Rethinking applications for the AI age
Hint: it's not about chatbots or autonomous agents!
This week, I was shown a new CRM system called NetHunt. It’s pretty decent, and like most modern CRM tools, it integrates directly with your email and tracks your conversations with your business contacts.
That’s good, but it got me thinking about how much better we should be able to do this in the age of AI. A lot of people are talking about the value of large language models as autonomous agents that interact with humans or with other machines, but there is also a large opportunity to apply these models to optimize and replace existing business workflows.
The core of my hypothesis is that a large language model can now replace most data entry tasks with access to existing datasets. As a minimum, a large language model should be able to take a good first stab at how the data should look and then confirm specific details with a human.
Instead of having to set up a list of accounts, leads, and contacts in the CRM, it should be relatively easy to get a large language model to do this for me. First up, we give it access to each sales perso’s email. We can quickly see who they’ve been emailing. From the contents of the email, it should be straightforward to get any LLM to establish whether this contact is a lead, a customer, or a partner, and classify them accordingly. We should then be able to extract more context based on the email address and any email signatures to get enough information to run some online searches about the contact and their business.
Some people will be accessible on LinkedIn, for others we’ll only be able to find their company, but the LLM should be able to parse enough information from the website to set up a strong record for the details of the account. Then all we need is to highlight any opportunities that are being discussed and we have completed our CRM dataset.
Why is this radical? Firstly, people who are good at selling are notoriously bad at administration. Reducing drag on the sales cycle is going to boost revenues. Secondly, salespeople are notoriously bad at estimating because they are natural optimists. As my friend Paul Donato noted in his excellent paper on LLMs for the ARF, these models are much less biased because they don’t have preconceived notions. Getting an AI to predict the probability of a deal converting based on conversations and historical precedent will be much more reliable than asking an optimist whose bonus depends on her ability to convert deals.
When you realize that our models can not only access our emails but also our Zoom or Teams transcripts, you can see that the vision of a future CRM is not so far away.
The opportunity is the same if we apply the same logic to other applications. Consider the massive metadata catalogs that support the TV industry. Currently, descriptions are being manually written for every show and ad that appears on TV. Getting an AI to extract features from video and combine them with audio transcripts is relatively straightforward. You then just need an LLM to be able to write all of that metadata automatically.
How about campaign planning? At the moment, there is massive overhead in entering campaign details into multiple systems. Shouldn’t an AI be able to work out the target audience for the campaign by analyzing the creative and looking at the brand's online and social media presence? Moreover, the AI should be able to look at historical success patterns and use those to optimize the placements.
We need to consider every data entry and manual intervention in every process and ask if we can replace them with an LLM. We’re just scratching the surface of what’s possible with these new large language models, but I think the real opportunity isn’t so much in autonomous agents or chatbots as in designing brand-new applications that have LLMs built in from the ground up.
I’ve seen this firsthand with the work that Megan and the team at MX8 have been doing with consumer research. It’s almost like the investments made in technology over the past five years are likely to become entirely obsolete when AI-enabled competitors completely rethink the workflow.
This has to be the real start-up opportunity for 2024.