Document classification answers what kind of material arrived and where it should go next. It becomes useful when categories correspond to real processing decisions. Before commissioning a classifier, ask the agency to help resolve ambiguous labels; a model cannot fix a category system that your own reviewers interpret differently.
Related service: AI automation development- 01Define category boundaries
- 02Label representative examples
- 03Test confusing pairs
- 04Approve routing rules
SECTION 01
Start with decisions, then names
List the actions that differ after classification: extracting delivery details, reviewing a quote, filing a receipt or requesting missing material. If two labels lead to identical work, splitting them may add maintenance without helping the operator. Conversely, visually similar documents may need different labels when their business meaning differs.
For each label, supply a positive example, a near miss and an owner who can resolve disputes. Microsoft's classification documentation distinguishes classifying a whole file from splitting it into document ranges. Require the agency to explain which unit your workflow labels: file, page, document or package. Do not infer that one PDF always represents one item.
SECTION 02
Illustrative example: quote versus order confirmation
Imagine a distributor receives two documents with the same supplier logo and product table. One is a quotation that needs commercial review; the other confirms an order already placed. A classifier based mainly on layout could route both into the order queue. The category definition must include the document's intent and supporting evidence.
The evaluation includes revised quotations, documents with both titles and a PDF containing several confirmations. Ambiguous material enters a review queue with its candidate category and source pages. An operator correction records why the label changed, so a later update can address a real boundary rather than memorize one supplier's design.
SECTION 03
Accept the operating rules as well as the model
Ask for results by confusing category pair, not just an overall score. A common easy label should not hide mistakes in a less frequent but consequential route. NIST's AI risk framework supports contextual evaluation; your release decision should reflect what a wrong route would actually do.
At handoff, request the label dictionary, sample ownership, routing map and procedure for adding a category. Test unknown files, mixed packages and changing templates. Compare simple rules for reliable identifiers with a model for variable language. A smaller, well-owned taxonomy can be a better first release than a broad classifier whose labels no one maintains.
SECTION 04
Frequently asked questions
Can we classify by filename alone?
Sometimes a controlled naming convention is sufficient. Test whether senders consistently follow it and define a fallback for exceptions before adding AI.
Should every file receive a known category?
No. Preserve unknown or review-needed outcomes so unfamiliar material is not forced into an inappropriate operational route.
Who should approve new categories?
A business owner who understands the downstream work, with the agency checking data and integration implications. Treat label changes as workflow changes.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
- Custom document classificationMicrosoft Learn
- AI Risk Management FrameworkNIST
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy