Prepare Your Service Catalogue for AI Lookup

“Orientation” does not identify one service

A colleague asks how long Orientation takes. Your catalogue has a 30-minute remote version and a 90-minute on-site version. Choosing the first matching name would lose the distinction the colleague needs.

A third row says “Orientation / on-site / 1.5 hours.” Converting hours to minutes establishes that its duration matches 90 minutes. It does not establish whether the row is an alias, another variant or a separate service with similar attributes. That needs an identity decision from the source owner.

The six rows below are entirely fictional. They contain non-personal service information and produce a preparation record. No importer runs, no AI product is queried and no service is booked.

Give the duration field one meaning

In fictional Guide A v1, “time” means scheduled facilitator-contact duration. It does not mean how long a learner takes to finish, how soon an appointment is available or when a service will be completed. Write that definition before mapping the field.

The field dictionary maps the raw time and unit to contactMinutes, and delivery to deliveryFormat. The preparation record retains raw values and source locations so a reviewer can see that 1.5 hours became 90 minutes. Do not silently treat every column called “length” or “time” as the same concept.

A stable record ID preserves identity. A display name helps someone recognize the service. A delivery format preserves the variant. These fields have different jobs even when one row happens to have a unique name.

Prepare the six-row catalogue

Read the six source rows below, choose the mappings and decide how to treat the possible duplicate. Make a preparation record that classifies every row as trial, held or excluded. The complete fictional data and expected decisions support a paper exercise. If the catalogue workbench is available on this page, choose the field treatments and use “Generate preparation record”; a possible duplicate remains held unless the supplied fictional owner decision supports the selected alias treatment. No records are imported.

The trial has a narrow scope: describe each service’s delivery format and scheduled contact minutes. A row needs either a known value or a source-confirmed not-applicable status. Unknown duration and unresolved identity stay outside this trial. Those are editorial rules for this example, rather than universal rules for every catalogue.

The catalogue coordinator owns each fictional row. SRV-01 includes a guided introduction and questions. SRV-02 and SRV-03 include a guided introduction and room walk-through. SRV-04 covers questions after orientation. SRV-05 is a standard written welcome with no scheduled facilitator contact. SRV-06 supplies a reference pack without a contact session. Retain these distinctions in the preparation record.

Unknown, zero and not applicable carry different meanings

SRV-04, Follow-up, has a blank duration. The source says the value has not been established. Store a null numeric value with an unknown status and keep the record held for this trial. Filling it with a plausible duration would add an unsupported fact.

SRV-05, Written welcome, explicitly has zero scheduled facilitator-contact minutes. Preserve the numeric 0 and known status. Its reading time is not given. SRV-06, Resource pack, is a resource-only item for which scheduled contact duration does not apply. Preserve null with not-applicable status.

W3C’s tabular metadata vocabulary describes datatypes, markers for null values and whether a cell is required. A separate valueStatus field in this exercise carries the business meaning of a missing number. Verify any future tool’s treatment of both fields; the vocabulary alone cannot prove an importer preserves that distinction.

Keep the possible duplicate until the evidence changes

With the worked choices, the first result has four trial rows: SRV-01, SRV-02, SRV-05 and SRV-06. Two remain held: SRV-03 for identity and SRV-04 for unknown duration. Nothing is excluded. The six inputs reconcile as 4 + 2 + 0.

A second, explicitly fictional evidence fixture is owner decision D1. It states that the coordinator compared identity and all variant attributes, confirmed SRV-03 is an alias of SRV-02 and chose SRV-02 as the retained ID. Record D1 and explicitly apply that owner-approved alias decision in a second version of the preparation record. The result becomes four trial rows, one held row and one excluded alias: 4 + 1 + 1.

Supplying the evidence alone does not apply the decision. Choosing an alias treatment without the evidence does not force a merge. The original row and its source-to-retained-ID relationship remain in the preparation history.

Write the questions the trial must handle

A useful trial checks more than whether a file was accepted. Ask for Remote Orientation and expect SRV-01 with 30 contact minutes. Ask for On-site Orientation and expect SRV-02 with 90. Ask only for Orientation and expect a question about delivery format.

Ask for the Follow-up duration and expect an honest unknown, with the held record visible to the source owner. Ask about Written welcome and expect zero scheduled contact minutes without a reading-time promise. Ask about Resource pack and expect not applicable. Preserve each row’s included activities and conditions alongside the answer.

When an actual import is later tested, compare IDs, counts, variants, original-to-normalised values, conditions and value statuses with this expected record. Record observed results separately. A neatly formatted source and a successful upload do not establish the correctness of the imported copy.

Make a readiness decision with limits

The Government Data Quality Framework distinguishes completeness, uniqueness, consistency, timeliness, validity and accuracy. A filled cell can still be wrong. Use these dimensions to explain why a catalogue is suitable for a particular use, and which unresolved issues remain.

The deliverable here is a field dictionary, a reviewed fictional sample and an unresolved-issues list. In a real catalogue, name the owner who can resolve each hold and record the evidence needed to release it. Keep that boundary explicit rather than importing everything because a few sample rows looked tidy.

For the wider decision, read How to Choose Your First AI Project. A well-defined catalogue may support a conventional search just as well; decide what the proposed AI component adds after the information itself is understandable.

Practical next step

Bring a field dictionary and held-record list to an AI consultancy discussion

Sources checked 4 October 2026

• UK Government: The Government Data Quality Framework
• W3C: Metadata Vocabulary for Tabular Data