Document extraction.
A deployed language capability that reads expense documents and returns structured fields. It interprets genuine ambiguity, holds no authority, and writes nothing — which is why it is not an agent.
Deployed capability — this configuration is running in production. Client identities are withheld.
Classified as a capability: it interprets and returns, holding no authority and routing nothing. It is a component other parts of a system build on, not an agent.
A model does part of the work here, under a fixed prompt. It interprets or generates; it is not what grants this entry its authority.
Problem it stands behind
Objective
Turn an uploaded receipt into structured expense data an operator can review, and demonstrate a model doing useful work without being granted authority.
Runtime
01
Receive
Accepts a document reference from a role-gated operator endpoint.
02
Read
Passes the document to a vision model with a fixed extraction prompt.
03
Structure
Returns amount, currency, date, category, tags and description.
04
Abstain
Emits null or empty values wherever the evidence is uncertain.
05
Return
Hands the result to the operator. Nothing is persisted.
Anatomy
Context
- The uploaded document
- A fixed extraction prompt
- A closed category vocabulary
Capabilities
- Extract structured expense fields
- Classify into a closed category set
- Abstain when evidence is insufficient
Activation
- Operator action on a role-gated endpoint
Authority
What can it change?
- Autonomy
- Recommend
- Horizon
- Operator-invoked
Read
- The submitted document
Prohibited
- Persist any record
- Create or modify an expense
- Return card details in extracted tags
- Invent a category outside the closed set
Escalate
- Uncertain field — returned null for operator judgment
- No expense found — returns empty
Composition and verification
Made from explicit parts.
Composition
- Role-gated API endpoint
- Vision model
- Fixed prompt with a closed taxonomy
- Operator review before any write
Verification
- Extraction accuracy is reviewed by the operator against the source document
- Accuracy of amounts and dates prioritised over coverage
- Uncertainty expressed as null rather than a guess
- Holds no write authority, so a wrong reading cannot become a record
No evidence is published for this entry. The items above are our own account of its source, read before publishing — not something you can check without taking our word for it.
Next step
What you can do with this.
This configuration runs in production, client identity withheld. To discuss the same for your operation, start with the problem rather than the system: a first call is thirty minutes and costs nothing.
Related entries