And once those tasks disappear, the role itself begins to look different. The future is not one with fewer people. It is one in which people stop spending their energy on work that machines can now do more reliably.

Burroughs ushered in modern finance. AI is about to usher in modern insights.

The Fear Isn’t About Performance

Clerks feared the adding machine not because it failed but because it worked. It threatened their identity. When your value is defined by meticulous manual skill, a tool that performs that skill instantly feels like a personal insult.

Researchers are feeling a similar strain. If an AI can write a first draft of a report, that undermines the belief that report writing is craftsmanship. If it can summarise an entire dataset in seconds, that challenges the analyst's role. If it can ask probing follow-ups, that presses on the boundary of what counts as qualitative skill.

The fear is not about performance. It is about status. And confusing task automation with value automation is a mistake we keep making.

What Remains

Once the mechanical work disappears, what remains is the work that always mattered most. Thinking. Interpreting. Challenging assumptions. Understanding people. Applying context. Making strategic decisions. Telling stories that change minds.

These are not things AI takes away. They are things AI finally allows researchers to focus on.

Projects will move faster because they are no longer limited by human throughput. Budgets will stretch further because the operational layer becomes dramatically cheaper. Clients will receive deeper answers because analysts are not drowning in pre-analysis sludge. Creativity returns. Strategy returns. The job becomes more human as the workflow becomes less mechanical.

The Real Question

The adding machine did not replace clerks. It removed clerking. What emerged afterward was a profession with higher status and higher value.

AI will not remove research. It will remove researching-as-data-processing. And in doing so, it will let researchers become the strategic partners they have always claimed to want to be.

Burroughs’ clerks discovered that their future was bigger than their past once they stopped fighting the machine and started using it. The same opportunity is now sitting on every researcher’s desk.

The only question is who picks it up.

Not more noise. More signal. Not more metrics. More context. Not more measurements. Better timing.

Thirty years ago, companies relied on consumer research because it was all they had: structured, grounded input that helped teams make choices. Then the internet came along, and we turned to behavioural data. Fast, passive, cheap, and good enough. Clicks, opens, views, bounces. It wasn’t deep, but it was there when we needed it. It kept up with the clock speed of marketing and product clock speed. And market research became secondary to KPIs.

But a lot of real performance indicators can’t be measured behaviorally.

AI is shifting that equation again, and survey research is coming back.

Because now, surveys aren’t slow. Fieldwork doesn’t take weeks. Logic doesn’t have to be manually validated. Open ends don’t sit unstructured in backlogs. Synthetic models let us explore spaces that humans won’t tolerate, and validate only what matters. The result is a data source that’s just as fast as behavioural signals, but dramatically more structured, more interpretable, and more useful.

AI doesn’t just make surveys cheaper. It makes them better and faster than the business expects.

And that’s exactly why the bar is rising.

“Good enough” data used to be defensible. We worked with shallow metrics because we didn’t have time for anything else. The message didn’t get tested. The segment didn’t get updated. The brand signal didn’t get refreshed. But the product shipped, the campaign launched, and the presentation went out. We told ourselves it was fine, because it had to be.

Now the excuse is gone. And the risk is higher.

If your team is still working off last quarter’s segment, you’re behind. If your message testing consists of ten variants while your competitors are testing a hundred synthetically and validating the best ten with real respondents, you’re not just slower. You’re structurally wrong. If you’re watching brand shifts in quarterly snapshots while the market moves weekly, your “tracking” is already a post-mortem.

The cost of bad data is growing. Not because data is harder to get, but because decisions are being made faster than ever. Product cycles are shorter. Creative iterations are constant. CX loops are continuous. Sentiment is dynamic. If your inputs are stale or vague, your decisions will be too.

And the tools don’t protect you. AI lets teams move faster, but only the strong ones will move smarter.

Weak message testing surfaces late in performance. Weak segments collapse under activation pressure. Weak brand tracking misses the shift until it’s already in the market. And weak insight culture fails to tell the difference between a confident-looking chart and a decision-ready signal.

This isn’t about research anymore. It’s about operational relevance.

The teams that win won’t be the ones with the best decks. They’ll be the ones who bring the right data to the right decision, fast. AI gives every team access to more inputs. Only a few will build the judgment to act on them.

The bar is going up. Your competitors are already raising their game. Survey data is no longer a slow alternative. It’s the fast lane to structured evidence. The businesses that learn how to use it again will outperform the ones who settled for “good enough.”

Fast decisions don’t require less input. They require better input.

Now you have the tools to get it.

Don’t believe me? Listen to how Wunderkind totally changed their content marketing with AI:

What mattered, though, was not just the physics. It was the software.

I joined at a moment of transition. The team had spent years working in a world where “software” meant everything. They wrote the graphics drivers. They wrote the operating system. They wrote low-level libraries. They worried constantly about memory limits, hardware quirks, and processor behaviour. When a new microprocessor or graphics chip appeared, the cycle was familiar: acquire the hardware, get someone who understood compiler compilers to build a compiler for their variant of C or Fortran, recompile everything, debug relentlessly, and only then add new functionality.

It was an impressive system, and the people I learned from were deeply skilled. They could reason about the entire stack in a way that is now genuinely rare. But most of their effort was spent just getting the machine to work at all.

I arrived just as they were migrating to Windows.

From the outside, that sounds banal. From the inside, it was transformative. Overnight, huge parts of the problem space simply disappeared. No more writing graphics drivers. No more home-grown windowing systems. No more worrying about entire classes of low-level issues that Windows absorbed and standardised.

What replaced that work was something else: focus.

Instead of asking “how do we get pixels on the screen efficiently?”, the question became “what does the customer actually need?” Instead of building infrastructure, we could build applications.

One of the first projects I managed was a bespoke gunshot residue analysis system. If you’re unfamiliar with it, the idea is that you can match the chemical signature found on a fired bullet with residues found on a suspect, establishing whether that person fired the weapon. Previously, this kind of analysis was slow, cumbersome, and limited to specialist labs. On Windows, we created a streamlined, purpose-built application that fit directly into forensic workflows.

Today, versions of these systems are used in police departments in cities around the world. At the time, that reach would have been unimaginable.

What also struck me was a subtle cultural shift. Some of the people I worked with didn’t quite think I was a “proper” software engineer. I had only once debugged code with an oscilloscope. I wasn’t fluent in the lowest levels of the stack in the way they were. I lived from Windows upwards.

They worried about everything. I worried about outcomes.

In retrospect, this wasn’t a loss of rigor. It was a change in where rigor applied.

I think about that period a lot when people talk about AI and software today.

