Skip to content
A PRACTICAL BUSINESS GUIDE

Prepare data for AI by checking one process, not the entire archive.

Start with one process. Check its sources, access permissions and exceptions.

Define the data needed for one decision

“We have lots of data” does not establish whether it can support the intended task. Preparing an order draft might require customer, product, quantity and unit. Confirming that order may additionally require commercial terms, availability and authorisation. These are different scopes with different requirements.

Read the details

Write down the expected output and work backwards to the information that supports it. For each item, identify its location, owner and the consequence of its absence. This inventory distinguishes a blocking gap from an optional enhancement. Do not ask the model to guess information the business process needs to verify.

Five checks before the first integration

Read the details

Scroll the table to compare all columns.

A source readiness record to complete for each source
CheckEvidence neededExample problem
AccessAuthorised account and permitted operationsThe demonstration works only with an administrator account
MeaningField definitions and units“Quantity 10” could mean items or packs
QualityComplete, incomplete and conflicting examplesThe same customer has different identifiers
FreshnessOwner and update frequencyAn archived price list still appears current
TraceabilitySource reference and versionNobody can reconstruct where an amount came from

Build a sample that reveals problems

Select recent cases and difficult cases rather than only the cleanest documents. Include different formats, rotated attachments, missing lines, abbreviations and new customers. An initial sample helps discover defects; it does not by itself establish future reliability across all production volumes.

Read the details

Keep the correct outcome and its reasoning for every case. If two experienced people disagree, clarify the business rule first. Calling an undefined organisational decision an AI error makes evaluation unhelpful. Limit shared data to what the test requires and use an environment with suitable access controls. Record which formats or business units the sample does not cover.

Illustrative example: ambiguous order quantities

Read the details
  1. Observation

    A request says “3 boxes”, while the catalogue holds a price per individual item. Master data does not consistently record how many items a box contains.

  2. Decision

    The project can extract the wording and prepare a draft. It cannot automatically convert the quantity when an approved mapping is absent.

  3. Targeted repair

    The catalogue owner completes conversions for products included in the pilot. Other cases go to review with an explicit reason.

Data preparation is more than file cleaning. It includes meaning, ownership and behaviour when information is unavailable.

Finish with practical lists and named owners

Avoid searching for a universal “AI readiness” score. The useful output is knowing which activities can be tested now and which depend on a specific repair. NIST stresses understanding context and documenting limitations; this worksheet is our operational application, not a certification. Review it again when a new data source or business unit is added.

Read the details
  • Ready: usable sources, defined fields and tested access for the initial scope.
  • To repair: issues with an owner, concrete action and verification criterion.
  • Excluded: cases the system must recognise and send to a person.
  • To monitor: quality and freshness that may deteriorate after release.
FROM IDEAS TO A BRIEF

A template to work from.

AI data readiness worksheet

Sources, defects, remediation actions and an approved sample.

Download the Markdown template

Document extraction schema

A schema with fields, validation, evidence and missing-data handling.

Download the Markdown template

Knowledge source register

Sources with valid versions, access rules, updates and retrieval tests.

Download the Markdown template

Practical questions

Must we digitise everything first?

No. Start with a scope that has sufficient sources. Document excluded inputs and do not automatically extend pilot results to different archives.

Does the data have to be perfect?

No, but errors must be recognisable and manageable. A missing field may trigger review; ambiguous information treated as certain can cause an incorrect decision.

Who should take part in the assessment?

Someone who performs the process, the source owner and whoever manages access or integrations. Include other specialists when the data and its intended use require them.

References and method

NIST — AI Risk Management Framework 1.0: Core

General reference on use context, measurement and documenting limitations. The checklist and example are Stolen Orbit operational proposals.

Which process should improve first?

Let’s discuss your process