Where AI is actually working in underwriting (and where it's still theatre)

That was the central finding from our joint webinar with Celent, "State of AI in Underwriting 2026: What's Working, What's Not, and What's Next." This article recaps the session and pulls out the five takeaways we think matter most for underwriting leaders.

The webinar brought together Karlyn Carnahan, Head of Insurance for North America at Celent, with hx VP of Strategy Jamie Wilson and Chief Solutions Officer Tom Clark. Carnahan presented original research from Celent's third annual GenAI survey of North American P&C carriers. Wilson shared what hx is seeing across its customer base in commercial and specialty markets. The discussion was grounded in survey data, production use cases, and a direct assessment of where the industry actually stands versus where it thinks it stands.

Here are the five takeaways we think matter most. You can also watch the full session on demand.

1. Document ingestion is table stakes. The race has moved on.

Two years ago, automated document ingestion and summarisation would have been market-leading. In 2026, it is a hygiene factor.

Celent's survey data backs this up. The top underwriting AI use cases in production are document analysis and summarisation (29%), case analysis (22%), and submission ingestion (20%). These are all bottom-left quadrant activities on Carnahan's differentiation framework: necessary, sometimes high-impact on the expense ratio, but they do not change market position.

Source: Celent, 2026

Jamie Wilson reinforced this with a specific data point from hx: in the last 30 days alone, over 250,000 data points were automatically ingested into the hx platform across customers. That volume would have been entered manually just two years ago. One hx customer is rolling out ingestion across all 20 lines of business in three months.

Ingestion matters. But if you are still treating it as a differentiator, you are behind.

2. The productivity gap between top-quartile and median carriers is about to widen, and it won't close.

Carnahan framed this as one of three predictions she is willing to be wrong about: by 2028, top-quartile carriers will operate with materially fewer underwriting touchpoints per policy. The AI will be handling more of the workflow. And the gap between those carriers and everyone else will become structural.

The mechanism is not overnight disruption. AI will gradually reallocate margins and shift broker expectations. Underwriting productivity will change, and the carriers that redesign around AI will be rewarded for it. Then one day, the gap will be too large to close.

This maps to the tech adoption curve, but with a difference. AI's compounding returns mean that slow adopters don't just fall behind. They fall behind at an accelerating rate. As Tom Clark put it during the session: you may be talking to the Nokias and Blackberries of the insurance world if you haven't started.

3. Most carriers are stuck between Stage 1 and Stage 2 on the intelligence curve.

Carnahan's intelligence curve places carriers across four stages: Experimental, Operational AI, Scaled Intelligence, and Adaptive Enterprise. No carrier is fully operating at Stage 4 today. Very few are at Stage 3.

Source: Celent, 2026. No carrier is fully operating at Stage 4 today.

The uncomfortable reality is that most sit somewhere between Stage 1 (pilots, proof of concepts, unclear ROI) and Stage 2 (AI embedded in defined workflows with measurable productivity gains).

Celent's adoption data tells the story. In 2023, 8% of US P&C carriers had generative AI in production. By 2024, that figure was 28%. By early 2025, it reached 44%. That growth is real, but Carnahan's challenge to executives was direct: where are you actually, not where your innovation team tells you you are?

US P&C carriers with generative AI in production: 2023: 8% → 2024: 28% → 2025: 44%

Source: Celent, 3rd Annual GenAI-oneers in P&C Insurance

The board at Stage 1 considers AI interesting but not material. Stage 2 delivers efficiency gains, which is welcome but not strategic. The jump to Stage 3 is where AI becomes cross-functional, embedded at the point of decision, and supported by aligned data architecture. That is where boards start to see strategic capability, and it is the jump most carriers have not made.

4. The winners are embedding AI at the point of decision, not bolting it on as a separate tool.

This was the strongest point of agreement between Carnahan and Wilson. The use cases that deliver value are the ones that reach the underwriter's hands, inside their existing workflow.

Wilson shared an anecdote from a conversation in Chicago the week before: one insurance leader said the most effective generative AI deployment mechanism his team had found was Microsoft Teams. The reason had nothing to do with AI capability. Everyone is already logged in. That is a deployment insight, not a technology one.

Carnahan reinforced this: if AI intelligence sits in a separate tool, a second or third or fourth screen the underwriter has to open, the impact is limited. Getting the decision support to the point of decision, embedded in the process itself, is where the leverage sits.

This is the design principle behind hx's approach to underwriting agents. The Underwriting Agent and Ingestion Agent operate within the same workflow where underwriters already do their pricing and risk assessment. External data summarisation, policy wording comparison, and risk context support are surfaced where the underwriter needs them, not in a separate application.

Source: Celent, 2026

5. The biggest AI failures will be organisational, not technical.

Carnahan's second and third predictions both point to the same conclusion: AI governance failures, not technical failures, will be the first to make headlines. And the widening gap between top and median carriers will be driven by the difficulty of operating model redesign, not by the difficulty of the technology.

Carnahan's three predictions for 2026-2028:

1. The productivity gap will widen. By 2028, top-quartile carriers will operate with materially fewer underwriting touchpoints per policy.

2. The first major AI failure will be governance-related, not technical. AI governance roles like Chief AI Risk Officer and AI Product Owner will formalise.

3. The biggest AI failures will be organisational. The gap between top and median carriers will widen because operating model redesign is harder than technology adoption.

The technology works. The data needs work, certainly, but it is tractable. The hard part is cultural. Underwriters who have relied on their own judgement for decades are not going to embrace AI tools without clear evidence that the tools help them, not replace them. Change management and adoption are the binding constraints.

Carnahan used a memorable analogy to close the session. If you are making a cake, mixing the ingredients, putting it in the oven, that is exciting. But if you are the egg, all that happened is you got cracked and beaten. The job of leadership is to help the eggs see the value of being part of a cake.

It is a funny line, and a precise description of the change management challenge that will determine which carriers capture the value of AI and which do not.

Where this leaves the market

The webinar confirmed what senior underwriting leaders already sense: AI in insurance has moved past experimentation. The use cases delivering value are not the flashiest ones. They reach underwriters inside existing workflows and compound over time.

The separation between carriers is no longer about whether you are using AI. It is about whether you are redesigning your operating model around it, or layering it onto yesterday's infrastructure. If you removed your AI tools tomorrow and your operating model looked the same, the AI is sitting on top. It is not inside the business.

During the session, Jamie showed four practical applications of AI that hx customers are using today: external data summarisation to surface adverse media and third-party intelligence at the point of underwriting, policy wording comparison to track terms deterioration across renewals, risk context support to enrich location-level data within the pricing workflow, and quick quotes to accelerate turnaround on simpler submissions. Each one operates inside the underwriter's existing workflow, not as a separate tool.

If you want to see how these work in practice, get a demo scheduled with our team.