Use AI to Track Selections and Preserve Approval.

A homeowner writes, “Love the warmer tile. Can we see it with the other grout?” An AI summary records “Tile approved.” The sentence is tidy, but it has removed the question that still needs an answer. For a remodeler, that can turn a design conversation into a purchasing mistake.

Use AI to assemble the evidence behind a selection and show what remains unresolved. Approval should come from the project’s accepted approval record, with the exact option and revision attached. A positive comment alone should not change that status.

Document the choice before asking for approval

Buildertrend’s case study on Curate Design + Build, updated September 1, describes a team gathering product information, installation details, and associated costs before clients approve selections. Its current help guide also treats client comments and approval actions as distinct parts of the workflow. These are published examples of selection management, not evidence that an AI can safely approve a choice.

Datum’s proposed AI role is the preparation work: find the relevant messages, match them to the selection record, and identify missing information for the designer. Keep the existing selection system as the record the team relies on. A generated review sheet should link back to it and show when it was checked.

Keep approval attached to the exact version

Consider a hypothetical primary-bath selection. The client has approved a particular wall tile and layout. A later showroom message suggests a different size from the same collection. The assistant should preserve the approved version and flag the new size as a proposed change. It should not transfer approval just because the collection name matches.

The designer still needs to confirm what the new size changes: the layout, trim, quantity, installation scope, or quoted amount. A model number copied from a vendor page does not establish which version the client saw. Link each important detail to the selection record or message that supports it. If the evidence does not identify the size or finish, leave it unresolved.

Anthropic’s guidance on trustworthy agents distinguishes uncertainty an agent can research from questions of human intent that need clarification. Applied to selections, the assistant can find the attached cut sheet. It should ask the responsible person to resolve “the warmer one” when several samples could fit that description.

Give the designer a review sheet they can act on

• Selection: project, room, selection ID, and the specific product or finish under discussion.
• Recorded approval: the approved option and revision, who approved it, when, and a link to the record. Say when no approval was found in the records checked.
• Later discussion: comments or proposed changes that may affect the approved choice, with their dates and source links.
• Open detail: the exact missing answer, such as grout color, tile size, or whether the revised price has been accepted.
• Next step: the designer or project manager responsible, the question to resolve, and the decision date supplied by the project team.

Keep the historical approval and the current open question visible together. A later question does not automatically revoke a prior approval, and an old approval does not authorize a new option. The responsible person decides whether to retain the existing choice, request clarification, or start the project’s change process.

Purchasing needs its own check. Before releasing an order, the responsible buyer should confirm that the current approved item matches the order and that any required cost or scope review is complete. During the pilot, give the assistant read access and have it produce a review sheet. Keep approval, order release, and client messaging with the team.

Test the cases that could produce a wrong order

Use resolved selection threads for a practice run. Have a designer mark the correct approval state and unresolved details before the assistant sees them. Include a positive comment without approval, a changed finish after approval, similar products in different rooms, and a missing attachment. Limit the assistant to the information available at the time of each decision.

Count false approvals, missed proposed changes, wrong-room matches, and missing details. Measure the time a designer needs to check and correct the output against preparing the same review manually. A proposed pilot requirement is zero unsupported approvals in the practice set. That is a suggested acceptance rule, not a measured result or a guarantee about future work.

If the review sheet takes longer to verify than the original thread, narrow the trial to one selection category. The useful output is a specific question the designer can resolve with the evidence already open beside it.

• Check the approved revision behind an AI citation
• Give an AI agent a change order process
• Explore practical AI training for your team

Next step, if this note maps to a problem on your desk: book a conversation with Adam. Thirty minutes, weekday evenings, Eastern time.