Right now, we still frame programming as a syntactic activity. We argue about languages, frameworks, tooling, and style. We train people to think in terms of exact instructions, rigid structures, and mechanical correctness. That matters today because humans are still responsible for translating intent into executable form.

But AI changes that boundary.

The next generation of people building software will not worry about syntax in the way we do. They won’t need to internalise every rule of a language or remember every API shape. What they will need to worry about is something else entirely: what the system does, whether it does it correctly, whether it is secure, whether it behaves well under edge cases, and whether it actually serves the user.

This is not a lowering of standards. It is a relocation of responsibility.

The Windows transition didn’t make software simpler in total; it made it possible to spend complexity where it mattered. As a result, applications became more bespoke, more usable, and more tightly aligned with real-world problems. Entire categories of software only became viable because the platform removed the need to reinvent everything beneath them.

AI is poised to do the same.

There is a common claim that AI will eliminate software development jobs. I’ve heard this argument before, in different forms. The historical record doesn’t support it. There were far more software developers in the 1990s than there were in the 1970s or 1980s, precisely because writing software stopped requiring you to also write an operating system and graphics stack.

When the barrier drops, demand expands.

If you can only build applications by mastering the entire machine, software remains a niche. If you can build on a powerful abstraction layer, software spreads everywhere. That is what Windows did, and that is what AI will do.

Yes, future developers will not know many of the things I know. In the same way, I never knew many of the things the engineers who trained me knew. They could see deeper into the machine. I could see further into the problem space. Each generation gives something up and gains something else.

Crucially, we did not make worse software. We made better software because it was more aligned with human needs.

My view is that AI is an enabler, not a destroyer. It will allow us to build more software for more specific contexts with higher expectations of usability and correctness. It will shift the centre of gravity in software engineering away from mechanical translation toward judgment, design, and responsibility.

I’ve seen this transition before. It didn’t end with fewer builders. It ended with software everywhere.

Not because the business doesn’t need that information. But because humans won’t tolerate the instrument required to produce it.

This is one of research’s open secrets: the most valuable questions are often the ones we can’t ask.

AI changes that. Carefully.

Synthetic models, when trained privately on your own historical survey data, open up a new category of research: the impossible survey. The kind that no human would ever complete, but that contains structure, logic, and signals worth exploring. Not invented data. Not simulated respondents from nowhere. But grounded extensions of your actual data, moving through spaces that would otherwise remain closed.

But it’s only useful if you treat it like a model, not a shortcut to truth.

Synthetic doesn’t create answers; it extends patterns. It doesn’t discover what’s true; it makes visible what your data already implies. That’s the line. And it matters.

Once you stay on the right side of that line, impossible surveys become not just feasible, but also necessary. They become essential.

Every researcher knows the human constraints. Real respondents fatigue quickly. They satisfice. They abandon long surveys. They flatten nuance under cognitive strain. They misread complex structures. They won’t tolerate sequential trade-offs or high-redundancy tasks. So we adapt. We drop ideas. We trim concepts. We simplify grids. We avoid the extremes. These are not methodological preferences. They are coping strategies. The data gets smaller because humans demand it.

Synthetic changes the constraint. When trained on a clean, consistent, well-structured dataset, a private synthetic model can handle what no respondent ever would. One hundred attributes. Two hundred concept permutations. Thousands of pairwise trade-offs. It doesn’t complain. It doesn’t drift. It doesn’t get tired or lazy. It just runs the space.

This isn’t attitude generation. It’s a pattern extension. Which only works if the patterns are present to begin with. If your past data contains real signal: how preferences cluster, how features cohere, how trade-offs shift across subgroups, the model can carry those patterns into new designs. You’re not asking it to invent sentiment. You’re asking it to map a structure. And when the structure is there, it performs.

This becomes particularly powerful at the edge. Most good strategy happens at the edges, not in the averages. But edge-case exploration is a nightmare for survey design: too niche, too complex, too repetitive, too weird. Which is why most of it gets dropped before the field. With synthetic, you don’t need to skip those steps. You can run the fringe, identify where patterns hold or collapse, and take the promising paths into live validation. You still need humans, but you only need them for the parts that matter.

It also helps resolve one of the most persistent tensions in insight: the backlog of ambiguous questions that stakeholders never stop asking. What happens if we push the idea further? What if we combine these features? Where does this concept start to break? How sensitive is this position to context, pricing, or sentiment? These are real questions. But most teams can’t afford to answer them: not in timeline, not in sample, not in budget. So we guess.

Impossible surveys let you stop guessing. You can stress-test ideas before running a tracker. You can model logical coherence before re-fielding. You can run boundary conditions, contradiction checks, or feasibility passes without spending a single cent on sample. And because it’s your data, not someone else’s model, you can trust the structure being extended.

This doesn’t just save money. It avoids bad research. It cuts waste, eliminates dead ends, and gives researchers room to work. The result is tighter fieldwork, fewer errors, better decisions, and far less noise.

None of this replaces real respondents. It just ensures they’re used where they’re needed. You don’t need a human to rank 200 attributes. But you absolutely need them to surface emotion, interpret nuance, and validate signals in the wild. Synthetic can’t do culture. It can’t capture variance. It can’t respond to the moment. That’s why the loop matters. Synthetic expands the space. Humans choose what’s important. The interplay is where the value lives.

But none of this works without one critical rule: synthetic models can only explore what you’ve already learned. You cannot generate truth from nothing. The training data has to be strong. The patterns have to be real. The respondent base must be large enough to capture variation. The model has to be private. And the guardrails have to be clear.

When that’s true, impossible surveys become one of the most powerful tools available to researchers. Not to predict the future. But to explore the shape of it.

I’m not proposing you stop running real surveys with real people. I’m just suggesting that you make sure you ask enough questions in those surveys that you can then run impossible surveys with synthetic respondents trained on exactly the responses you just got.

The future of research isn’t just faster. It’s deeper. Traditional surveys reveal what people are willing to tell us. Impossible surveys reveal what people can’t: because the instrument was never tolerable to begin with.

And now, for the first time, we can run them. Not to replace the human layer. But to make it sharper, faster, and vastly more focused. The best questions were never unaskable. They were just impossible to answer. Until now.

It worked, for a while. Annual cycles gave the illusion of rhythm. They fit neatly with fiscal planning and budget approvals. But they weren’t designed around customers or decisions; they were designed around operational constraints. And now the constraint is gone.

AI has collapsed the lifecycle. The steps that used to define insight production - survey build, programming, QA, fieldwork, cleaning, coding, reporting, analysis - are increasingly compressed or automated. What took six to ten weeks now takes hours, sometimes less. Not always, and not perfectly, but reliably enough to change how we think about pace.

