The AI Record Type Detection action reads the first page or two of a document, decides which Record Type it is from the Record Types defined on your site, and assigns the Record to it — automatically, as part of a workflow.


Use it at the front of an intake process, when documents arrive without anyone telling the system what they are. A common case is a shared scan or email drop-box that receives invoices, purchase orders, packing slips and applications all mixed together: this action looks at each one and routes it to the right Record Type so the rest of the process knows what it is dealing with.

It is the natural first step of an automated intake pipeline. Once the Record Type is set, the actions that follow — AI Generic Split, AI Document Classification, AI Generic Extraction and Generic Indexing — all configure themselves from that type, because each of them reads the Record's own Record Type to find its categories and fields.


AZURE ONLY: This is an AI action and runs only on Azure-hosted / AI-licensed sites.


How it differs from AI Document Classification

The names sound similar but the two answer completely different questions, and they are usually used together in that order:

  • AI Record Type Detection decides which Record Type this document belongs to — Invoice vs. Purchase Order vs. Application — choosing from the Record Types on your site. It works on the whole document and sets the Record's type.
  • AI Document Classification decides what each page is within one Record Type's Category / SubCategory tree, and splits the pages of a single Record by type.


In short: Record Type Detection picks the Record Type; Classification sorts pages into categories once the type is already known. Run detection first, then classification.


How the Record Type is assigned

This is worth understanding, because it explains the one setup requirement the action has.

A Record in DocMgt has no "Record Type" field of its own. A Record's type is worked out from its data: each Record Type has fields that carry a Filter Value, and a Record belongs to the first Record Type whose filter values all match the Record's data. So "assigning" a Record Type really means writing the data values that make that Record Type match, and that is exactly what this action does.


Two consequences follow:

  • A Record Type is only assignable if at least one of its fields has a Filter Value set. If the detected Record Type has no filter values anywhere, the action reports that it cannot be assigned instead of failing silently — add a Filter Value to the Record Type to make it assignable.
  • Filter values that describe a shape rather than a value cannot be written. Negations (~ or !), ranges (Filter Value To), and "must be empty" tests (# or [EMPTY]) are reported rather than guessed at. Plain literal filter values are the ones that work.


After writing the values, the action re-checks which Record Type the Record now resolves to and reports the result. If another Record Type's filters happened to match first, it tells you so rather than claiming success.


What it works on

  • The action inspects one document on the active Record — the active document, or the first document matching the Category / SubCategory / Name you specify.
  • No OCR is required. The action reads the pages with vision AI directly.
  • By default only the first two pages are sent, because a document's type is almost always obvious from its first page. This keeps the token cost low no matter how long the document is. You can override this on the Document tab.


Options

The action's settings are grouped into four tabs.

Document

  • Active Document — When on, the action inspects the active Record's document if one is available.
  • Cat/Sub/Name — Restrict which document is inspected by Category, SubCategory and Name. The first match is used. Leave blank to match any.
  • Pages — Which pages to send. Enter 1 for the first page only, or a range such as 1-3. Leave blank to use the first two pages. Sending a long document just to identify its type costs tokens without improving the answer.

Detection

  • Candidates — Which Record Types the AI may choose from. Select nothing to offer every Record Type, which is the usual setting. Select specific types to scope the detector to one department — this also improves accuracy, because removing look-alike types from the list removes the chance of confusing them.
  • Minimum Confidence — How certain the AI must be before the Record is assigned.
    • High — assign only when the document names itself or is unmistakable.
    • Medium (default) — assign on a clear, well-reasoned fit.
    • Low — assign on the best available guess.

Below this level the detection is still recorded on the Record, but nothing is assigned. A wrong Record Type sends the document into the wrong workflow, fields and security, so leaving it for a person is usually the cheaper mistake.

  • Assign Record Type — On by default. Turn it off to have the action only record what it detected without changing anything, which is the safe way to test a detector against real documents before letting it act.
  • Lookup Types — Record Types marked as Lookup are excluded from the candidate list by default, since they hold reference data rather than documents.

AI Level

  • Quality — Identifying a document type from its appearance needs a capable model, so only two levels are offered:
    • Standard — good for most standard business documents.
    • Advanced — higher token cost; best for poor scans or highly similar Record Types.

Economy is intentionally not offered for this action.

Advanced

  • Extra Instructions — Free-form hints for this particular workflow, for example "Invoices we receive are Vendor Invoices; invoices we issue are Customer Invoices" or "Treat a remittance stub as a Payment, not an Invoice."


Teaching the AI what your Record Types are

The AI chooses from a list built out of each Record Type's Name and Description. The Description is the single biggest influence on accuracy, and it is the same Description field you already fill in on the Record Type — its prompt even says it is used to help AI work with the type.


A name alone tells the AI very little. "Invoices" does not explain how an invoice differs from a packing slip or a statement. A good description names the distinguishing traits: "Bills received from suppliers requesting payment. Always shows a vendor name, an invoice number and an amount due. Not to be confused with purchase orders, which we issue to vendors."


If detection is picking the wrong type, improve the descriptions on the Record Types before anything else. It is faster and more durable than adding instructions to each action.


What the action writes to the Record

Every run records what happened, whether or not it assigned anything. These values are useful for building follow-up rules and review queues:

  • AIRecordTypeDetect_RecordType and AIRecordTypeDetect_RecordTypeID — the Record Type that was detected.
  • AIRecordTypeDetect_Confidence — high, medium or low.
  • AIRecordTypeDetect_Assigned — True or False, whether the Record was actually assigned.
  • AIRecordTypeDetect_Outcome — a plain-language explanation of what happened, including why nothing was assigned when that is the case.
  • AIRecordTypeDetect_Reason — the AI's reasoning for its choice.
  • AIRecordTypeDetect_AlternateCount plus AIRecordTypeDetect_AlternateRecordType_1, _AlternateConfidence_1, _AlternateReason_1 and so on — the runner-up types the AI seriously considered, which make a close call quick for a person to resolve.


TIPS

  1. Write a real Description on every Record Type you want detected. This matters more than any other setting.
  2. Test with Assign Record Type turned off first. Run a batch of real documents through and read AIRecordTypeDetect_Outcome before letting the action change anything.
  3. Build a review queue on AIRecordTypeDetect_Assigned = False. Those are the documents a person needs to look at, and the alternates tell them what the AI was torn between.
  4. Give at least one field on each Record Type a Filter Value so it can be assigned. A Record Type with no filter values anywhere cannot be assigned by this or any other automatic means.
  5. Use Candidates to scope a detector when a process only ever handles a few types. Fewer look-alikes means fewer mistakes.
  6. Leave Pages blank unless you have documents whose type is genuinely not evident until later pages.

NOTE ON (AI): Reading the document with AI consumes tokens and may incur extra fees. Workflow tokens are billed separately from End-User or Admin tokens — check with your sales rep for more information.