The Missing Half: Why Insurers Struggle to Move AI Agents from Pilot to Production

Walk into almost any insurer today and you'll find an AI pilot. Probably several. A claims bot here, a document reader there, a promising underwriting assistant someone demoed at the last offsite. Ask how many of them are running real business, with real money and real policyholders on the line, and the room gets noticeably quieter.

That silence is the honest state of agentic AI in insurance right now. The technology works. Getting it out of the lab is where things stall.

Adoption is everywhere. Scale is not.

The numbers are blunt. More than half of insurers say they have adopted generative AI in some form, yet only around one in five have taken a single initiative beyond the pilot stage. Camunda's research shows the same pattern across industries: most organizations are experimenting with AI agents, but only a small fraction of those use cases reached production last year. Gartner goes further, expecting more than 40% of agentic AI projects to be scrapped by the end of 2027 as costs climb, value stays fuzzy and risk controls fail to keep pace.

Notice what's missing from that list. Nobody is saying the models weren't smart enough. The journey from pilot to production isn't blocked by an intelligence gap. It's blocked by an operating model gap. 

Two capabilities, usually bought separately

Here's what we believe many carriers are missing. Agentic automation and orchestration are two different jobs, and most organizations are investing heavily in only one of them.

Automation is the work itself. An agent reads a repair estimate, pulls out the line items, checks them against the policy and drafts a recommendation. It's fast, tireless and increasingly accurate.

Orchestration is the operating discipline around that work. It decides which agent acts next, which decisions need a human signature, what threshold triggers a review, and what happens when something looks off. It's less glamorous. It's also the reason a regulator will let you keep the automation running. 

Think of a busy airport. Every pilot may be highly skilled, but nobody boards a plane at an airport without air traffic control. Skill was never the issue. Sequencing, clearance and coordination are. AI agents work the same way. Ten capable agents without a shared control layer won't give you a faster claims operation. They'll give you ten new ways for things to go wrong at scale.

The reverse fails too. A beautifully designed control layer with nothing meaningful to control is just an expensive diagram. Insurers need both, designed together, on one platform.

The analyst community has landed in the same place. Gartner's BOAT framework folds process orchestration, integration, low-code and agentic automation into a single category. Forrester's view of adaptive process orchestration describes platforms that blend predictable, rules-based logic with agent-led decisions in pursuit of a shared business outcome. If your roadmap still treats these as separate purchases, the market has already moved on.

What it looks like on a real claim

Abstract arguments only go so far, so let's take an ordinary auto claim. No injuries, moderate damage, one other driver.

In a traditional setup, that claim spends most of its life waiting. It waits in a queue for an adjuster. It waits for photos and an estimate. It waits for coverage to be confirmed, a reserve to be set and payment to be approved. Each handoff adds a day, sometimes more, and a claim that needed perhaps an hour of real thinking drags on for a week or two.

With claims automation running on an orchestrated platform, the flow changes shape. One agent captures the loss and confirms coverage on arrival. Another reads the estimate and photos the moment they land. A rules engine sets a reserve within pre-agreed limits. If the claim sits inside the platform's authority, it moves to payment without anyone touching it, and every step is logged as it happens.

Claims that don't fit that profile are never forced through. A disputed liability, an injury, or photos that tell a different story from the estimate go straight to an adjuster. The difference is that the adjuster opens a file that's already assembled, with coverage checked and a reserve suggested, rather than starting from zero.

That's what straight-through processing really means in practice. It isn't about removing people. It's about sending people only the work that genuinely needs them.

The results are already visible. McKinsey has documented one carrier that put more than 80 AI models to work across its claims function. It cut complex liability assessment time by 23 days, lifted routing accuracy by 30% and reduced complaints by 65%. No single heroic model delivered that. Many narrow ones did, working inside a coordinated, governed flow. 

Underwriting follows the same pattern

Underwriting automation tells a similar story. Picture a small commercial submission arriving from a broker. Today it usually sits until an underwriter has a free moment, then crawls through rekeying, report ordering, more waiting, appetite checks and manual pricing. The clock runs mostly on idle time, not on judgment.

Orchestrated agents compress that dramatically. One agent cleans up the submission and checks eligibility. Another orders and reconciles the third-party reports. A pricing agent applies the rating logic. If the risk is within appetite and authority, a quote can go out in minutes. If it isn't, the underwriter receives a prepared file and puts their expertise where it earns its keep. McKinsey links this kind of operating-model change to double-digit gains in conversion and premium growth.

Why architecture matters more than the model

There's an appealing picture of agentic AI as one brilliant model handling a case end to end. Durable production systems aren't built that way, and I'd be wary of any vendor who suggests otherwise.

Good multi-agent orchestration looks more like a well-run team. Separate agents handle intake, interpretation, decisioning and escalation, while a governance function watches all of them. They communicate through events rather than calling each other directly. That design choice pays off over time. You can swap in a better document-reading model next year without disturbing your decision rules, your guardrails or anything downstream.

It also means every decision leaves a trail. When an examiner or a policyholder asks why something happened, you can show them.

Depth is the real differentiator

Almost every vendor now describes itself with some mix of "automation" and "orchestration," so the words alone tell you very little. What's genuinely hard to copy is insurance depth: FNOL triage, reserve logic, subrogation and regulatory reporting built into the platform from day one, not retrofitted onto a horizontal tool. That depth shows up where it counts, in how quickly you see value.

The question for your next steering committee

Running AI pilots no longer sets anyone apart. The insurers who pull ahead will be the ones who learn to run AI agents on live business, safely and at scale, before their competitors do.

So here's the question I'd put to any leadership team. Are you building automation and orchestration together, or planning to bolt on control later? In a business built on trust, "later" usually arrives too late.