And when the timeline shifts, the model sitting on top of it starts to break. If a clean dataset can be ready tomorrow, why are we planning to deliver insights six months out? If brand perception can move in two weeks, why are we still measuring it once a year? If customers respond to creative within days, what are we doing waiting twelve months to review it?

The gap between the speed of behaviour and the speed of insight used to be excusable; it no longer is. Most teams already feel that lag; they just haven’t had the tools to do much about it. Now they do.

What comes next isn’t a faster version of the same system. It’s a different one. You don’t get an accelerated tracker; you get a living brand signal. You don’t re-run a segmentation; you let the segments shift as behaviour changes. You don’t pre-load a shortlist and hope it holds; you test ideas as they’re formed. Insight stops being a milestone and becomes an input. It moves out of slide decks and into infrastructure.

And no, this doesn’t make researchers obsolete. Quite the opposite. Automation reduces manual lift; it does not reduce the need to interpret what comes back. If anything, the constraint just moves. Instead of being limited by the number of studies you can fund or field, you’re limited by how well you can read what the data is telling you. Insight becomes abundant; the bottleneck is judgment.

Legacy teams were designed around scarcity. One brand wave, one customer deep dive, one planning input. Not because it was ideal, but because it was the only thing that could be done. That scarcity defined the headcount, the timelines, and the relationship to the business. When that scarcity disappears, everything else starts to shift. The questions multiply. So does the need for people who can work upstream from the data.

What’s dying isn’t research. It’s the architecture that made it annual. It made sense in a world where data arrived slowly, where cleaning and coding ate up calendar time, and where budgets were locked to quarterly meetings. That world is gone. The workflows built to survive inside it are going with it.

No one is going to ask when your next brand tracker is due. They’re going to ask why you didn’t see the shift already. And if your calendar doesn’t let you answer that, it’s not a research cadence. It’s a risk.

Annual research had a good run. But it was built for a slower market, a slower process, and a slower response time. None of those conditions applies anymore. The future isn’t annual, and it isn’t even quarterly. It’s continuous, and the teams that learn how to work that way will be the ones that last.

The result wasn’t marginal. Yarn production soared; costs collapsed. The bottleneck in textile manufacturing—hand-spun thread—suddenly cleared.

Before the Jenny, weaving teams had to wait. After it, weaving accelerated to keep up. The industry didn’t contract. It scaled.

But the technology wasn’t the most important part of the story. The reaction was.

Fear always arrives first

Hargreaves’ invention didn’t land softly. Four years after it appeared, rioters broke into his home and destroyed his machines. They believed, correctly, that the economics of their work had changed. They just couldn’t yet see what would replace it.

He fled to Nottingham. The machines spread anyway.

Decades later, the Luddites would do the same. New productivity meant new fear. That part doesn’t change.

Today, it’s AI. The conversation is familiar. People working in knowledge roles—researchers, strategists, analysts—see automation showing up inside their workflows. The instinct is protective. But the long-term pattern rarely follows the fear.

Yarn got cheaper. Cloth got cheaper. Demand increased. The textile workforce didn’t vanish. It grew, in size and in variety.

Automation removed the bottleneck. Then the industry expanded into everything that had been waiting behind it.

Research has been spinning by hand

For the past twenty years, insight work has followed the structure of its tools. Surveys had to be scheduled. Cleaning was manual. Charting was slow. Each step dictated the rhythm of the next.

We built long cycles and inflexible plans not because those were optimal, but because the process required them. We ran annual trackers because real-time measurement wasn’t feasible. We tested small concept lists because testing more took too much effort. We published reports for decision-makers because real integration into decisions wasn’t yet practical.

Most research wasn’t designed around the needs of the business. It was designed around the constraints of delivery.

Now, those constraints are breaking.

Insight production is collapsing in cost and time

Surveys can be created and deployed in hours. Data cleaning and coding can be instant. Charting and summary analysis are already there.

What used to take weeks can happen by the end of the day.

And when the constraint lifts, behaviour shifts. Research teams start testing more ideas. Supporting more hypotheses. Asking better questions. Moving in closer step with the business, not just reporting on it.

It’s not that curiosity was ever lacking. It’s that the process couldn’t keep up.

Now it can.

What grows is not the machine, but the demand

The yarn didn’t matter. The weaving did.

In research, it isn’t the automation that carries the value. It’s what becomes possible once the friction disappears.

Insight teams can stop explaining their timelines. They can stop defending their headcount. They can stop positioning themselves as an expensive, slow-moving gatekeeper.

Instead, they can run at the speed of the questions. Not all of them. But more of them. Enough to matter.

As research gets faster and cheaper, its reach inside the organisation expands. So does the need for people who know how to handle it.

Not just to execute. To interpret.

And that’s where the job changes.

We won’t need fewer researchers; we’ll need better ones

Automation is already handling the mechanical parts: formatting, scripting, filtering, tagging. It will continue to get better.

But that doesn’t remove the need for humans. It clarifies where they matter most.

We will need researchers who can judge what’s useful, not just what’s true. Who can synthesise contradictory signals. Who can design better questions. Who can move fluidly between quant, qual, and behavioural. Who can bring context into the conversation, not just data.

AI can do the repetition. It can’t do the thinking.

That doesn’t eliminate roles. It changes what those roles are for.

The real risk isn’t automation. It’s inertia.

The organisations most exposed aren’t the ones adopting AI. It’s the ones that fail to change their research architecture in response.

If your team is still working on cycles shaped by what used to be hard, you are already lagging behind.

The gap will not be between companies that use AI and companies that don’t. It will be between companies that changed their internal model to match, and companies that left their workflows intact.

Just as spinning didn’t survive without the Jenny, insight teams won’t survive if they stay built around manual production timelines.

There is no structural advantage to being slow.

This is the expansion moment

AI will change the way research operates. But not by making it smaller.

It will make it too central to ignore.

The real shift isn’t technological. It’s architectural. A cheaper, faster, more flexible insight engine invites a different kind of business culture. One that asks more, listens more, and adjusts more quickly.

Research becomes not just a checkpoint, but a habit.

The Spinning Jenny didn’t kill the textile workforce. It made it larger, more varied, and more economically powerful.

This will do the same.

The only open question is whether your team will grow with it.

We start with Awareness. You show up in someone’s feed, interrupt their podcast, or plaster your logo on a taxi. You’re not solving a problem yet, you’re just getting on the radar. Think reach campaigns, outdoor media, and sponsorships. You measure it with aided recall or CPMs and hope someone remembers your name later. Most don’t.

