Two things happened this week that belong in the same frame. On Chubb's third-quarter earnings call, Evan Greenberg said the industry has to "put our arms around" AI liability — a "growing and emerging exposure" that is "not contemplated in casualty underwriting to date" — and went further: he can imagine certification standards, "best practices in the use and management of AI and agentics," that would "assist in insurance underwriting." The day before, an essay by insurance-AI writer Srinivas argued the same thing from the architecture side. His line should be taped to every carrier CTO's door: "The next generation of insurance legacy systems might not be traditional applications. It might be disconnected ecosystems of intelligent agents."
Srinivas lays out the road in three stages: the workbench-led insurer, where a person coordinates and technology supports tasks; the agent-orchestrated insurer, where authorized agents coordinate retrieval, validation, enrichment, appetite checks, pricing, and risk-assessment prep while the underwriter keeps judgment, negotiation, and the binding decision; and connected market services, where broker, insurer, syndicate, and reinsurer agents exchange structured information and invoke authorized services across organizations. The catch is that every vendor stack is shipping agents with their own identities, permissions, data models, and audit rules. "Model portability is not the same as architectural portability." His proposed answer is a verifiable AI agent mandate — delegated authority for software, modeled on the binding-authority agreements and coverholder approvals the market already runs on. It names the entity represented, the permitted activities and limits, the human approvals required, the evidence of authorized execution, and how the mandate is suspended or revoked. The key sentence: "An agent should never gain underwriting authority simply because another system accepts its technical credentials."
An agent should never gain underwriting authority simply because another system accepts its technical credentials.
What it means for carriers
This is the authority question from three days ago, raised one level up. Essay Nº 003 was about the decisions inside one carrier. This is about the decisions between carriers, brokers, and markets — and the good news is that insurance does not need to invent the discipline. It is the industry's oldest one. Binder authority, coverholder approvals, referral limits: the market has governed delegated human authority for a century. The agent mandate is that playbook extended to software, and London is quietly laying the rails for it — Lloyd's March 2026 pivot from Blueprint Two as originally conceived toward process simplification and common data standards; the Core Data Record expanded on October 1 across pre-bind, placement, claims, and delegated authority including FNOL; ACORD's GRLC Generation 2.0; electronic placement above 90% for in-scope Lloyd's business. Meanwhile Moody's reported today that carriers keep raising AI budgets before seeing returns — 55% selectively measure returns and expect benefits within three years, while 81% cite talent, 71% data quality, and 62% legacy integration as the blockers. Money is flowing into an architecture that still has no authority model.
From the operator's chair
If I were running the book this week, I'd do three things.
First, stop buying workbenches; buy governed workflow. The verdict in Srinivas's piece matches what my own research concluded this month: the winning layer is vendor-neutral, governed workflow wrapped around the tools a carrier already owns — not another workbench with its own agent ecosystem inside it. Before signing any platform, ask the three-year portability question: what remains reusable, and what evidence of authorized execution survives, if you replace it?
Second, write the mandate before you deploy the agent. Treat software authorization like binder authority: named principals, scoped permissions, required human approvals, a ledger of what the agent actually did. The rule still applies — agents pre-qualify, a licensed person confirms, the ledger keeps the tape — and now it applies to the agents too. Greenberg's instinct is the market's instinct: certification and best-practice standards are coming. The carriers who wrote their own authority models early will set them; the carriers who didn't will rent them from someone else.
Third, distinguish automating the existing process from redesigning it. Automating submission ingestion inside each carrier separately rebuilds the same information N times — that is the work of the connected-market stage, not the workbench stage, and it belongs on common data standards like the Core Data Record rather than inside any single vendor's pipeline.
Access is not authority. A technical credential is not underwriting authority. The industry's oldest governance question is about to be its newest — and this time, the agent it governs doesn't have a desk.
— Krishna