In an insurance brokerage, artificial intelligence first serves the document back-office. It reads inbound letters and PDFs, finds the matching policy, checks regulatory deadlines and routes the file to the right handler. The duty of advice stays with the broker.
Tandem works with more than 50 companies on these topics, insurance brokers among them. The projects that survive in production all look alike. They automate the standard file and escalate the exception, never the other way round.
What does AI actually handle in a brokerage?
Four families of tasks come back in every brokerage engagement. They share one trait. They are repetitive, high volume, and their output can be checked line by line.
- Reading inbound requests. Cancellations, quote requests, claim documents, certificates. The flow arrives as PDFs, scans and mail bodies, in formats that never match.
- Matching the policy. Finding the policyholder and the contract from a misspelled name or a partial reference.
- Checking deadlines. Confirming a request is admissible. French law lets a health policyholder leave at any time after one year, under article L113-15-2 of the insurance code.
- Drafting the first reply. Preparing a letter or mail already filled with the right references, which a handler reviews and sends.
The claims flow deserves its own mention. AI opens the file from the declaration and its attachments, checks consistency between documents, then routes it to the right track according to complexity. It also flags inconsistent files. It never decides the payout.
This split matches what Planète CSCA published for its members. The French brokers' union launched a programme called Vision IA in May 2026, covered by the trade press. It bundles a model AI charter, a guide of ten brokerage use cases built with consultancy Eürus, and training eligible under the insurance distribution directive. The selected cases cover market monitoring and quote comparison on motor and multi-risk policies. They reach commercial lease analysis, claim estimation and anti-money-laundering compliance.
How a workflow processes a cancellation letter
An inbound request travels through five steps. The fifth is the one that matters, because it decides who makes the call.
On the Santiane engagement, reading runs as two parallel passes. One tuned for scanned documents, one for native text files. Both streams merge before analysis. That redundancy costs a few cents and clears most unreadable scans.
Extraction then guesses nothing. Every field comes out with a confidence score, and uncertain values are flagged rather than filled with an assumption. The full build is public in the Santiane case study.
Why a workflow beats an autonomous agent on a regulated flow
I set out this distinction in a post on AI agents in October 2025. Almost every enterprise use case I see is a workflow. Very few are agents. An agent picks its own steps, which makes it non-deterministic and hard to test. On a regulated flow, that autonomy adds risk without adding value.
The consequence is concrete. Business rules and legal deadlines are applied by a deterministic module, never by the language model. The model reads and proposes. The code decides. That separation is what makes a file defensible under inspection.
What compliance forbids you to automate
In May 2026, the sanctions committee of the French prudential regulator issued a reprimand and a 20 million euro fine against Société Générale, acting as an insurance intermediary. The decision targets failures in pre-contractual information and in the duty of advice on a bundled offer sold since 2018. The bank argued that bundling exempted it from distributor obligations. That argument was dismissed.
The ruling matters to any firm tooling up its advisory work. The regulator's 2024-R-03 recommendation, in force since late December 2025, applies to every insurance distributor, brokers included. It extends the duty of advice and requires monitoring that a policy still fits the client throughout its life.
What automated processing costs per file
Cost is settled per document. On an inbound mail flow the main line item stays the extraction model calls, and it runs in cents.
| Line item | Approx. price | What moves it |
|---|---|---|
| Reading one document | A few cents | Page count, scan quality, dual pass |
| One complete file | Under ten cents | Cumulative extraction calls, API included |
| Workflow hosting | The firm's own server | Self-hosting forced by sensitive data |
| Implementation | Around eight weeks | Business rules to formalise, sovereign environment |
These benchmarks come from our engagements and from the Santiane case, where the target cost stays under ten cents per processed file, API included, for eight weeks of implementation. I detailed the general mechanics of automated extraction and data entry in my newsletter on use cases. The stack that comes back most often pairs n8n for orchestration, an optical character recognition engine for reading, and a small model for processing.
Health data changes the stack
A health insurance broker handles health data. That single constraint rules out most off-the-shelf tools, because they route documents through a third-party API hosted outside the European Union.
At Santiane the workflow therefore runs self-hosted, on the group's own infrastructure, and no health data ever leaves the controlled perimeter. The same reflex applies to automated quote processing, where volume demands keeping control of the execution environment.
Where AI gets a document wrong
A model reading a document does two things. It extracts text according to the format, then analyses images and tables, before merging both readings. Errors almost always hide in that merge.
In that walkthrough a receipt shows a fourteen euro total while a discount voucher brings the amount actually paid lower. The model grabs the most visible figure. Move the case to a certificate or a claim statement, and the error becomes a dispute.
The guardrail is one rule. A field that commits the firm never ships without a check. Either the confidence score clears a threshold, or the file goes to a handler as a draft. Firms that skip this step end up doing the work twice.
Where to start in a brokerage
Start with the highest volume, most standardised flow. In most firms that means cancellation requests or simple claim documents. Measure monthly volume and time per file first, before talking about tooling.
Three numbers are enough to decide. Files received per month, average handling time per file, and the share of files that fall outside the standard frame. That last one gives you the volume that will stay in human validation. A firm handling two hundred requests a month at fifteen minutes each frees up roughly three weeks of handler time a year.
Sector appetite is real. The 2035 outlook study published in November 2025 by Planète CSCA and EY reports that 75 percent of brokers have already committed AI-related investment. The union claims close to 2,400 members. The gap now plays out on execution.
Should you automate a brokerage back-office?
Yes, on the document flow, under one condition. Routing stays deterministic and human validation is kept on every file with something at stake. A firm that automates reading and policy matching wins back handler hours, without touching what carries its liability.
An autonomous agent on a regulated flow remains a poor choice in 2026. The framed workflow wins because it can be tested, traced and defended. To scope a first perimeter, our insurance engagements and our custom automation offer always start from the same place, real volume and a written rule. For the sector overview, AI in insurance takes the subject from the other end.
