AI agent for document intake

Documents arrive the way the sender finds convenient: a receipt photographed in bad light, a contract as a PDF, source files in a zip, a signed consent form as a photo. The agent takes them at the door, works out what each one is, puts it where you will go looking for it later, and chases whatever is still missing.

What a document agent does

Takes documents wherever people send them: a chat, an inbox, a shared folder.
Works out what it is looking at — a receipt, an invoice, a contract, a source file, a signed consent.
Files it by client, project and period, following the naming rules you already use.
Sees what is missing from a set and asks the person who owes it for the rest.
Marks which version is the current one without touching the earlier ones.

Processes that run themselves

A playbook isn't a button — it's a process with a condition that starts it. These three get set up most often around incoming documents.

A document lands in a chat

A client drops a photo of a receipt or a PDF into the chat with no message

Work out the type of document and pull out its details
Establish which client, project and period it belongs to
File it under a name someone will actually find it by
If anything can't be read with confidence, set it aside with a question rather than guess

A promised document hasn't arrived

The date by which a document was due has passed and nothing has come in

Check it didn't arrive by another route and isn't already in the folder
Gather everything outstanding for that client into one list
Send a single request instead of three separate ones
Follow up once at the agreed interval, and hand it to you if it stays quiet

A new file contradicts one already on record

A document arrives that disagrees with one already filed: a different amount, a different date, a different version of the layout

Don't overwrite: place it alongside and keep both
Show exactly where the two versions differ
Hold anything downstream that depends on this document
Bring you the decision about which version stands

Where the agent stops

A document it can't read with confidence is never guessed at — it goes aside with a specific question.
It doesn't decide the legal or accounting classification of a document: it files and flags, and the professional interprets.
Where a signed consent or authorisation is required first, its absence stops the process rather than being worked around.
Nothing is deleted or overwritten: a new version sits beside the old one, never on top of it.
What a document means, and what follows from it, stays with the professional. The agent is only answerable for it being found and being in the right place.

See it on your own documentsSetup is a conversation, 15–30 minutes

FAQ

What does it do with a receipt shot in the dark at an angle?

It takes what reads clearly and stops at the rest. A total that is only half visible is not reconstructed from context: the document goes into a separate pile with a specific question — can't read the amount, no date visible. A minute of your attention costs less than the wrong figure booked.

Will it post the documents into the books itself?

It gets them to the state of «filed, and here is what's missing». The classification — which expense this belongs to, whether it falls inside the period — stays with the bookkeeper. The agent shows where every field came from, so checking takes minutes rather than an evening.

Will it overwrite or lose my files?

No. A new version is placed beside the old one rather than over it, and nothing is deleted. Tidying up on its own judgement is the fastest way to lose the version you needed, so the agent only files and marks which one is current.

Do we need a new storage system for this?

No. The agent works inside the folder structure you already have and repeats your naming rules. Moving somewhere new is a separate decision, and it's yours rather than the agent's.

Where this shows up

Walkthroughs where taking in documents removes the bulk of the evening's sorting.

Bookkeeping & outsourced finance