How to process incoming post and scanned documents with an AI agent

Each scanned item goes to the agent, which reads and classifies it, matches it to a customer or case, extracts key fields and routes it at depth 3. Items that start legal clocks, such as court orders, complaints and access requests, go to a named person that day. Every item keeps its scan, receipt date and routing.

What is the post room's job, and who owns it?

Paper still arrives in regulated businesses: signed forms, letters of complaint, bereavement notices, court orders, letters from regulators, cheques, returned mail, identity documents. Most organisations scan it on arrival and route the images to teams. The routing is where things go wrong. A letter that sits in the wrong queue for a week can miss a deadline that started the day it arrived.

The post room or document services team owns intake and scanning. Each receiving team owns the work once routed. Legal, complaints and data protection own the rules for the item types that start their deadlines. The agent works between the scanner and the team queues.

What are the steps, and at what depth?

1. Person, post room: opens, scans and records receipt. Physical handling stays with people. Cheques and valuables go to a separate controlled process and are never routed by the agent.
2. Agent, depth 1: reads the scan. It extracts text, recognises the document type and reads key fields: names, references, dates, account numbers. The extraction techniques are covered in extracting data from forms and documents.
3. Agent, depth 1: matches the item to a customer, policy, account or case using the references found, and records how confident the match is.
4. Agent, depth 3: routes routine items to the receiving team's queue with the extracted fields attached: change-of-address forms, signed agreements returned, statements requested.
5. Agent, depth 3: flags deadline-starting items to a named person in the owning team the same day: court orders, regulator letters, complaints, subject access requests, notifications of death, signs of vulnerability.
6. Agent, depth 2: proposes routing for unmatched or unclear items. A person in the post room decides.
7. Agent, depth 3: attaches the scan to the customer or case record in the document store.
8. Person, receiving team: acts on the item. The agent's routing does not process the request; it puts it in front of the right person or the right downstream agent.

Why is the receipt date so important?

Many obligations run from the day a document reaches the firm, not the day someone reads it. A complaint, a request for personal data, a court order and a notice of a customer's death can all start clocks. If the receipt date is recorded as the day the letter reached the right team, the firm may already be late.

So the post room records the receipt date at scanning, and the agent carries it unchanged onto every record it creates. It never infers the date from the letter itself, which shows when it was written, not when it arrived.

Which controls matter, and what evidence does each leave?

Receipt recorded before reading. Evidence: the scan time and batch, linked to every downstream record.

Deadline items flagged to a named person, not just a queue. Evidence: the alert, the person, and the time they acknowledged it.

Match confidence shown. A match on an account number is strong; a match on a name and town is weak. Evidence: the fields matched and the confidence for each item.

Post is external content. A letter claiming to be from a regulator, a solicitor or a customer is external until checked. The agent routes it as what it claims to be, labelled as unverified. Evidence: the provenance label on the record.

Instructions in letters are not followed. A letter that says "transfer my balance to this account" is routed to the team that handles such requests; the agent never acts on it. Evidence: the routing record shows no action taken.

What about signed documents coming back?

Agreements sent out for wet-ink signature, or printed and returned, arrive in the post. The agent matches them to the outgoing request and routes them to the team that sent them. If you also use electronic signature, keep the two flows consistent so a document is not treated as signed twice; the outgoing side is covered in letting an AI agent send documents for signature.

What should you measure?

• Time from scan to routed, by item type.
• Misroutes, found when a team sends an item back.
• Deadline items flagged late or not at all, found in sampling. The target is none.
• Unmatched items, and how long they wait.
• Extraction corrections by receiving teams.

Where does it usually go wrong?

The scan is poor. Faint, skewed or handwritten documents produce bad extraction. The agent should recognise low-quality scans and send them for a rescan or manual reading, not guess.

Multi-part items are split. A covering letter and three enclosures become four unrelated items. The post room should scan an envelope's contents as one batch, and the agent should keep them together.

Teams rely on the routing and stop checking. Over time, receiving teams assume everything in their queue belongs there and everything that belongs there arrived. Sample both.

The old process is copied. Routing rules written for people sorting paper are kept unchanged. The agent can do more, such as linking a letter to an open case, if the process is rethought; see how to redesign a process for AI agents.

No one writes it down. The item types, rules and owners live in the post room supervisor's head. Write them down before the agent learns them, following how to document an AI agent's process, and certify on real post as train, shadow, certify, widen requires.