Investing in AI Takes More Than Money

The way I would characterize it is that AI is highly capable but not yet competent. In order for AI to get over the hump of just good enough into the world of higher-quality and demonstrable impact, it really does require a mix of deep domain knowledge plus what the technology can do. At the end of the day, this is not a technology problem. It’s fundamentally a business process challenge and to some extent an organizational issue.

What resources are you applying? How are you actually treating it as a living, breathing organism, if you will, that needs to be taught, trained, nurtured, and managed? The expectation around what AI can deliver needs to be tempered with the reality of how it’s actually progressing against real world information processes and structures.

One, there is a whole lot of work that’s required to become really AI-ready. Not just from the technical side, but from the business side first. Where am I getting the data? Is it clean? Do I understand it? Is it verifiable? Those sorts of things. The data cleanup that you see in any large-scale insurance company or brokerage is very real.

The second is how does the process itself work? Am I going to apply AI into a highly fragmented, highly complex, essentially customized process or can I have some degree of standardization?

The cost of the setup, the cost of the preparation work, typically goes unappreciated, and people are fast-forwarding to the “And the magic happens here, where I apply AI” moment. When you deploy AI, you need to treat it as either a distributed workforce or even another employee. Employees need to be managed, they need to be trained, and they need to be improved over time.

The verification gap is the distance between what AI produces and what a human expert can stand behind. In insurance, it’s regulatory issues, it’s financial, it’s reputational risk, and it’s in a high-volume, high-risk scenario. What companies need to do when we talk about the verification gap is not just take output as structured. There are stories, of course, of all the hallucinations. But when a company is choosing to say “Yes, this is covered,” or “It’s not covered,” or “This coverage matches what we’d actually said,” the traceability is hugely important.

The provenance of how those decisions were made is even more important. Why would I choose this answer over that answer, or this result over that result, or trust this schedule of values over here versus something else that may have been produced—that’s the verification gap. While the machines are really competent in just showing you the discrepancies, the decision trace and the logic that sits behind that is where the verification really needs to live.

The value of what happens in agentic processing is that there is a decision trace and there is a data exhaust. The challenge is, how do you actually harness them and then learn from them? What you want to be able to do is look at decisions that are getting made, steps that are being taken, and then constantly reevaluate. Is that working? Is it not working? Can we improve it?

Some people view the human in the loop as the last checkpoint before something goes out the door. The expert in the loop is creating this reinforcing learning that says “We built the process to look like this. The AI system operated this way. My traces and logs and decisions showed that this is what we got. Here’s how we improve that now.” And that’s a big difference from simply running a process, sending it to a human, pressing a button, and it moves down the path. It’s much higher order and a key element that makes AI so potentially valuable.