“AI-enabled delivery” can describe a useful workflow, a collection of tools, or a premium label placed on ordinary production. Before paying more, verify what the agency actually does, what evidence it can show, which decisions remain human, and how claims, rights, disclosure, policy, and measurement are controlled.
This is not a benchmark of tools or a legal opinion. It is a diagnostic for the buying decision: continue the evaluation, run a bounded test, negotiate scope, or hold the premium.
Define the capability claim
Write the exact claim: faster creative iteration, audience analysis, media optimisation, production support, localisation, reporting, or another job. Ask what changes for the client, what remains the same, and which result the agency can influence.
Separate a tool licence from a capability. A capability includes people, process, data, review, integration, accountability, and a way to stop or correct the output.
Reconstruct the workflow
Request a walk-through from brief to prompt or input, source data, generated output, human review, platform upload, measurement, and correction. Ask which steps are automated, assisted, sampled, or manual. Require an example that can be inspected without exposing another client’s confidential data.
Record the owner of each step and what happens when a model is wrong, unavailable, or changed. A workflow that cannot explain failure handling is not ready for a premium claim.
Test claims and evidence
List every claim in the agency’s pitch: faster, better, safer, more personalised, lower cost, higher performance, or fewer errors. Ask for the baseline, comparison, time window, cohort, what the agency controlled, and what remains unknown. Do not accept a case study that attributes every result to AI without separating offer, audience, media, creative, and client effort.
If the proposed output contains ad claims, require a source, date, scope, limitation, and claim owner. The IAB AI Transparency and Disclosure Framework provides an industry framework for responsible advertising transparency; it does not prove an agency’s implementation or replace local review.
Review rights and provenance
Ask how the agency records the origin, licence, permission, and usage rights for images, music, data, fonts, voices, likenesses, testimonials, and generated content. The IAB AI intellectual-property and transactions playbook is useful for framing rights and transaction questions; contracts and jurisdictional advice still govern.
Request the tool terms, commercial-use boundary, attribution requirement, training or retention setting, and client ownership of outputs. “AI-generated” is not a substitute for a licence or consent record.
Inspect human approval and disclosure
Identify who reviews claims, identity, brand safety, privacy, accessibility, language, and destination consistency. Ask how the reviewer sees the rendered asset, not only a prompt or text layer. Record version, date, exception, and sign-off.
Decide when the audience or platform needs an AI, synthetic-media, sponsored, or material-connection disclosure. Make the decision placement-specific. Hidden or inconsistent disclosure is a control failure even if the creative is approved.
Check platform and data controls
List platforms, formats, audiences, geographies, automated optimisation features, and data inputs. Review current platform requirements. Google’s Display & Video 360 AI-generated content labelling guidance illustrates why platform labels and internal approvals are separate controls.
Ask where customer data is stored, who can access it, how long it is retained, and how a deletion or correction request is handled. Do not grant broad data access merely to demonstrate a workflow.
Trace measurement and commercial handoff
Walk one output from release to impression, action, lead acceptance, opportunity, and outcome. Separate production speed from media performance and from revenue. If the agency claims better optimisation, require the event definition, attribution boundary, comparison, and stop rule.
Check that the client owns or can export the asset register, source evidence, approvals, logs, and campaign definitions. A report that cannot be audited after the retainer ends is a dependency, not a verified capability.
Compare premium with reversible value
List premium cost, setup, tools, review, integration, usage rights, data work, and client-side time. Compare the cost with a bounded job the agency can demonstrate. A smaller pilot may be more informative than a long premium term.
Record dependencies on a named specialist, model, vendor, or platform. If the capability disappears when one person leaves, the agency should price and govern that dependency honestly.
Ask how the agency trains and reviews its own staff. A repeatable capability should include a current playbook, a quality sample, an escalation route, and a way to retrain or replace a workflow when the tool changes. Treat a one-person demonstration as a pilot signal until the process can be replayed by another authorised team member.
Find the first evidence break
| Layer | Evidence | Diagnostic question | | — | — | — | | claim | baseline, comparison, scope | what exactly is better? | | workflow | input, tool, output, review | can the process be replayed? | | rights | source, permission, terms | can the client use the result? | | human control | reviewer, version, exception | who can stop a bad asset? | | disclosure | placement-specific decision | could the audience be misled? | | platform | current policy and label | will the delivery path accept it? | | data | access, retention, deletion | is the data boundary controlled? | | measurement | event, outcome, export | can the premium be evaluated? |
Decide what to buy
Pay a premium only when the agency can demonstrate a repeatable workflow, evidence-backed claim, rights and data controls, human approval, platform compatibility, measurable handoff, accountable owners, and a reversible test. Choose ordinary delivery or a diagnostic when the AI label adds no distinct capability. Hold when the agency relies on secrecy, unsupported results, vague automation, unowned data, or guarantees.
Keep the diagnostic local and non-indexable until current AI policy, provider overlap, contract, technical, and editorial review are complete. Verified capability is a process the buyer can inspect—not a fashionable word in a proposal.
How did this article land?
Choose one reaction. You can change it anytime.
