Learning AI · Practical guide
Define the fields before asking for extraction.
Turn a fictional note into reviewable data. No integration is required: start with text, a field contract and a result check before copying anything into another tool.

The process at a glance
Text
Request S-17: three notebooks
Fields
Quantity: 3 · cost centre: null
Check
Verify each value against its passage
01Use a small note with one missing value
Fictional note: “Request S-17. Paula asks for three notebooks. Collection on 12 October 2026. Cost centre still to be confirmed.” Fields are identifier, requester, item, quantity, collection date and cost centre.
Define each field’s meaning. The date concerns collection, not creation. Cost centre must stay empty; the text does not let you infer it from Paula’s name.
02Copy an extraction contract
Instruction: “Extract only the defined fields from this note. Include a supporting passage for each value. Use null when unspecified. Do not fill gaps using outside knowledge. Return quantity as an integer and the explicit date as YYYY-MM-DD.”
An instruction proposes the output format; reviewing that output remains part of the work. If the text contains an ambiguous date, add a rule requesting clarification rather than guessing its interpretation.
03Compare with the expected answer
Expected: identifier S-17; requester Paula; item notebooks; quantity 3; date 2026-10-12; cost centre null. Converting “three” to 3 is allowed by the contract; assigning a cost centre is not.
Check each value and its evidence. A correct field paired with an unrelated passage makes auditing harder. Keep the original text associated with the extraction while reviewing it.
04Test conflicts before reuse
Change the note to mention three notebooks and then five without saying whether this is a correction. Expect a conflict or review marker. Next, include two requests in one text to check that their fields stay separate.
Validate required fields and review exceptions before importing. Preserve unresolved rows with reasons rather than silently converting them into complete records.
Reference [1]: NIST · AI 600-1
Sources and further reading
These references expand on the concepts indicated. The examples and exercises are original editorial material.
[1] NIST · AI 600-1
Generative AI Profile ↗
Context on unsupported content.
Back to the related section
Frequently asked questions
Is a correctly formatted value already verified?
No. A date can match the requested format but refer to another event. Check its meaning and supporting passage as well as its shape.
Keep reading
Learning AI
How to check an AI assistant’s answers before using it ↗
Build a simple test with questions, evidence and acceptance criteria. Use a worked exercise to find unsupported answers before relying on an assistant.
Learning AI
How to prepare a CSV before importing it into an app ↗
Check columns, identifiers and missing values with a fictional sample. Decide what to preserve, correct or set aside before importing.
Learning AI
Turn meeting notes into tasks with AI: a checkable exercise ↗
Practise extracting commitments, owners and deadlines without inventing them. Includes fictional notes, a reusable instruction and a reference answer.