Now you’ve earned their Interest. Maybe they click. Maybe they linger. Product detail pages, explainer videos, TikToks with unusually long watch times. This is where curiosity starts to crack through. Metrics? Time on site, scroll depth, click-through rate. The goal isn’t to convert. It’s to intrigue.

The shortlist moment. They’re not just aware, you’re under consideration. Your retargeting kicks in, your comparison charts, your user testimonials. Maybe a lead magnet lands in their inbox. You track engagement, bounce rates, and email opens. You start to matter.

Next, the intent is getting serious. They’re hunting for pricing. They’ve added to the cart. They’re on your “Book a demo” page. Your CRM is watching, and your conversion pixels are twitching. You fire promos, urgency copy, one-on-one chats. Attribution models light up. It’s close.

Purchase! The glorious transaction. The Stripe webhook sings the thank-you page loads, and a new row hits your sales dashboard. This is where performance marketing claims victory, but only because the other stages did their job. Measured by revenue, ROAS, and caffeine highs.

Advocacy is often forgotten, but powerful. The customer becomes the channel. Referrals, reviews, unprompted TikTok praise. This is where NPS lives, and so does lifetime value. If awareness is the start of the funnel, advocacy is where it loops back, turning one buyer into your next campaign.

The funnel was never just a diagram. It’s a system of attention and trust. And for a long time, it worked.

But what happens when the buyer isn’t a person anymore?

How the Funnel Breaks Down in the Age of Agents

Let’s walk through a near-future (or if you’re like me, a present-day) purchasing journey. I need a new set of noise-cancelling headphones. Instead of hopping between websites or wading through sponsored posts, I write a short brief to my AI assistant:

“Find the best over-ear noise-cancelling headphones under $300. Prioritize comfort, battery life, and sound quality. I’ll use them mostly for work and travel.”

The AI does the rest. It scans product specs, review sites, Reddit threads, and maybe even YouTube transcripts. It summarizes trade-offs, flags user complaints, and ranks the top three options with citations.

This isn’t speculative. I already use ChatGPT this way for B2B research: evaluating vendors, exploring frameworks, and even drafting RFPs. The leap from enterprise to high-consideration consumer decisions (think phones, TVs, vacations) is minimal.

Now step back and look at what just happened.

The AI did almost the entire funnel.

• Awareness? Not needed—it’s crawling the universe of options anyway.

• Interest? Sorted—products are grouped and scored based on my use case.

• Consideration? It’s in full swing—features, reviews, and trade-offs synthesized.

• Intent? It presents a shortlist.

• Action? I click one.

Except here’s the thing: I still make the final decision. And that’s where the funnel, or what’s left of it, collapses.

In the Moment of Truth, Awareness Resurfaces

Because now I’m scanning the AI’s top picks. And my reaction?

“Oh, I’ve heard of Bose. I trust Bose.”

Click.

The funnel didn’t matter. The years of programmatic optimizations didn’t help. The display ads I ignored weren’t even in the room.

What tipped the decision?

Brand awareness. Familiarity. Trust. Recognition.

If I’ve heard of the brand, and the AI hasn’t flagged any red flags, I’ll go with it. If I haven’t? I pause. I Google. I try to figure out if this unknown brand is real, reputable, or just a hallucination. Suddenly, awareness isn’t just a soft metric—it’s the decisive edge.

AI Will Kill the Funnel. And Resurrect the Top.

AI agents won’t need convincing. They need inputs. But I still need trust. And I trust what I’ve already heard of.

This means awareness moves from optional to existential. In a world where agents compress or automate most of the funnel, awareness becomes the only piece left for the human to feel good about the final call.

Years of funnel optimization narrowed everything toward performance. But AI is going to swing the pendulum back. The question now isn’t how to retarget the buyer—but how to make sure you’re already in the room when their AI agent draws up the shortlist.

How do we drive awareness in that new room, where the agent is the gatekeeper, researcher, and recommender? That’s a post for another day. But the message is clear: the brand is coming back.

And this time, it’s algorithmically essential.

Let’s get real. AI doesn’t write good code. It writes plausible code. It guesses. Sometimes those guesses are decent, sometimes they’re nonsense. That’s not a bug; that’s what it was trained to do. So when you drop an LLM into your dev pipeline and expect a productivity revolution, what you get - if you’re doing it properly - is an experienced engineer checking every line. And if you’re doing it properly, that engineer is cross-referencing against documentation, test cases, edge conditions, and business context.

And when you do that? Yes, the productivity gains are there. The boring stuff flies by. Unit tests, boilerplate, translations, documentation snippets—AI’s pretty good at that. It’s like having a room full of junior engineers ready to churn out first drafts 24/7, and you’re just walking the floor checking their work.

Which sounds brilliant.

Until you realise what it really means: you’re walking the floor. You’re not building. You’re not creating. You’re supervising.

And here’s the rub: great engineers don’t get into engineering to supervise code. They do it because they love to build. The joy of creation, of staring down the blank file and making something out of nothing, is hard to beat. And when you replace that with a loop of “check, fix, confirm, repeat,” it grinds. Fast.

So yes, AI is like a room full of juniors, but let’s not pretend that senior engineers want to spend their lives cleaning up after interns.

That’s the heart of the challenge.

AI can be a powerful tool, but only if we use it to amplify the creative joy, not dull it. That means shifting the AI conversation away from “How do we automate more?” to “How do we get more satisfaction out of the time we spend?” It’s not just about productivity metrics; it’s about designing workflows that leave engineers feeling like they still own the magic.

We’re not quite there yet. Most AI tools are still built to optimize throughput, not experience. They’re productivity engines, not joy engines.

But they could be. And if we want to stop AI from sucking the joy out of creative work, that’s where the focus has to shift: not replacing the work, but elevating the humans doing it. Let the machines take the grunt work. Keep the spark for us.

Because otherwise, vibe coding becomes just that: a vibe. All flash, no feeling.

Let’s talk about why we create in the first place.

Professionally, most of my creativity is channeled into coding. Away from the office I make a lot of music and am working on a couple of novels, but that’s a separate story.

So let’s talk about how I create at work. When I started out I spent all my time coding in C. And I loved it. The challenge of getting things working and the satisfaction when they did. Then Visual Basic came out, and I dropped Borland C for UI work almost overnight. Not because VB was a “better” language—it wasn’t. C had the performance, power, and precision. But Visual Basic had something else: ease. Suddenly I wasn’t wrangling pointers just to get a button to align. I wasn’t hand-coding pixel offsets to center a label or calculating window positions in my head.

