Insurtech 2026 Verdict: Legacy Architecture Is the AI Bottleneck

Insurtech Insights 2026 at the Javits Center produced an unusually consistent message across two days and 400 speakers: the constraint on insurance AI is the data layer, not the model layer. Celent puts 48% of global insurers running generative AI in production. Conning's C-suite survey puts enterprise-wide scaling at 7%.

That 41-point gap is the number the conference was organized around, and nobody on the main stage attributed it to model capability.

Key Takeaways

• 80% of an adjuster's time goes to shuttling data between disconnected systems, per Snapsheet CEO Andy Cohen. The claims AI prize is recovering four hours in five, not a faster algorithm.
• 41 points separate production use from enterprise scale: 48% of insurers run GenAI in production, 7% have scaled it. Conning's survey also shows LLM adoption jumping from 18% to 63% in a single year.
• 28 operational domains were mapped in the Swiss Re, AXA XL and Stibo panel, which found the fastest movers deliberately started on lower-risk use cases to build the data governance track record first.
• The Insurtech Impact Award went to Quantexa, a connected-data platform rather than a model vendor, with MJP Insurance taking honorable mention for 84% efficiency gains against market averages.
• $100 billion in cumulative industry earnings over three years is the capital cushion funding this, and the panel named interest rates the single most impactful variable over the next 18 to 24 months.

The 80% That Frames Everything Else

The most-quoted figure of the conference came from the "Claims Without Friction" panel, where Snapsheet CEO Andy Cohen described adjusters spending 80% of their time as "switchboard operators", moving data between systems rather than exercising judgment.

Cohen's framing was that the goal is to make adjusters "superhuman," not to replace them. The operational point underneath is a cost one. Microsoft's February 2026 analysis found adjusters typically need one to three days simply to gather, read and interpret documents before substantive work begins. Against more than 30 million U.S. personal auto claims filed in a year, that lag is a loss adjustment expense line, not a workflow annoyance.

Sedgwick's Sidekick Agent, built with Microsoft, showed better than 30% improvement in claims processing efficiency. The qualifier attached to it every time it was cited: the improvement requires clean data moving through connected systems.

Conning's adoption numbers give the gap its shape. LLM use among U.S. insurers went from 18% to 63% in one year, with 90% of respondents somewhere in generative AI evaluation and 55% in early or full adoption, against 7% who have scaled. Our cross-carrier claims AI tracking shows the same split: the deployments that reached 75% cycle-time reductions consolidated data infrastructure first.

The Golden Record and What It Does to a Triangle

The analytical center of the programme was the underwriting track panel with Samrat Dua of Swiss Re, Suraj Tiwari of AXA XL and Mark Blake of Stibo Systems, which mapped AI capability across 28 operational domains.

Blake's term for the target state is the "Golden Record": one consolidated, validated version of each customer's data. His term for the current state is the "Frankenstein customer," the same policyholder living in five or six systems with conflicting addresses, coverage histories and claims records. A model querying that inherits the inconsistency and returns an answer that is correct on its inputs and wrong on the world.

The panel's finding was a sequencing one. Organizations that advanced fastest took moderate and lower-risk use cases first, building governance evidence before touching pricing, reserving or solvency work.

That sequencing has a measurable consequence for reserving. Carriers that finish data modernization get faster first notice of loss intake, cleaner claims records and automated routing, which compress early development factors and cut late adverse development from reopened claims. Carriers that have not, do not. Industry loss development triangles now blend both populations, so pooled factors carry heterogeneity that did not exist when every carrier ran comparable intake.

The award slate ratified the same thesis. The Insurtech Impact Award went to Quantexa, a connected data intelligence vendor rather than a model vendor, judged on business process clarity, measurable financial impact, uniqueness and breadth. MJP Insurance took honorable mention on 84% efficiency gains against market averages, which the committee noted required operational redesign alongside the technology.

McKinsey's April 2026 work put ranges on the payoff once the layer is fixed: 20% to 50% productivity improvement on discovery and reverse engineering of legacy policy logic, and 15% to 90% on data mapping, conversion and testing, with the marginal cost of agent reuse falling sharply once the first agents are governed.

What the Data Fix Does Not Reach

Two things constrain the thesis, and the conference surfaced both without resolving either.

The first is that the money funding this is cyclical. The "Funding the Future of Insurance" panel put roughly $100 billion of cumulative industry earnings over three years behind the current capital cushion, and named interest rates the single most impactful variable over the next 18 to 24 months. Elevated rates let investment income subsidize technology spend through a soft underwriting cycle. Falling rates put data modernization back in the discretionary expense queue, mid-build, at carriers whose payoff arrives two steps later.

The funding mix compounds the exposure. Global insurtech funding rose 19.5% in 2025 to $5.08 billion on a record 162 re/insurer private technology deals, and in Q1 2026 AI-focused companies took 95.2% of it, up from 77.9% in Q4 2025, at $1.55 billion across 68 deals. Capital that concentrated moves as a block when the rate environment turns.

The second constraint is that the fragmentation does not stop at the carrier boundary. Agentic AI is migrating into producer-facing tools, and independent agents work across dozens of carrier systems with their own formats, submission requirements and quoting interfaces. Cohen's 80% has a distribution analog in producers re-keying the same information into different portals. A tool sitting on fragmented carrier APIs inherits those APIs, so a carrier can complete its own Golden Record and leave the channel exactly where it was.

That is the gap Capgemini measured as a 42% non-measurement rate among P&C insurers. Insurtech Insights CEO Kristoffer Lundberg closed on the line that "the staircase has been built," which is accurate and also the point: a staircase is sequential, and most of the industry is still on the first step.