Filing and naming scanned documents with AI

How the bot works

Nine steps. The naming, the folder arithmetic, the duplicate check and the audit trail are ordinary code and cost almost nothing to run; the model is only used to read the page.

1. It watches where documents actually land. The scanner's output folder, the shared drive everybody dumps into, a mailbox, a photo somebody took on a phone, a portal download. All of them, rather than one tidy inbox that only exists in a diagram.

2. It reads the page, not the filename. Scan0001.pdf tells you nothing, so it is ignored. The document is read — text layer where there is one, OCR where there isn't — for what kind of document it is, who sent it, who or what it concerns, and the date on the page rather than the date it was scanned.

3. It names it by your rule. The pattern is yours and it is written down: which fields, in which order, in which format, with which separator — and what to put where a field is genuinely missing, which is the part every naming convention forgets.

4. It works out the folder from your own tree. By client and year, by property and unit, by entity and document type — whatever your structure already is. It does not invent a new hierarchy, and it does not move anything into one of its own.

5. It scores how sure it is, and stops under your floor. Two scores, not one: how sure it is what the document is, and how sure it is where it goes. Either below your floor and nothing is filed — the document waits, with both proposals and the words it read them from.

6. It catches the same document arriving twice. The emailed copy, the scan of the printout and somebody's phone photo of the same page are one document. It compares content, not file size, and keeps the copy your rule says wins while recording the others.

7. It files into your storage, readable without us. Your drive, your folders, your filenames. Someone who has never heard of us can open that tree and find things, and if you switch tools in three years nothing has to be exported — the files were never anywhere else.

8. The uncertain ones come to a person as a short queue. The page on one side, the proposed name and folder on the other, and the reason for each. Confirm, correct, or reject. A correction is worth more than a confirmation: it is where your rule gets sharper.

9. Every move is recorded, and every move is reversible. What the file was called, where it came from, what it is called now, where it went, which rule and which prompt version decided that, and who confirmed it. Any run can be read back and undone.

What you configure

It can be configured to run in any business. This is what that actually means — the list of things that are yours, not ours.

Your naming rule

The pattern, field by field: date format, which party comes first, the document type, any reference number, the separator, the case — and the fallback for a field that is genuinely not on the page.

Your folder structure

The tree as it already is, which level each document type lands at, and what happens when a document belongs to two places at once.

Your document types

The list, in your words, and what each type must have before it can be filed at all — a date, a party, an account, a signature.

Your confidence floor

The score below which nothing is named or moved, set separately for what the document is and where it goes. Start it high; the dry run tells you where it belongs.

Your retention rule

How long each type is kept and where, which types are under a hold, and who is told when something reaches the end of its life — which is a person's decision, always.

Your intake points

Which scanner folders, which mailboxes, which drives and which portals are watched, how often, and what happens to a file that arrives in the wrong one.

Your duplicate rule

What counts as the same document, which copy wins when there are three, and whether the losers are kept somewhere or just recorded.

Your review queue

Who confirms the uncertain ones, how quickly, and where a document waits in the meantime — because a holding folder nobody looks at is just a worse inbox.

Your storage and access

Where the files live, who can see which part of the tree, what we are allowed to read, and whether anything leaves your own storage at all.

Transparent, and configurable by you

The dark band below is what the system will not do. This is what you can see and change — and it is how we build everything, not something invented for this page.

You see what it did, and why

Every filing shows the two scores, the words on the page it read them from, the rule that produced the name and the rule that chose the folder. You can see why a document is where it is without opening it.

Versioned prompts you can tune yourself

The prompts behind reading a page are yours to read and yours to tune. A change is a reviewable diff, replayed against documents you have already filed by hand so you can see exactly what it would have named them, and rolled back the same way.

A human in the loop wherever judgment actually matters

Reading, naming and filing the confident ones is automatic. Confirming an uncertain one, deleting anything and every retention decision are a person's — that boundary is a property of the system, not a promise about how carefully we'll use it.

Every run logged, along with what it cost

One row per document and one per move: where it came from, what it was called, what it is called now, where it went, the scores, the model and prompt version, and who confirmed it — plus token usage and cost for that run.

It can be configured to run in any business. The rule and the tree are yours: your naming pattern, your folder structure, your document types, your confidence floor, your retention. The scanned-paperwork wiring on this page is one configuration of it, not the limit of it.

Works with your tools

This runs on top of the scanner, the drive and the mailboxes you already have. We're not moving your documents into a system of ours.

• Google Drive
• SharePoint or OneDrive
• Dropbox
• A document management system
• A scanner or multifunction printer
• Microsoft 365 or Outlook
• Google Workspace
• DocuSign or a similar tool
• An office file server
• A spreadsheet index you keep

Works well for

• Law firms — Everything belongs to a matter, and a document in the wrong matter is the one nobody finds again.
• Accounting and tax firms — Client paperwork arrives in a flood with a deadline on it, named however the client's scanner felt like.
• Insurance agencies — Policies, endorsements and claim correspondence all look alike and all belong to a different file.
• Medical and dental practices — Referrals, results and forms arrive by fax, scan and portal, and each one belongs to one chart.
• Engineering and design firms — Drawings and revisions live or die by the filename, and the convention is already written down.

