AI Agents need the racetrack

Readingtime: 3 min

A year ago, we announced that we would only offer our AI solutions as AI agents. At the time, this was uncharted territory. Today, it is still far from standard practice, but it has become part of the verbal repertoire of discussions about AI in the insurance industry. What AI agents can do and why they are so relevant for insurers in particular is also now being discussed. However, surprisingly little has happened in practice so far.

A recent study by the MIT Center for Information Systems Research (Generative AI in Business: Early Experiences with Large Language Models, MIT CISR Research Briefing, Vol. XXIV, No. 7, July 2024) shows that a lack of implementation is not a problem specific to the insurance industry. Companies from a wide range of industries were surveyed for the study. The result: 95 percent of respondents are dissatisfied with their results to date in dealing with generative artificial intelligence (AI).

Interestingly, 80 percent have already experimented with large language models (LLMs), of which 50 percent have launched pilot projects and 40 percent have implemented LLMs. However, only 5 percent have used the models in a business-relevant, focused way. I suspect that these are precisely the 5 percent who are not dissatisfied.

This is actually the key: The semantic and argumentative abilities of LLMs may elicit admiration—and even amusement—but in the real world of business, they are of little use. A Formula 1 car is also exciting when it is parked on a dirt road, but it only unleashes its full potential when it is driven on the race track for which it was designed.

The same applies to AI in the insurance industry. Powerful AI needs the “race track.” And that is the workflow. The “car,” the AI agent, should be specifically tailored to the respective workflow.

Our experience proves this impressively: the use of general LLMs of the current generation achieves an accuracy of around 70 percent when applied to insurance matters. Specialized AI agents already achieve over 95 percent. And optimized, tailor-made AI agents that are deeply integrated into workflows achieve values of over 99 percent.

Conclusion: The difference between experimentation and business value lies not in the technology itself, but in how it is embedded in workflows. Those who merely test AI agents are driving a race car on a dirt road. Those who embed them in the race track—i.e., insurance workflows—take productivity to a new level, for example, by significantly improving loss and expense ratios.

Investors are also attaching great importance to this: Black Rock recently emphasized that the company is specifically looking for industries and companies that use AI in a concrete and productive way.

German Version: KI-Agenten brauchen die Rennstrecke

Read more: KI-Agenten brauchen die Rennstrecke

Fitting to the topic: