Postbot and claims intelligence

A proposal for One Claim Limited, for discussion and correction

Status: Draft 0.4, proposal stage Date: 23 September 2026 Prepared by: Custodia UK

One Claim handles motor fleet, home and casualty claims for broker partners and their clients, from first notification through to settlement. Two members of staff currently spend their time reading incoming email, deciding how urgent each one is and filing it into the right folder. The claims system, Keylane SiteView, does not give the team the data they need to do this, and there is a separate need to read medical reports from the Official Injury Claim portal to understand how many claims are whiplash only and how many carry injuries outside the whiplash tariff.

This site sets out what we propose to build, starting with Postbot: an assistant that reads each incoming email and its attachments, works out which claim it belongs to, sorts it into Category 1 (urgent) or Category 2 (routine), files it, and explains its reasoning. The second piece reads medical reports and produces a consistent injury breakdown with page references. Both are built on fixed rules wherever a rule can do the job. Claude, Anthropic's language model, is called through an API key only to read the parts a rule cannot, such as a free-written letter or a medical report, and turn them into facts. Every decision is then made by the rules, which stay under One Claim's control.

Nothing here is fixed. Each page marks our assumptions in grey and the questions we need One Claim to answer in yellow. The open questions page gathers them in one place. The first thing to agree is the demonstration: what it should show and to whom.

Phase 1, first priority

Postbot email triage

Read, match, prioritise and file incoming email. Draft replies for a person to send. Learn from staff corrections.

Part of phase 1

Rule editor and dry run

One Claim's supervisors change the rules themselves and see the effect on the live mailbox, read-only, before publishing.

Phase 2

Medical report analysis

Extract every injury from Official Injury Claim medical reports, separate tariff from non-tariff, and cross-check against the claim.

Supports both

Keylane integration and reporting

How we link to SiteView, what we do if there is no usable interface, and how we fill the reporting gap.

Later

Other applications

Twenty further workflows built on the same engine, from credit hire checks to fleet risk insight, rated for value and feasibility.

Delivery

Project plan

Stages, approval gates, who does what, risks, and how it is run after go-live.

Next step

The demonstration app

A working demonstration using invented claims and emails, so One Claim can see it before any real data is involved.

What we found about One Claim

This is from public sources only. Please correct anything that is wrong or out of date.

Topic	What we found
Company	One Claim Limited, company number 13606052, incorporated 7 September 2021. Operates from First Floor, 6 Atlantic Square, Station Road, Witham, Essex.
What it does	Claims management support for broker partners and their clients "from initial notification through to settlement". Motor, home and casualty, with a specific focus on motor fleet claims.
Ownership	Acquired by Specialist Risk Group in November 2024, shortly after the group acquired Capulus, a motor fleet managing general agent with offices in Witham and Caerphilly. The two teams work closely together. Kevin Smith is named as a director of both.
Claims system	SiteView, originally from 360Globalnet. Keylane acquired 360Globalnet and continues to sell the product as "SiteView, a Keylane product". It is described as a no-code, modular platform that can be integrated with other core systems. We found no public interface documentation.
Injury claims	Low-value road traffic injury claims in England and Wales go through the Official Injury Claim portal, run by the Motor Insurers' Bureau for the Ministry of Justice. Organisations acting for insurers (called compensators) can work in the portal manually or connect their own systems to it.

What this suggests about the work

Fleet claims generate a lot of correspondence from many parties for each incident: the broker, the fleet operator, the driver, third-party insurers, credit hire companies, solicitors, engineers, repairers and medical agencies. Much of it is routine, but a small share carries deadlines that cost money if missed, such as court documents, liability decisions on the Official Injury Claim portal (30 business days, after which liability is treated as accepted) and offer responses. The value of triage is less about speed on routine mail and more about never letting an urgent item sit in a general inbox.

Being part of Specialist Risk Group also matters. Any tool that reads claim correspondence and medical reports will need approval from the group's information security and data protection functions, so we have written the architecture page with that review in mind.

Assumption

One Claim acts on the compensator side of the Official Injury Claim portal, handling third-party injury claims made against its clients' fleet drivers. If One Claim also acts for claimants, the medical report analysis changes emphasis, so this needs confirming (question Q11).

What we heard in the first conversation

First priority: a bot that looks at incoming emails, triages them and sorts them into folders. This is currently done by two people.
Emails are sorted as Category 1 or Category 2: urgent or not urgent.
The claims system is Keylane, formerly 360 SiteView. Does it have an interface we can use?
The system does not provide the data they need.
Separately, read all medical reports on the Official Injury Claim platform to work out how many are whiplash and how many include a non-tariff injury, cross-referenced against other information.

How we propose to approach it

        Rules first, inference only where rules cannot work. Anything that can be decided by a fixed rule is: known senders, portal notifications, reference numbers, claim lookups, thread history, category, folder, tariff amounts, dates and counts. Claude is used only to read unstructured text (free-written emails, letters, scanned attachments, medical reports) and turn it into facts. The software then makes every decision from those facts using rules One Claim can read and change.
Assist first, automate later. Postbot starts by proposing a decision next to what staff actually do. It only files email on its own once the agreement rate has been measured and One Claim is satisfied with it.
The model reports, the software decides. Claude returns extracted facts in a fixed structure, never a folder or a category. Our code validates the facts, applies the rules and confidence thresholds, and only then acts. The model cannot send, delete, forward or move email.
Every decision is explained and logged. Each email gets a short reason, the evidence it relied on and the rule it triggered, so a supervisor can audit it.
Urgent means cautious. Where Postbot is unsure whether something is urgent, it treats it as Category 1. A missed deadline costs far more than an extra look.
No real data until it is approved. The demonstration uses invented emails, claims and medical reports.

    

Proposed order of work

Summarised here; the project plan has the detail, approval gates and risks.

Stage	What happens	Output
0. Specification	Agree this document. Sit with the two staff who triage email today and capture their rules.	Signed-off specification and category rules
1. Demonstration	Working app on invented data showing triage, claim matching and the medical report reader.	A demonstration One Claim can use themselves
2. Accuracy test	Run Postbot over a sample of real historic email (with approval) and compare with where staff filed it.	Measured agreement rate by category
3. Shadow running	Connected to the live Microsoft 365 mailbox in read-only mode. Staff see Postbot's suggestion but file as normal. Supervisors refine the rules with dry runs against current mail.	Weekly accuracy report, list of rule changes
4. Assisted filing	Postbot files routine email it is confident about. Everything else goes to a review folder.	Reduced manual triage; supervisor dashboard
5. Medical reports	Same pattern: test on historic reports, then run on new ones.	Tariff and non-tariff breakdown with evidence