How a rollout works

Nothing is renamed on day one. The first milestone is a list of what the bot would have called every document and where it would have put it, next to where they actually are.

1. 01Discovery

   The quote call, then a working session. We go through your naming rule as it exists today, your folder tree, your document types, what each one needs before it can be filed, your retention and holds, and which intake points are real. You get the naming specification, the folder map, the exception classes and a written scope.

2. 02Dry run

   The bot reads documents you have already filed and proposes a name and a folder for each one. Nothing is renamed and nothing is moved. We compare its proposals against where they actually live, and every disagreement is either a rule we have not captured or a genuine judgment call. This is where your own numbers come from, and where the floor gets set.

3. 03Assisted mode

   Live on one intake point or one document type. Confident filings go through; everything under the floor arrives in the review queue with its proposals. Filing stops being typing and becomes confirming. The control doesn't change.

4. 04Optional wider rollout

   Optional. Once a type has run clean for a while, more types and intake points come on, with every hard stop in force, no delete rights ever, and the kill switch. Plenty of teams keep the queue on everything, and that's a good place to stop.

What we don't claim

• No numbers we haven't earned. No accuracy figure, no time-per-document, no claim about what finding a file costs you. We haven't read your documents. The dry run over paperwork you have already filed is where those numbers come from, and they'll be yours.
• No customer claims on this page. No names, no case studies, no logos.
• It cannot read what is not there. A crooked scan with the date cut off, a handwritten note, a third-generation fax: it will tell you it cannot read them rather than guess a date. Paper quality is the real limit on this workflow and no model fixes it.
• It is not a records policy. The retention periods, the holds and the question of what has to be kept where are yours. We turn them into rules the bot respects and alerts it raises — we will not tell you that your obligations are handled.
• Fully autonomous filing isn't on offer. There is a floor under which nothing moves, there are hard stops that never clear themselves, and there is no delete. That's deliberate, and it's why the boundary is written down.

Questions people ask

Can AI file and name our scanned documents?

Yes - and the part that matters is what it does when it is unsure. An agent can read the page rather than the filename, work out what the document is, who it concerns and what date it carries, rename it using the pattern you wrote, and place it in your own folder tree. Where either of those two judgments falls below the confidence floor you set, nothing is renamed and nothing is moved: the document waits in a short queue with its proposed name, its proposed folder and the words it read them from, for a person to confirm or correct.

What naming convention does it use?

Yours. The pattern is configuration, field by field - the date format, which party comes first, the document type, any reference, the separator and the case - and so is the fallback for a field that genuinely is not on the page, which is the part most conventions never specify. If you do not have a convention written down yet, the discovery session is where we write one, and it is yours whether or not you keep us.

What happens to a document it cannot place?

It stays exactly where it arrived, untouched, and appears in the review queue with the reason. The hard stops that can never clear themselves are an unreadable scan, a document type it has not seen before, a page with no date it can trust, two candidate folders it cannot choose between, and anything matching a pattern you have marked sensitive. None of those are resolved on confidence alone, in any configuration.

Do the files stay in our own folders?

Yes, and that is deliberate. The files are named and filed into your own storage, in your own tree, under your own filenames - not copied into a system of ours with your folders reduced to tags. Somebody who has never heard of us can open that drive and find things, your backups keep working, and if you replace us in three years there is nothing to migrate because the documents were never anywhere else.

Does it delete anything, or handle retention?

It deletes nothing, ever - not a duplicate, not a blank page, not a file it cannot read. There is no configuration that gives this bot delete rights. On retention it does the useful half: it knows the period you set for each document type and which hold a document sits under, and it tells a person when something has reached the end of its period. The decision is always a person's, and your records policy stays your responsibility.

Can it read a bad scan or a photo taken on a phone?

Often, yes - a phone photo of an invoice on a desk is a normal input, and so is a scan that came out crooked. But it reports what it could not read instead of filling the gap in: a date cut off at the edge of the page, a handwritten amendment, a fax of a fax. Those come through as exceptions with the image attached. Paper quality is the real limit on this workflow, and no model removes it.

What about the same document arriving twice?

It is treated as one document. The emailed copy, the scan of the printout and somebody's phone photo of the same page are compared on content rather than on file size or name, the copy your rule says wins is the one that gets filed, and the others are recorded so you can see they existed. Which copy wins - the first to arrive, the highest resolution, the one with the signature - is your rule.

Does it also pull the numbers out of the document?

No, and that is a deliberate line. This bot names a file and puts it in a folder. Reading the line items off a supplier's invoice and posting them as a bill, and pulling dates and obligations out of a signed contract, are separate jobs with separate boundaries and separate approval steps. Both exist - get a quote and we will scope them properly rather than bolting them onto a filing bot.

Is there a record of every move?

Yes, and any run can be undone. Every move records where the file came from, what it was called, what it is called now, where it went, both confidence scores, the rule and the prompt version that decided it, and who confirmed it where a person did. Retention of that log is yours to set, access is scoped to the folders in scope, and nothing is used to train anything.