It was amazing. The UI just worked. Drag, drop, done. Want a form? Drop it on the canvas. Need an event handler? Double-click the button. You could think about what you wanted to build instead of how to juggle the plumbing. The low-level stuff was still there, and if you needed it, you went back to C. But for 90% of the job, Visual Basic removed the friction. It didn’t kill the joy of coding—it gave it room to breathe.

That’s what AI needs to become.

Right now, it’s like we’ve built an AI-powered Borland C++. It’s impressive, powerful, and vaguely terrifying. But it still makes you write your mallocs. It still hands you vaguely helpful guesses and expects you to clean up after it. And worse—it inserts itself into the part of the process that people enjoy the most: the creation.

Here’s the shift we need: stop thinking of AI as the junior engineer. Start thinking of it as the layout manager.

You don’t hire a junior to draft all your interfaces and ask you to debug them. You hire a tool to make it easier to lay things out so you can focus on how it feels, how it flows, and how it works. That’s what Visual Basic did: it took away the pain of UI layout so you could get to the logic. So the experience.

AI should do the same for creative work. Take the parts we all dread:

• The repetitive copy-pasting.

• The placeholder content generation.

• The fifty ways of rewording the same headline.

• The boilerplate tests.

• The draft emails.

• The asset resizing.

If we can offload that to AI—reliably—then the creative process starts feeling like creation again. Not just supervision.

But we’re not quite there yet. Most current tools still sit in the uncanny valley between helpful and hindrance. They need to be watched. Checked. Reviewed. They still carry just enough uncertainty that you can’t trust them to own a task from start to finish. So instead of freeing you up, they trap you in a loop of “Did it get this right?” And that loop is joyless.

We have to build tools that, like Visual Basic, just work—at least for the stuff that’s meant to be easy. That’s how we regain the joy. Not by pushing humans out of the loop, but by building tools so good they disappear.

So here’s the challenge: stop asking how we replace people with AI. Start asking how we make the work feel better with it.

When VB showed up, I didn’t feel threatened. I felt liberated. That’s the bar. Let’s build toward that.

I know, I know. Broadcast is supposed to be in decline. Linear is dead. Cord-cutting is the future. Digital killed the local star. But if we look back at why broadcast has been losing ground, the answer is surprisingly simple: it wasn’t the content or the reach. It was the targeting and the measurement.

Broadcast has been losing advertising dollars to digital for two reason: digital targeting and attribution are better. Marketers don’t love banners and skippable videos—what they love is knowing who saw their ad, what that person did next, and whether they were in the correct cohort. It’s not about content; it’s about connection.

But what if broadcast could do that, too?

That’s the promise of ATSC 3.0, or NextGen TV. It connects the TV to the internet. And once the TV is connected, everything changes.

At Run3TV, we’re already trialing retargeting with the households we have. And even at our current footprint, the results are compelling. But the real inflection point will come with scale.

If the FCC follows the NAB’s recommendation for a mandatory transition to ATSC 3.0 in 2028, then by 2029, we’ll have enough volume to support not just trials but a fully scaled, data-driven broadcast ecosystem. That means addressable advertising, real-time retargeting, and attribution that doesn’t stop at the screen.

None of this requires new business models. We’re not reinventing anything—we’re just unlocking capabilities that are already working in cable and CTV and bringing them to local broadcast. It’s the same playbook, just a bigger field.

So, what does that look like in numbers?

• By 2029, we project 14.3M RUN3 TV-enabled households, growing to 25M by 2032.

• That growth alone could drive $2.5B in incremental addressable linear revenue.

• Add $0.9B in new retargeting revenue.

• Plus $0.5B in data-driven linear ad sales.

That’s $3.9B in total incremental ad revenue—on a media supposedly in decline.

Even with conservative assumptions, we’re talking about 40%+ growth in broadcast advertising revenue. That’s not a rounding error. That’s a renaissance.

So yes, I will be talking about agentic AI at NAB. But I’ll also talk about why now is the time to rethink what we mean by “broadcast.” Because when the pipe gets smart, the whole platform changes.

I look forward to seeing you there.

In 2024 alone, CTV ad fraud is expected to cost the industry over $1 billion—and that’s just the fraud we know about. The real number? Probably much higher.

CTV is attractive to fraudsters for one simple reason: high CPMs. Unlike display ads, where fraudsters have to work hard for small margins, CTV ad rates can reach $20+ per thousand impressions. That means if you’re a scammer faking ad traffic, CTV is where the big money is.

Invalid Traffic (IVT) is skyrocketing – In 2024, invalid traffic on programmatic CTV ads in the U.S. was 22%. That’s higher than most other digital ad categories, making it a prime target for fraudsters.

Fraudsters use bot networks and device emulators to generate 650+ million fake ad requests per day. And that’s just from one known scheme. And now CTV fraud is outpacing detection – New fraud techniques (AI-driven bots, SSAI spoofing, MFA apps) are constantly evolving, making detection an endless game of cat and mouse.

How Does CTV Fraud Work?

CTV fraud isn’t just one thing—it’s a collection of tactics designed to trick advertisers into paying for ads that no real person ever sees.

First up, fraudsters create botnets that pretend to be real CTV devices. You can read extensively about how the Pareto botnet infected 1 million Android devices and spoofed 6,000+ CTV apps, generating 650M fake ad requests per day. These bots can fake being a Roku, Fire TV, or smart TV, generate millions of ad impressions per day, and trick advertisers into thinking they’re reaching real households

We also get fake CTV Apps. (It seems MFA isn’t just a problem for digital). Pixalate found 2% of all CTV apps in Q2 2024 were MFA apps, CTV apps exist solely to serve ads—no real content, no real users. These Made-For-Advertising (MFA) apps are eating up $58M in ad spend, all the while runing ads in the background (even when the app isn’t being used), spoofing more popular apps to trick advertisers and generating millions in waster spend.

The final fraud move is premium inventory spoofing: buy cheap traffic and spoof it to look like CTV. Fraudsters make mobile web traffic appear as premium CTV impressions, so advertisers pay 20x higher CPMs for garbage inventory. The DiCaprio scheme used fake apps and domains to impersonate Grindr’s CTV inventory and sell non-existent ad slots.

And don’t believe that server-side ad insertion is sage. Fraudsters exploit it to generate fake ad impressions by running ads on fake SSAI servers (ads “play” but no one sees them), they then spoof device IDs to make it look like real TVs are watching. The StreamScam scheme used fake SSAI servers to create millions of phony CTV ad requests daily, costing advertisers hundreds of millions.

Why Isn’t it Stopping?

Advertisers aren’t paying attention – Many assume CTV is as fraud-proof as traditional TV. It’s not.

Platforms lack transparency – Many ad exchanges don’t provide full transparency into where ads are running.

