The agentic insurer.
Where AI agents actually fit in core life and P&C systems — and where they don't. A field study drawn from eleven engagements across insurers worldwide, written by the engineers who shipped them.
The short version,
for those who need it.
Most agentic deployments in regulated insurance are over-claiming the value and under-instrumenting the risk. The honest map is narrower than the keynote slides suggest — and a great deal more interesting in the three workflows where agents do earn their keep.
This paper draws on eleven production engagements across insurers worldwide between 2024 and 2026. We interviewed eighteen contributors — engineers, underwriters, claims managers, and one regulator — and we instrumented two workflows for four months apiece. The findings are pragmatic, not promotional. They argue for staged adoption, in-the-loop discipline, and an evaluation practice the team will actually run.
The reference architecture proposed here treats agents as a layer on top of an institution's existing system of record — never as a replacement, rarely as an orchestrator. We will be wrong about parts of this within twelve months. The intention is to be wrong in ways that are inspectable and reversible.
“A good agent isn't the one that decides for you. It's the one that spares you the trivial decision — and leaves you the one that matters.”
Four numbers,
measured.
Drawn from instrumented production engagements. The full methodology sits in the appendix of the PDF.
Ten chapters,
twenty-eight pages.
Each chapter is anchored to a production case. Read in sequence, or jump.
The people
behind the byline.
Four named editors and a wider group of practice contributors and reviewers. Full list inside the PDF.
The agentic insurer: where AI actually fits in core life systems.
Reactit Research, Q3 2026. CC BY 4.0. https://reactit.tn/research/agentic-insurer