Private-cloud agentic models for insurers and banks, priced per token
Every board cycle, a bank or insurer's Head of Automation is asked for an agentic model the CISO will let near claims files.
Shortlisted vendors print no price or deployment sheet, so the board declines the pilot. Cahier is the row that board can read.
Cahier is a model system shipped as one single-tenant container image. The image holds a fine-tuned open-weight language model with a 32k-token context window, an inference runtime that answers UiPath's LLM connector, and a policy gate. A governance console issues each version's evidence pack, mapped to the NIST AI Risk Management Framework.
Cahier's model reads the prose in claim letters, broker emails and KYC files, which varies with every sender.
Templates and regular expressions take header fields, reference numbers and dates; the model extracts the rest from free text and proposes the robot's next action, scoring each field and action. Per workflow, the evaluation report counts the held-out cases rules resolved alone, those the model resolved, and the measured error.
Every action Cahier's model proposes is checked against a per-workflow allow-list, and anything off it is dropped. Below a per-field confidence threshold the customer sets in the console, the case goes to its review queue in ServiceNow; no action that legally affects a person is executed at any confidence. Reviewer decisions, kept beside the proposals they corrected, form the customer's labeled set, which fine-tunes that customer's next version on written request and no one else's.
Cahier's GPU work peaks at four moments: the pilot evaluation, the fine-tune at version time, the quarterly re-evaluation and each customer's batch window.
An evaluation of three to five hundred held-out cases takes an afternoon, a fine-tune hours; an insurer's batch runs overnight over a small daytime baseline. Each Paris tenant runs on GPUs of its own, so Paris capacity grows one customer at a time; a self-managed customer provisions two GPU nodes.
Insurers and banks may not send claim or KYC documents to a hosted, multi-tenant model endpoint, and Cahier runs none.
A customer with CISO-approved GPUs runs the image there, and only the signed token meter leaves. One without, whose data must stay in the EU, gets a single tenant on cloud GPU capacity we control in Paris, replicated nowhere outside the EEA.
Cahier Ltd was registered in London in March 2023 and spent eighteen months building its runtime and fine-tuning pipeline.
Twenty customers run Cahier today; the company is bootstrapped, with backing from local investors. Its Paris GPU capacity has served private-cloud inference since early 2025. Two changes come in 2027: embedding retrieval inside each tenant, on the GPUs serving its model, and quantized builds for self-managed customers with fewer or smaller GPUs.
The founders
Edmund Thorley
President
Ran the automation Center of Excellence at a UK general insurer and lost three board cycles in a row to a vendor row that was a press release with no page behind it.
Celia Mendenhall
CTO
Served fine-tuned open-weight models inside a bank's private cloud and had to explain, version by version, why a model that changed under a running pilot could not be signed off.
Rhys Ainsworth
CISO
Wrote a NIST AI RMF mapping into a vendor questionnaire for a regulated buyer and got a SOC 2 report back; he started Cahier so that the pack would exist before the question.