Verification isn’t universal – Some advertisers still don’t use fraud detection tools, leaving them exposed.

And of course, fraudsters innovate faster – Every time a scam is caught, a new one pops up. CTV ad fraud doesn’t have to steal your budget. The industry is making progress—ad verification tech is improving, platforms are cracking down on shady players, and more brands are waking up to the risks.

But if you’re running CTV ads, it’s on you to be vigilant. Without the right protections, there’s a good chance you’re paying for ads that no real person is watching.

Time to stop funding fraudsters and start demanding better.

Jevons’ Paradox states that making something more efficient doesn’t mean we use less; instead, we use more.

Jevons figured this out in 1865 for coal consumption in the UK. As steam engines became more fuel-efficient, people assumed coal demand would drop. But the opposite happened—coal became so efficient that it powered more industries, expanding the economy and driving demand through the roof. Efficiency didn’t lead to conservation. It led to acceleration.

I’ve been thinking about this since the DeepSeek announcements. Swap out “coal” for “AI,” and you get a pretty good idea of where we’re headed. There’s this persistent fear that AI will eliminate jobs and make human labor redundant. But if Jevons’ Paradox holds, AI won’t lead to a jobless dystopia. It’ll supercharge productivity, create new industries, and—counterintuitively—increase the total work.

And as we find more efficient ways to run these LLMs, we’ll run them more.

Let’s break it down. AI makes everything faster, cheaper, and more scalable. If we follow Jevons’ logic, that won’t mean fewer workers—it’ll mean:

1. More Growth, Not Less Work – More Inventory, More Experimentation, More Complexity

AI isn’t just cutting advertising costs—it’s expanding what’s possible. Programmatic bidding, dynamic creative optimization, and AI-driven media planning reduce inefficiencies, allowing brands to launch more campaigns, test more variations, and reach more audiences. But instead of shrinking ad teams, this explosion of possibilities means more strategists, data scientists, and creative specialists are needed to fine-tune messaging, optimize engagement, and keep up with the ever-growing complexity of AI-driven campaigns.

Take retail media networks—AI has made them scaleable, turning every major retailer into an ad platform. More efficiency hasn’t led to a slowdown; it’s triggered an arms race in commerce-driven advertising, forcing brands and agencies to invest in AI-powered tools to stay competitive.

2. New Problems to Solve – Entirely New Media Markets

The internet didn’t just digitize old media—it created new economies around content, influencers, and on-demand entertainment. AI is following the same pattern, spinning up entire sub-industries that didn’t exist a few years ago.

Take synthetic media. AI-generated content is turning creative production into a scalable, automated process, but that raises new challenges: How do we ensure brand safety in a world of deepfake influencers? How do advertisers navigate copyright and ownership when AI remixes assets on the fly? These aren’t problems we had five years ago, but they’re shaping entire business models now.

And then there’s contextual advertising. With privacy changes killing off third-party cookies, AI-powered content analysis fills the gap. The result? A new wave of startups specializing in AI-driven brand suitability, contextual targeting, and sentiment analysis—industries that wouldn’t exist without AI-driven ad disruption.

3. Work Will Look Different – Reshaping, Not Replacing Roles

AI isn’t eliminating media jobs—it’s evolving them. Just as the rise of digital advertising didn’t wipe out marketers but shifted them from print to programmatic, AI is transforming how advertising professionals work.

For example, media buyers aren’t disappearing—they’re becoming media strategists who oversee AI-driven bidding algorithms. Copywriters aren’t being replaced—they’re becoming AI editors, fine-tuning machine-generated messaging for different audience segments. Creative directors are leaning on generative AI to spin up hundreds of ad variants, but their role in storytelling and brand vision is more critical than ever.

Look at ad measurement. AI is automating lift studies, MMM, and attribution modeling at a scale that is impossible to manually perform. But instead of making researchers obsolete, it’s forcing them to upskill—moving from raw data crunching to higher-level strategic analysis and insight generation.

Jevons’ Paradox tells us AI won’t just shrink the labor market—it’ll transform it. But the big unknown is who will benefit from that transformation.

• For businesses, AI brings efficiency, but efficiency alone doesn’t guarantee shared prosperity.

• For workers, adaptability is the new job security. The ability to work with AI, not against it, will define career stability.

• For policymakers, the challenge isn’t stopping automation—ensuring AI lifts more boats than it sinks.

If history is any guide, AI won’t be the end of jobs—it’ll be a massive shift in what counts as work. Jevons’ Paradox tells us that efficiency gains don’t mean society slows down—they mean we speed up. AI will make businesses bigger, make some jobs disappear, and create new industries we haven’t even considered. (Who knew retail media was going to be a thing?)

The real risk is not AI itself but failing to adapt fast enough. The future isn’t about fighting automation—it’s about keeping up with what happens next.

They were curious—not about the usual quarterly earnings chatter, but about the broader picture. “Can they stay on top of this AI boom?” they asked.

I told them I thought they would, at least until one of two things happened:

• someone else figured out how to make silicon as good as theirs for training AI models

• Or someone cracked the code on training those models more efficiently.

Eight months later, it feels like we’ve hit that second milestone.

This is not bad news for AI, in fact it’s amazing news. The AI race has never really been about who can make the best models. It’s about who can put those models to work in smart, meaningful ways.

The winners in this space won’t be the ones building from scratch—they’re the ones finding creative ways to integrate AI into existing systems or developing tools that let others do the same. It reminds me of the dot-com boom, where companies like Interwoven and Mediasurface were the darlings of the early web. Their valuations soared on the promise of proprietary content management systems that were supposed to transform how businesses managed their online presence. These platforms were complex, expensive, and designed to lock customers into a single ecosystem. For a time, it worked. But as the dust settled, it became clear that the value wasn’t in owning the entire stack—it was in providing flexible, accessible tools.

Eventually, open-source alternatives emerged, offering modular, customizable solutions for free—or at a fraction of the cost. What these open-source tools lacked in initial polish, they made up for with adaptability. They became the foundation for an ecosystem of plugins and integrations that let businesses tailor solutions to their needs. Interwoven and Mediasurface? They were replaced, left behind as the industry shifted toward tools that prioritized accessibility over exclusivity. The companies that thrived were the ones that saw the value in enabling users, not controlling them.

In the same vein, as we get better at developing frontier models, the best open-source ones are always going to learn from the leaders. I don’t agree with Elon Musk much these days, but he’s right that OpenAI was better as a not-for-profit.

So who are the AI companies I think are doing the right thing?

