How a UK accountancy firm built an AI agent that reads, understands and files the post

The firm's post room received a daily stack of scanned letters: HMRC correspondence, VAT notices, Companies House mail, bank letters and everything else that arrives when you act for more than 1,500 live clients. Every item had to be opened, read, attributed to the right client and filed in the practice's document management system.

That work fell to people. Someone read each scan, worked out who it belonged to, renamed it, filed it and moved on to the next one. It was slow, it queued behind busier work, and a misread letter could sit unfiled while a statutory deadline ran quietly in the background.

The volume made manual sorting expensive, but the risk made it dangerous. A VAT penalty notice filed against the wrong client, or not filed at all, is not an admin slip. It is a missed deadline with a client's name on it.

Any AI touching this post had to meet the standards of a regulated firm. Data residency for client correspondence had to be assessed and signed off by compliance before go-live, and every automated decision had to be traceable afterwards. The firm wanted the time back, but not at the price of a black box.

The design brief was strict about failure modes before it said anything about features.

• A black box that filed documents with no record of what it decided or why.
• Client post processed before data residency had been assessed.
• An automation that guessed when it was unsure rather than asking a person.
• A pipeline that fell over whenever a letter did not match the expected pattern.
• A rollout without compliance sign-off behind it.

Incoming scanned post lands in one place. AI document intelligence reads each scan, splits bundled pages into individual letters and identifies the sender and the subject. Each letter is then matched against more than 1,500 live clients drawn from the firm's CRM, so attribution runs on current client data rather than a static list.

A matched document is filed into the document management system under its client code, with a fallback copy saved to SharePoint. The agent writes an entry to an audit register recording what arrived, what it decided and where the document went, then posts to a Teams channel so the team can see the day's post without opening a scanner folder.

End to end, a document takes about two minutes from scan to filed, logged and visible.

Not every letter can be filed with confidence, and the design treats that as normal rather than as failure. Letters naming multiple group companies, VAT penalty notices that need careful handling and post from senders the system cannot identify are all routed to a human review queue instead of being guessed at.

That queue is what earned the team's trust. People stopped sorting the whole stack and started reviewing the short list that genuinely needed judgement, with everything else already filed and logged behind them.

The unglamorous parts carried the project. Data residency was assessed and compliance signed off the design before any client document was processed. The audit register was built as a first-class feature rather than bolted on afterwards, because a regulated firm has to be able to show what happened to a piece of client correspondence and who, or what, handled it.

It is also why the firm can honestly call this an agent. The job is defined, the knowledge sources are named, every action is logged and a person handles the judgement calls. That is the standard any AI in a practice should be held to.

Post that used to wait for a person now moves from scanner to filed in about two minutes, whatever else the team is busy with. Human involvement is reserved for the exceptions that deserve it.

The firm also gained something it never had when people did the sorting: a complete, timestamped register of every document that arrived, where it went and why. The audit trail improved because the AI needed one.