Why the agentic era will break insurance systems

The ACORD 125 is on the left screen, the loss run is on the right, and the appetite guide is a PDF from March. That is a normal Tuesday, and the tools in it were chosen by people who were right.

A policy system is built to open one account and keep it straight: the terms, the endorsements, the premium. A warehouse answers what the book did last month, once someone has assembled the file. A bordereau is the file a carrier is owed, assembled when the period closes. Each one assumes a person who finishes a task before the next one starts.

Git earned the same kind of trust, for the same kind of reason. It assumes a developer who takes a complete copy, works offline, and shares the change when it is ready. Anthropic's engineering post on redesigning source control, published on 13 October 2026, says that is the right design for the vast majority of repositories. It is a large part of why git became the standard.

The received view in an MGA is the cousin of that one. Buy the system that is excellent at one job, connect it, and let people move between screens. It worked, because a person can only hold one account in their head. That stops when the accounts arrive faster than a person can open them.

Authority is the other half, and it is a different argument: what an agent is allowed to change, and where a person still has to see it. This piece is about what the agent can know when the question arrives.

Anthropic published what happened when the work outran the tool. In May 2026 the company merged about 2,500 code changes a day into its monorepo. By September it was more than 30,000. Writing the code had stopped being the slow part. The bottleneck was source control, the builds, and the checks that run before anything lands.

They were using git the way most companies do, as one authoritative copy. Nearly all of the clients were agents and jobs, and those clients do not wait their turn. A tool that assumes one developer finishes a change before the next one starts cannot absorb a day that size.

Git assumes a client that can work offline with a complete copy, and Anthropic's clients were online by definition. That is what broke for them.

They built Scion, an in-house version control system, and moved the monorepo onto it in five months. The detail that travels is not how the files are stored. A job can ask a question about the repository without keeping a local copy of it. The server answers. The client fetches what that question needs.

Anthropic is plain about the boundary. Scion is not a suitable replacement for git in general. The ideas make sense for an enormous repository, or for swarms of agents writing code, and Anthropic spends more on this internally than another business would choose to. Do not rebuild git. Notice which of your own tools still assume yesterday's volume.

Insurance is heading for that kind of day for a different reason. More product lines means more submissions that do not look like the ones last year's bordereau was built around. A broker's agent does not send one habitational account on the Florida coast to one underwriter and wait. It submits the same risk to more carriers, and it does it again before lunch. The MGA's own desk, and the carrier's, pick up agents of their own.

The result is more business on the same book: more submissions, and more operations, in the same day.

A platform that still answers from a report assembled after the day is over decides after the submission has moved on. Appetite, the share of insured value already in that county, the last three terms, what you charged last year, whether the quarter's bordereau has closed. Those used to be errands a person could run if someone stayed on the line. The broker's agent will not stay, and there will be more of them. The platform that can score the risk, see the concentration and answer while the submission is still open takes the account. The one that exports a file and pivots it loses it to a competitor who answered first.

An agentic operation is the work, with an agent in it: intake, the decline, the referral, the renewal call, the number you carry into a carrier meeting. Two properties decide whether that loop can close while the person is still in it.

The first is the book as it stands now, including the parts nobody has opened. Coastal concentration. The open tasks on the policies in front of you. Premium that moved when a policy bound an hour ago. A mid-market book is not a small attachment, and an agent that can only see the file on the email will answer about the wrong population.

The second is that any question is allowed. A scheduled report covers the questions someone already knew to ask. The submission that just landed asks the other kind. Largest building limits on the property renewals due in ninety days. Share of insured value in one postcode. Which of the ten submissions on the desk would push a bucket you have already filled.

If that question has to become a ticket, the agent is decoration. The person who needs the number has to be the one who asks, and the policies behind it have to come back with it. A figure you cannot open is a figure you cannot take into the room.

The fair objection is that the data is already there, and an agent can be pointed at the warehouse. Anthropic's caveat belongs beside it. They do not think every team should leave git. Most repositories are fine as they are. Most policy systems are fine at the thing they were bought for, which is opening one account and keeping it straight.

A warehouse fails a live submission for a simpler reason than git failed Anthropic. It is a copy that arrives after the fact. This afternoon's submission is not in yesterday's file. The policy that bound an hour ago may not be either. The agent will answer confidently about a book that has already moved, and the answer cannot be written back onto the risk the underwriter opens next, because the warehouse is not where that risk lives.

There is a second cost, and it shows up on Thursday rather than on the call. Appetite in a guide, a report in another file, and a question in a chat window are three books. The referral says one thing, the chart says another, and the sentence just produced says a third. Nobody is lying. They read different moments. An agent on top of that makes the disagreement faster. It does not make it one book.

The day that is coming has more submissions than a person can walk, and the platform that cannot decide inside that day loses them. We built Riska as the Agentic Operations Platform for Insurance so the record a person binds and the record an agent asks are the same book, at the same moment. The four surfaces below are how that shows up when a broker's agent has already moved on to the next carrier, and none of them is worth having if it reads a copy.

The Data Mesh is what ends the argument about which system is right. A figure lands once, so the quote, the referral and the month-end number move together. When an agent proposes a change, a person approves it before it lands, and the history names the conversation it came from. Thursday stops being a reconciliation.

Appetite keeps the hour on risks you might write. A submission is scored when it lands. A knockout or a referral sits on the risk, so the argument about habitational declines is a count you can open, not a feeling in the room. A bucket you have already filled can show up on the next quote that would fill it further, while the capacity decision is still a person's.

Reports & Metrics is the pack you owe people, still true this afternoon. The ninety-day renewal list is there on Monday because it is the book, not a file someone remembered to refresh. Written premium moves when the next policy binds. The bordereau a carrier expects comes from those same records, and a line that comes back outside tolerance is waiting for someone to look at it.

Ask Riska is the question nobody scheduled. The underwriter on the call asks the way they would ask a colleague, and the policies behind the number come back with it. Preparing the renewal is one question rather than five screens. A follow-up task can be opened before the call ends. It acts as the person asking, so giving it to the whole team does not invent a second set of access to govern.

An agent is only as current as the book it can ask. The book is only useful if the decline, the chart and the sentence agree. A platform that reconciles those three on Thursday will watch the submission bind somewhere else. That is the job the Agentic Operations Platform for Insurance is built for, and it is the job the previous generation of systems was never given.