Eric Sibony is Shift's Chief Scientist and Chief Product Officer. In this installment of Four Questions we get Eric's take on Agentic AI, how insurers should be thinking the different types of AI that can be applied to their key processes, the ideal use cases for Agentic AI and how Shift is bringing Agentic AI to its products.
The easiest way to think about agentic AI is that it is artificial intelligence designed for complex automation. I like to use the following example to help illustrate what I mean. Airports around the world use trams without operators to take passengers from one terminal to another. This is automation, however it’s simple automation. Yes, the train stops at the right place and the doors open, but in reality it doesn’t take that much to make this happen reliably.
Agentic AI is more like what is happening with the development of autonomous vehicles. We’re no longer talking about going from point A to point B in a straight line or even a loop. Now we’re navigating city streets, avoiding other vehicles and pedestrians, stopping for red lights, and following all the other rules of the road. This is really complex stuff. Every step of the journey requires accurate analysis of data inputs—is that a stop sign or a yield sign—to make the right decision about how to proceed. The decision in this case: does the vehicle stop or slow down to assess its surroundings before continuing its journey? This level of complexity is what we’re asking agentic AI to address.
So now let me talk a little more specifically about how agentic AI is different from other forms of AI insurers may be familiar with. Machine learning (ML), which is sometimes referred to as predictive AI, is really good at taking data from multiple sources and providing a score—a suspicion of fraud for example. It can also report the factors that impacted how the score was generated. However, it’s not really great at knowing what to do with the score after it's presented to the end user.
Generative AI is all about analyzing data to produce something. For insurers, that something could be a summary of the documents included with a claim, a set of next steps for an SIU investigator to follow if a claim is found to be suspicious, or an email to a policyholder to provide an update on the claim they’ve filed. And while Gen AI is certainly capable of more complex automation than that of ML, the tasks it can automate are still relatively simple with their basis in data extraction and classification.
Agentic AI truly changes the game. Think of everything that goes into buying an airline ticket. You research convenient flight times, compare costs, and reserve and pay for seats. Most of us still do those steps manually, even when we’re using an aggregator site. With agentic AI you just ask, “can you buy me a flight ticket from Paris to Boston on January 1, 2026?” Granted you may need to put some additional parameters around the request, but fundamentally the AI agents are able to identify the flights that meet your requirements, secure the reservation and provide proper payment information. The AI behaves almost the same as if you had asked someone to buy a ticket for you.
Fundamentally, we end up with one big difference between an LLM (AI assistant, chatbot, co-pilot, etc.) vs. agent. With LLMs, you ask it something and you get an answer. With agents, you request something and it undertakes an action. I would say that's the easiest way to think of AI agents and what they are capable of.
Based on what agentic AI is capable of, insurers should be considering it as “AI for complex automation.” By that, I mean that it is able to reason, chain multi‑step actions, and act on both structured and unstructured data—both the insurer’s own and that from external sources.
Ideal use cases for agentic AI in insurance can be found throughout the claims process including initial triage, recommending best next steps, reacting to timing/urgency, facilitating subrogation pathways, and those tasks requiring making connections across documents and past outcomes. What it really comes down to is that agentic AI can be truly powerful when applied to tasks that require adaptation and judgment, not just a single prediction.
With that being said, agentic AI is less appropriate for simple deterministic workflows. Think about requesting a police report or undertaking pure scoring tasks where traditional ML is less expensive and more precise.
What is crucial is that insurers should be thinking about how to pair agentic models with domain-specific controls, an “I don’t know” fallback—meaning that the agent doesn’t simply make up an answer or required action, and human‑in‑the‑loop checks. This approach helps to prevent hallucinations and ensure reliability.
Shift views agentic AI as powerful automation that can reason, execute multi‑step actions, and bring together unstructured and external data. We also know that AI, especially for insurance, is not a “one size fits all” proposition. We need to make sure that we’re applying the right AI to the right insurance problems.
Our approach is designed to make it easy for insurers to adopt agentic AI: embed purpose‑built agents into Shift Claims so insurers don’t have to build them; pair generative/agentic models with traditional ML for scoring; and wrap everything in an insurance‑specific reliability layer that prevents hallucinations. That layer enforces checks, enables an explicit “I don’t know” fallback, and human‑in‑the‑loop exit ramps. We focus agents on complex, high‑variance claims and triage, not simple deterministic tasks, and continuously refine models from feedback to improve safety and accuracy.
Agentic AI in Shift Claims is well suited to claims transformation because it handles the complexity and variability of real-world claims. It can ingest and analyze unstructured data such as notes, documents, emails and even photos and images. Agentic AI can compare current claims to historical outcomes, and execute actions/instructions containing multiple steps, and then recommend the precise next course of action with associated timing and urgency. That makes it valuable for triage, case management and complex claims where tailored decisions drive material savings.
The result of agentic AI in Shift Claims: more effective, consistent decisions, faster throughput, and measurable loss reduction while keeping critical oversight in human hands.