OPENLOCK · PRACTICAL GUIDE
How do you assess whether business data is ready for AI?
Business data is ready for AI when it can support a defined task with approved access, sufficient coverage, agreed meaning, current information, and answers that can be checked against evidence. Readiness is a decision about a particular use case, not a universal score for a company’s data.
1. Define the decision before profiling the data
Write down the question, the person responsible for the decision, and the action an answer could inform. “Which components are missing for approved orders due next week?” needs different evidence from “What did we sell last quarter?” Specify the time period, facilities, products, source systems, allowed users, and expected output.
Select known examples with the business owner: a normal case, a missing record, a changed revision, an ambiguous identifier, a stale source, and a user who must not see the result. Keep a separate set of examples for acceptance, so passing only the records used to build a mapping is not mistaken for general reliability.
2. Inspect seven readiness dimensions
| Dimension | Evidence to collect | Acceptance question |
|---|---|---|
| Access | Approved source owner, access method, and repeatable extraction. | Can the scoped data be obtained again without unapproved access? |
| Meaning | Field definitions, units, time zones, and source authority. | Does “available stock” exclude reservations and quarantine? |
| Relationships | Identifier crosswalks, approved matches, and rejected matches. | Can an order be joined to the correct product revision without duplication? |
| Coverage & quality | Missing-key counts, duplicate checks, and reconciled totals. | Are all in-scope records accounted for, including exceptions? |
| Freshness | Source event time, extraction time, and a task-specific age limit. | Will a late refresh qualify or stop the answer? |
| Governance | Named owners, permission tests, and change authority. | Can restricted information enter a joined answer or model context? |
| Traceability | Source record identifiers, transformation versions, and calculation inputs. | Can a reviewer reproduce the result from the recorded snapshot? |
3. Set gates that reflect the consequences
Hypothetical acceptance example: a manufacturer evaluates a weekly component-shortage report. Its operations owner requires every included order line to have an approved product and bill-of-materials revision. Unmatched lines enter an exception list and prevent the report from being labeled complete. The owner chooses a maximum stock-snapshot age of 24 hours for this planning exercise; that threshold is an example, not an OpenLock service commitment.
Reconcile demand quantities to the source export. Check that joins have not multiplied order lines. Recompute several component totals independently. Test that a user without access to a facility cannot retrieve its stock, even through a combined report. Introduce a late source refresh deliberately and confirm the report shows the age and withholds an unqualified answer.
A high overall match percentage can conceal one missing high-priority order. Report both counts and the operational effect of exceptions. Thresholds should be agreed by the person accountable for the decision; security or required-source failures should not be averaged away by a good score elsewhere.
4. Record a clear outcome
- Ready for the scoped task: agreed checks pass, limitations are visible, and the business owner accepts the evidence.
- Ready with restrictions: a narrower scope is useful, and excluded records, users, or actions are explicit.
- Remediation needed: missing sources, unresolved definitions, failed permissions, or unreliable freshness prevent the intended use.
The assessment should produce a source register, a glossary and mapping decisions, a test pack with expected answers, an exception backlog, and a recommendation to proceed, narrow, or pause. A sample assessment cannot establish full production coverage. New sources, seasonal changes, and new users require further checks; a passing test also does not guarantee that a model will always interpret an answer correctly.
OpenLock’s engagement approach uses these kinds of records to define scope and acceptance. See the manufacturing walkthrough for an inspectable calculation.