• Zapier: The ultimate API factory. Not an AI company as such, but they have the plumbing to connect apps and automate workflows, putting AI in the center of this to find information, trigger actions, and keep things running without constant manual oversight is truly fabulous. I use it to handle most of my incoming newsletters.

• MX8 Labs: Full disclosure, I’m a shareholder, but they’ve created a survey research platform with APIs built specifically to work with large language models. This means the AI can do 90% of the survey programming and analysis, reducing effort by an order of magnitude compared to historical platforms.

• BOLT.ai: Instead of just offering coding tools, BOLT has gone the extra step of wrapping coding into an infrastructure that supports full application hosting. You can build with AI and seamlessly go from idea to execution, making it faster to iterate. If you’ve got a non-technical consumer proposition you want to build, this is where I’d go.

• GitHub Copilot: A tool that doesn’t try to replace developers but works alongside them, adding just enough intelligence to make coding faster and smoother. It’s not the smartest AI, but it’s tightly integrated with a toolset that pretty much everyone uses. It will be interesting to see whether the integrations of Gemini into Google Workspace and Copilot into MS Office have the same impact.

• Synthesia: Focussed purely on how to put human-like avatars for e-learning and marketing to work, they’ve turned AI into a tool for better communication—not just flashy tech for its own sake.

Each of these companies focuses on how AI can help in day-to-day tasks. Much of it is pretty unexciting but has a significant business impact.

Meanwhile, the press focuses on what can be done with AI instead of asking what should be done. Automatic screen control by AI? It’s futuristic and flashy, but what problem is it really solving? Platforms already run on APIs, and AI doesn’t need to “see” a screen to do its job. It’s a redundant layer that adds complexity without clear value.

In your AI journey, focus on how you can leverage AI to either do things more quickly or entirely eliminate processes that you previously needed to do. The winners won’t be the people with the highest-ranked large-language model. That will be some Chinese or Indian university, closely followed by a raft of other open-source models.

The winners will be about who can put AI to work. That’s much harder to measure, but that’s where the money will be found.

But what happens when those metrics are set by players with a vested interest in the outcome?

Digital Platforms Skew the Game in Their Favor

As an industry, we’ve let digital platforms define what success looks like. From click-through rates (CTR) to impressions to views, the metrics built into their systems aren’t just reflective of outcomes—they’re designed to favor their platforms. Consider how social media platforms tout video “views” counted after a mere three seconds, whether the user was engaged or simply scrolling past. Or how programmatic ad platforms boast high “reach” without accounting for ad clutter or how often ads appear in low-quality environments.

These metrics create a feedback loop in which investment flows toward platforms that “perform” according to their own rules. The irony is that these rules are designed to make platforms look good, not necessarily to deliver genuine value for brands.

MMM still proves out TV Effectiveness

Contrast this with television, where Marketing Mix Modeling (MMM) studies have repeatedly demonstrated its unparalleled ability to build brand awareness and reach.

For example, Nielsen's studies consistently find that TV generates a higher return on ad spend (ROAS) for brand-building campaigns than most digital platforms. Similarly, Thinkbox, the marketing body for commercial TV in the UK, has shown through MMM that TV accounts for 71% of total ad-generated profit over the long term despite representing just 54% of media spend.

The reason is apparent: TV delivers unmatched reach and impact. Ads appear in trusted, high-quality environments with engaged audiences. Unlike digital, where ads are often missed, skipped, or lost in the clutter, TV ads benefit from the shared cultural moments they create—whether it’s a prime-time show or a major sporting event.

But here’s the catch: television’s effectiveness isn’t always immediate. Its results play out over months and years, while digital platforms are optimized to deliver fast—but often superficial—wins.

Making 2025 the Year of Balanced Metrics

Unless we’re going to let TV die a slow and certain death , 2025 must be the year the industry reclaims its metrics. It’s time to move beyond metrics skewed by the digital tilt and focus on measures that reflect genuine impact across the entire funnel.

For instance:

• Top-of-Funnel Metrics: Let’s emphasize reach and brand awareness, but focus on how well platforms actually reach attentive audiences. Quality matters more than raw numbers.

• Mid-Funnel Metrics: Engagement metrics like video completion rates should account for the context in which ads are seen—separating genuine engagement from passive exposure.

• Bottom-of-Funnel Metrics: Sales lift, return on investment (ROI), and customer lifetime value (CLV) must be included to ensure we’re measuring conversions that matter, not just vanity metrics.

The auto industry provides a great example. Toyota’s long-term investment in TV-led brand campaigns helped it weather economic uncertainty, while competitors chasing short-term digital wins saw market share erode. Similarly, FMCG giant Procter & Gamble has publicly committed to a more balanced media mix, acknowledging that TV remains essential for driving mass awareness and trust.

New Year’s Resolutions

Marketers, advertisers, and business leaders: 2025 is our chance to hit reset. It’s time to look beyond the shiny metrics digital platforms have sold us and focus on outcomes that matter. Let’s combine the best of both worlds—TV’s proven effectiveness and digital’s measurable immediacy—into a more balanced and strategic approach.

The metrics we choose will define the future of our campaigns, our brands, and our industry. Let’s make them count.

Highlights for me:

• It started with a conversation with Justin Evans from the Data Story at CES about how we anthropomorphize AI and large language models. Later in the year, there was a lot more of this.

• The big news for me this year was William Wang finally getting his exit on Vizio to Walmart. I was cynical about whether the deal would close, but I was finally proven wrong in the year's closing weeks. Congratulations to everyone on both sides of that deal. I’ll be fascinated to see how Walmart leverages the fabulous data assets they’ve acquired.

• I’ve written several times about brand vs performance advertising, and I still think the industry is ripe for a resurgence in brand advertising, or maybe we’ll start talking about" “brandformace” as a new category. Either way, we’re too focused on only buying what we can measure.

• AI is the more significant trend, and the increases in speed, performance, and cost have opened up AI to the masses. I’ve created some creative videos and even copy-tested them to see if people can tell the difference between real and AI videos. Oh, and make sure you listen to the AI-generated podcast if you haven’t already. Just please be ethical about how you use AI.

• The market research industry is changing to follow suit. MX8 Labs, the AI research company I’m involved with, is growing nicely.automate manual processes seems to resonate. However, The proposition of using AI to I remain cynical about replacing consumer research with AI, and it also seems we’re not quite ready for voice to replace online surveys.

• Finally, I’ve been getting back involved in Broadcast TV through Run3TV, a company central to enabling smart broadcasting on the next generation of TVs currently rolling out across the USA. Most of the industry is so focused on streaming that they miss a big opportunity. What’s not to like about free football?

I’m now nesting away with the family for Christmas. Like 2024, 2025 will start at CES. I hope to see many of you there.

