An advertising agency can list AI tools without demonstrating a repeatable capability. A premium decision needs measurements for workflow, claims, rights, disclosure, human review, platform compatibility, data controls, exports, cost, and correction. Build the ledger before paying more.
Define the capability job
Record the claimed job: creative iteration, localisation, analysis, optimisation, reporting, or production support. State the client problem, expected change, time window, comparison, and decision. Separate a tool licence from an operating capability that includes people, process, data, review, and accountability.
Measure the workflow
For a representative asset or campaign, capture input, data source, model or tool, prompt or configuration, output, human review, platform upload, measurement, and correction. Mark each step automated, assisted, sampled, or manual. Record owner, timestamp, version, and exception.
Ask whether another authorised team member can replay the workflow. A one-person demonstration is a pilot signal, not proof of a scalable capability.
Measure the claim and baseline
List “faster,” “safer,” “more personalised,” “higher-performing,” or “lower-cost” claims. For each, record baseline, comparison, cohort, time window, what the agency controlled, and what remains unknown. The IAB AI Transparency and Disclosure Framework provides an industry framework for responsible advertising transparency; it does not verify a provider’s numbers.
Keep production speed, media performance, quality, and commercial outcomes as separate measures. A shorter drafting time can coexist with more review, rights, or correction work.
Measure rights and provenance
Track source, permission, licence, tool terms, commercial-use boundary, attribution, training or retention setting, output ownership, and client export for images, data, music, voices, fonts, likenesses, testimonials, and generated text. The IAB AI intellectual-property and transactions playbook helps structure rights and transaction questions; contracts and local advice govern.
Use “unknown” when provenance is incomplete. Do not treat generated as unrestricted or client-owned without evidence.
Measure human review and disclosure
Record reviewer, version, claim check, brand check, privacy, accessibility, language, identity, destination, approval date, and exception. Confirm the reviewer saw the rendered asset. Measure the share of assets with complete approval rather than counting approvals from a template.
Decide when the audience or platform needs AI, synthetic-media, sponsored, or material-connection disclosure. Keep placement, language, and format fields. A hidden label is not equivalent to a visible disclosure decision.
Measure platform and data controls
List platform, format, audience, geography, automated feature, data input, access role, retention, deletion, and export. Google’s Display & Video 360 AI-generated content labelling guidance illustrates why platform-specific labels are a separate control.
Record disapprovals, policy changes, blocked uploads, and the fallback process. Measure whether the client can retrieve asset, source, approval, and campaign logs after the engagement ends.
Measure commercial handoff
For any claim of better advertising performance, trace output to impression, meaningful action, accepted lead, opportunity, outcome, and cost. Keep event, CRM, and accounting definitions separate. If the agency cannot expose the handoff or the client cannot export it, mark capability evaluation incomplete.
Record client-side effort, expert review, media, production, rights, tool, and correction cost. A premium is not justified by a gross speed claim that ignores these costs.
Add a confidence field to every claimed improvement: observed in the client account, sampled from a small set, self-reported by the agency, inferred from a case study, or unknown. Keep the source and reviewer beside the field. A case study can inform a question without serving as this buyer’s baseline.
Use the capability ledger
| Layer | Measures | Required qualifier | | — | — | — | | job | problem, claimed change, comparison | cohort and period | | workflow | input, tool, output, review, correction | replay and owner | | claim | baseline, evidence, limitation | controlled vs uncontrolled | | rights | source, permission, terms, export | client use boundary | | human control | reviewer, version, exception | rendered asset | | disclosure | placement, language, label | current policy | | platform/data | format, access, retention, export | policy and privacy | | outcome | action, lead, opportunity, revenue status | attribution boundary | | cost | premium, tools, review, corrections | actual or estimate |
Measure resilience and training
Record whether the process survives tool changes, staff absence, policy updates, and a failed output. Track playbook currentness, reviewer training, sample quality, escalation, retraining, and replacement path. A capability that works only while one specialist is present should be priced and governed as a dependency.
Test one correction path: a wrong claim, unusable image, missing disclosure, broken label, or rejected upload. Measure time to detect, owner, action, re-review, and whether the original version remains traceable. The response to failure is part of capability.
Review the ledger after a tool, model, platform, or staff change. Preserve the old workflow, terms, and approval record so a new speed or quality result is not created by silently changing the comparison.
Add a named owner and review date for each change.
Re-run the relevant approval and export tests after every material change.
If the result cannot be reproduced, downgrade the claim to unknown and pause the premium decision until the owner documents the repair.
Record that pause in the provider decision log.
Set the premium gate
Pay a premium when the agency demonstrates a repeatable workflow, evidence-backed claim, rights and data controls, human approval, visible disclosure decisions, platform compatibility, exportable records, measurable handoff, and a reversible test. Choose ordinary delivery when the AI label adds no distinct value. Hold when claims, provenance, data, or cost cannot be verified.
Keep the ledger local and non-indexable until current policy, provider overlap, contract, technical, and editorial review are complete. Measurement is ready when the buyer can audit the capability after the impressive demo ends.
How did this article land?
Choose one reaction. You can change it anytime.
