I build AI for a living. Liberate's AI agents answer millions of insurance calls and handle claims without a human in the loop. I’ve seen AI systems operate inside real workflows, in real moments. I know what AI can do. I also know what AI should never do.
AI is extraordinary at the 95%
Most people think AI’s value comes from the conversation. But that’s just 5% of the real value. The other 95% is what happens behind the scenes. It’s verifying policy data, receiving first notices of loss, triggering repair workflows, and delivering Certificates of Insurance. That orchestration layer is where AI isn’t just a little better than humans, it’s fundamentally better. AI works faster, more accurately, and at a fraction of the cost. That's the fundamental reason AI belongs in the insurance industry.
Three things that need to stay human
#1 Decisions you can’t take back
Some decisions need to be owned by humans. In insurance, that means the decisions that carry regulatory weight, legal exposure, and direct impact on someone’s financial life. AI can still surface the data, flag the signal, and prepare the recommendation, but if a claim is wrongly denied and a family loses their home, someone has to answer for that. That someone should be a person. Accountability has to live somewhere, and it shouldn’t live in a model.
#2 When someone needs to feel heard
A policyholder calls at 3 am because their family home was just flooded. They’re scared, and they aren’t just looking for a claim number. They need to feel that one of the worst moments of their life actually matters to another human being.
AI can handle the intake. It can verify the policy, open the claim, dispatch the mitigation crew. What it can’t do, and what I don’t want it to do, is be the final word in a moment of genuine human distress.
The best AI deployments I’ve seen plan for human escalation from day one. Not as a fallback for when things go wrong. As a deliberate design choice for when the conversation moves somewhere technical efficiency can’t reach. Escalation should be a feature, not a failure.
#3 Strategic judgment under genuine uncertainty
AI is trained on the past. It’s extraordinarily good at recognizing patterns from data. But it’s not equipped to navigate situations where the right move requires synthesizing events that have never happened before, like entering a new market, responding to a regulatory shift nobody anticipated, or deciding which bets to make when climate change is creating weather we’ve never seen.
These decisions require lived judgment. The kind that comes from years of operating inside a specific industry, getting things wrong, and understanding how to improve.
I spent nearly four years at Metromile running insurance operations before co-founding Liberate. Operating inside a regulated, complex domain builds something models can’t replicate. You earn judgment. You can’t download it.
Why this matters more in insurance
Insurance is a promise. Policyholders pay premiums for years, and in return, we promise to be there when the worst thing happens. That promise is fundamentally human. AI can help keep it. But it can’t make the promise.
The carriers and agencies doing this well aren't replacing their people with AI. They're redirecting them. When routine work goes to AI, a claims adjuster who used to spend three hours a day answering claim status calls spend time accelerating claim resolution and improving NPS.
Drawing the right line
There’s a real risk in how the AI conversation is framed right now. On one side, there are people who say AI will do everything. On the other, there are people who say AI can’t be trusted with anything sensitive. They’re both wrong. The line isn’t AI versus human. It’s about what kind of work deserves what kind of attention.
Some work is repetitive, rule-bound and speed-dependent. Policy renewals, coverage lookups, status updates, FNOL intake. The outcome is the same regardless of who handles it. Give all of that to AI.
But a policyholder disputing a denial after a major loss isn't in a workflow. The outcome genuinely depends on how that insurance agent shows up at that moment. That work has to stay with people. Not to protect jobs, but because the alternative produces a worse result.
Carriers who get that distinction right will build something most of the industry hasn't achieved: lower costs and better customer experience, at the same time, without trading one for the other.
This is why it matters to me: Insurance has always wanted to be there for people at their worst moments. With AI, we’re finally able to make that possible at scale.
