12 questions procurement should ask an AI-augmented agency

If an agency says AI makes it faster or cheaper, the buyer should ask exactly how, where the data goes and who is accountable.

Ask about the operating model, not the buzzword

AI can mean anything from a copywriting subscription to a governed internal system. Procurement should ask what work is augmented, what remains human, what data classes are permitted, which providers can touch client material and where approval sits.

Ask what trained means

If the agency says its system is trained or grounded on a communications canon, ask how that knowledge is represented, how source provenance is controlled, how the canon is updated and how live evidence overrides model memory.

Ask how quality is evaluated

A fast answer is not a quality standard. Ask what makes material review-ready, what gets rejected, how unsupported claims are caught and whether the agency red-teams consequential work.

Ask where the saving comes from

Lower cost should come from fewer paid production hours, not from deleting accountability. Ask which tasks are compressed, which human roles remain, and how the efficiency appears in the commercial model.

Ask who answers for the work

There should be a named human owner. Ask who approves strategic recommendations, reputation-sensitive copy and final external release.

The twelve questions

1. What exactly is AI-augmented in this brief? 2. Who is accountable? 3. What data may enter the system? 4. Which providers process it? 5. Is client data used for model training? 6. What is retained and for how long? 7. How are sources checked? 8. What happens when the system is uncertain? 9. What requires human approval? 10. How is output quality evaluated? 11. How does augmentation reduce our fee? 12. Can the process run with tighter AI restrictions if our policy requires it?

New business

TELL US
THE PROBLEM.

Two fields. That is enough to start.

Thanks. We have it and will reply by email.

Keep confidential, health, financial, security-sensitive or legally privileged information out of this first message. See privacy.