In the world of data science, this tension is known as the explore-exploit conundrum. It’s a mathematical problem that models decision-making under uncertainty, where one must balance the need to explore new opportunities (which may yield higher rewards) with the desire to exploit proven strategies (to maximize immediate returns)

The multi-armed bandit problem is a mathematical model for understanding the explore-exploit tradeoff. Imagine a gambler faced with a row of slot machines (or “bandits”), each with an unknown payout probability. The gambler aims to maximize their total reward over a series of pulls. To do this, they must decide how to allocate their attempts—exploring different machines to learn about their payout rates while exploiting the machine that offers the best return. The dilemma arises because every pull used for exploration is not spent exploiting a potentially better-known machine, creating a tradeoff between gathering information and reaping the rewards.

There are many solutions to this problem; the Upper Confidence Bound algorithm is an easy-to-understand one. This approach balances exploration and exploitation by prioritizing machines with high average payouts and high uncertainty about their performance. It calculates a confidence interval for each machine’s average reward and selects the machine with the highest upper bound of this interval. This approach ensures that lesser-tested machines are occasionally played, reducing the risk of prematurely committing to suboptimal choices. Over time, the algorithm converges on the optimal strategy by systematically narrowing the uncertainty, making it a powerful tool for decision-making in uncertain environments.

For marketers, this framework translates directly into strategic decisions. Should a company experiment with untested, potentially high-impact tactics like emerging platforms or innovative messaging? Or should it double down on proven channels that guarantee immediate conversions? The answer depends on the organization’s appetite for risk, the length of its planning horizon, and its ability to endure short-term uncertainty for long-term gain.

CMOs and CFOs often find themselves unwitting participants in a multi-armed bandit game as they allocate budgets across marketing strategies. Short-term tactics like pay-per-click advertising or promotional campaigns represent a simpler version of the problem. Here, feedback loops are quick—marketers can run A/B tests or analyze immediate ROI, exploring various strategies rapidly before committing to the most effective ones. However, the game is significantly more complex when considering long-term investments like brand-building, where the “arms” (strategies) may take years to reveal their actual value.

Research consistently highlights the profound impact of brand equity on consumer behavior and business outcomes. Strong brands drive price premiums, reduce customer acquisition costs, and enhance customer loyalty. Studies, such as those by Les Binet and Peter Field in their seminal work on the long-term effects of advertising, reveal that brand-building efforts create compounding returns over time, even as short-term sales activation delivers immediate but often fleeting results. The challenge is that these long-term strategies require exploration over extended periods, during which leaders must resist the pull of short-term exploitation.

The fundamental problem is that short-term marketing strategies have an inherent advantage in measurability. Metrics like click-through rates, conversions, and immediate ROI provide quick and tangible feedback, allowing marketers to assess the effectiveness of campaigns almost in real-time. This creates a natural bias toward short-term tactics, as their outcomes can be tracked, optimized, and reported confidently. In contrast, long-term strategies like brand-building operate on delayed timelines, making their impact harder to quantify. Metrics such as brand equity, consumer loyalty, and lifetime value are abstract and often require years of consistent investment to materialize. This disparity creates tension in resource allocation, with short-term approaches usually winning out in environments where quick wins are valued over enduring growth.

This focus on short-termism also seems to be playing out in the boardroom. CMOs in consumer-facing Fortune 500 companies face average tenures of approximately four years, with over 20% serving for a year or less. This limited time horizon amplifies the bias toward short-term strategies, as CMOs must deliver measurable results quickly to secure their position. Meanwhile, CFOs, who typically remain longer in their roles, naturally emphasize immediate performance metrics, further entrenching the preference for tactics that yield fast and quantifiable outcomes. This discrepancy in priorities often leaves CMOs between needing to drive short-term results and wanting to invest in brand-building strategies that may not pay off until well after their tenure.

The result is a self-reinforcing cycle. Many CMOs don’t stay in their roles long enough to see the benefits of brand investments, leading them to prioritize strategies that deliver results within their time horizon. This short-term focus may, in turn, contribute to higher turnover as CMOs face pressure to demonstrate immediate impact.

The question then becomes: do CMOs avoid brand-building because they won’t be around to reap the rewards, or is their failure to focus on the long term a reason for their early departures? Either way, the structural incentives in many organizations appear misaligned with the enduring benefits of a strong brand, raising critical questions about how leadership stability and strategic focus could better support long-term marketing success.

1. AI Should Enhance, Not Replace, Human-Centric Insights

The core purpose of market research is to uncover fresh consumer behaviors, attitudes, and reactions. AI, trained only on existing data, cannot discover the new; it can only enhance what already exists. Researchers must ensure that consumers remain at the heart of all insights.

2. AI Should Empower, Not Replace, Researchers

AI is a tool—not a substitute—for the skilled professionals who craft the narratives and insights that guide marketers. Researchers are responsible for ensuring that findings are accurate, ethical, and actionable. AI serves to accelerate and deepen their insights but must never take over this critical responsibility.

3. AI Will Eliminate Inequitable Shortcuts

Traditional constraints on time and budget have often forced researchers to make compromises, such as excluding underrepresented groups or foregoing necessary translations. AI will remove these barriers by making tasks like survey translation, inclusivity, and large-scale fielding both accessible and affordable.

4. AI Will Increase the Demand for Consumer Insights

The pace of business and societal change is accelerating, and AI will only heighten this velocity. With faster cycles of innovation, businesses will require more frequent and detailed consumer insights to adapt effectively and responsibly.

5. AI Will Amplify Budgets

Market research budgets will go further with AI. Tasks like manual programming, re-fielding surveys, and coding open-ended responses—once time-consuming and expensive—will be streamlined, allowing researchers to focus resources on strategic goals rather than operational constraints.

6. AI Will Enable Unprecedented Depth of Research

AI will transform the complexity of research into an opportunity, not a hurdle. Researchers will be able to adapt surveys in real time, test intricate hypotheses, and analyze data at scales previously unimaginable. AI will allow us to ask the hard questions and find the nuanced answers.

7. AI Will Extend the Reach of Market Research

Smaller companies and startups, historically excluded from robust market research due to cost or speed barriers, will gain access to the same depth and quality of insights as larger brands. AI will level the playing field, enabling businesses of all sizes to make data-driven decisions.

Commitment to Ethics and Progress

By adhering to these principles, we ensure that AI becomes a force for good in market research—one that amplifies human insight, promotes inclusivity, and empowers businesses to understand and serve their consumers better than ever before. Let us embrace AI as a partner in progress, not a replacement for the human creativity and curiosity that define our field.