What Are AI Agents—And Who Profits From Them?

I work in an industry (integrity/enhanced due diligence) which is absolutely ripe for this type of AI agent takeover. I think the first company that manages to nail the "ai glue" element will basically kill off our entire industry (or at least 90% + of it) immediately. There are two types of main EDD/IDD work currently being done; the base, KYC/compliance work that is a regulatory requirement in a lot of jurisdictions and sectors, and the more reputational DD, which is usually more complex and is in support of some sort of transaction and investment, therefore going beyond just fulfilling a regulatory requirement.

LLMs are already adept at learning to mimic writing styles and at canvassing large amounts of information. Every firm in our sector keeps every past EDD/IDD report on file (for our small firm it's thousands and thousands of reports, a bigger firm like Kroll will have many more), so the model would have a huge amount of existing internal data to pick through, learning the house style, structure of reports etc. I don't think it will take too long before you have an AI that can go through large amounts of data online (google results, database search results either manually or through APIs, etc) and use its training on identifying relevant results, conducting "fuzzy searches", and analysing the results. Once you have that, having another AI agent which would write the report based off the existing internal repository of reports (most companies have a house style and structure of report) is also likely not very far away.

So basically "all you need" is that ai glue. To give you an idea, reports have at minimum a 2-3 working day TAT and the cheapest reports that have human analyst input cost thousands of dollars. An AI would write this in minutes, with commensurate savings in costs. Almost all EDD/IDD reports are a pure cost to the client and they are currently used to these TATs and costs... If you were the first firm on the market and your product actually worked, you would basically be unassailable on efficiency and cost (because the marginal gain of minutes, or a couple hundred dollars on an EDD report, is irrelevant to all but the absolute largest global bank's compliance team), and it would not be entirely clear whether the additional investment from competitors on making a tool that would provide marginally "better" (more detailed, even more accurate) reports would be worth it.

In fact, it would be interesting to see just how nuts the disruption around professional services firms would be. Already the largest corporate law firms are experimenting with AI to conduct doc review, which alone is like millions of billable hours a year.

Great article. Just like the term "AI", "AI agent" is overloaded and everyone seems to have a slightly different definition for it. Yours is as accurate as it can be as of today, in my opinion.

One comment on "Last year Anthropic told its investors it was preparing to create a model 10 times better than GPT-4". I can't recall exactly when they made this claim (about a year ago?) but am pretty sure they've already missed the mark on it. Claude 3 is arguably better, but not by much. We're seeing diminishing returns on "mega" LLMs (1T+ parameters) and scaling itself won't continue to yield proportional gains. Just as the auto industry started scaling down engine sizes in the 70s (while boosting efficiency), we now need to come up with smaller and "smarter" model architectures. That, or something akin of nuclear fusion 😬

Sorry for the naive question...are the agents you are referring to the same as GPTs? I don't hear much about GPTs since the initial announcement, so I am curious as I continue to learn AI and applications.