Practical guide

How to measure applied AI capability in advertising

Applied AI capability in advertising is the ability to use AI within a real advertising workflow while preserving the claim, audience, source, disclosure, approval, and measurement boundaries that make the work defensible. It cannot be inferred from content volume, tool adoption, prompt fluency, or campaign performance alone. This guide uses CLAIM: context, library, AI action, inspection, and.

By Two Prune

Applied AI capability in advertising is the ability to use AI within a real advertising workflow while preserving the claim, audience, source, disclosure, approval, and measurement boundaries that make the work defensible. It cannot be inferred from content volume, tool adoption, prompt fluency, or campaign performance alone. This guide uses CLAIM: context, library, AI action, inspection, and measurement statement.

Source basis and limits

The US Federal Trade Commission states that express and implied objective advertising claims need an appropriate evidential basis before dissemination, and that an advertisement must be considered as a whole. IAB's 2026 AI Transparency and Disclosure Framework V2 provides a risk- and materiality-based approach to disclosure for AI-generated and AI-assisted advertising content. The 2025 IAB and Media Rating Council attention-measurement guidelines emphasize defined measures, data quality, disclosure, and limits when attention data are used. These sources support representative evidence checks; they do not create one global advertising rule, prescribe a campaign strategy, or establish that a Two Prune assessment demonstrates legal compliance. FTC substantiation policy IAB AI transparency framework IAB/MRC attention guidelines

Sources: [1] [2] [3]

The CLAIM measurement record

| Element | Question | Evidence artifact | | --- | --- | --- | | Context | What audience, channel, objective, jurisdiction, and approval stage define the task? | Approved task and audience brief | | Library | Which claims, brand rules, product facts, consent records, and source assets are authoritative? | Versioned source and claim register | | AI action | What may AI generate, transform, classify, analyze, or recommend? | Instruction, model, and output record | | Inspection | Which human checks and specialist reviews are required before release? | Review checklist and exception log | | Measurement statement | What did the participant demonstrate, with what confidence and limits? | Evidence-linked assessment statement |

Representative activities

### Claim-grounded creative development

Give the participant a product brief, approved evidence, prohibited claims, brand rules, and channel constraints. Ask for creative concepts and final copy. Measure whether the participant identifies objective and implied claims, links them to adequate sources, preserves required qualifications, and escalates claims that lack support. A fluent headline is not strong evidence if it changes the meaning of the approved claim.

### AI-assisted asset review

Supply generated text or imagery with a realistic disclosure, provenance, accessibility, or brand problem. Observe whether the participant notices the material issue, checks the applicable policy, records the correction, and routes uncertainty to the right owner. Do not award credit merely for recognizing that AI was used.

### Audience and media analysis

Provide a bounded campaign dataset with definitions, exclusions, and known limitations. Ask for a recommendation. Measure whether the participant checks denominators, attribution assumptions, sample coverage, channel comparability, and uncertainty before using AI to summarize patterns. Campaign lift cannot be attributed to a person's AI capability without a separate causal design.

### Performance narrative challenge

Give an AI-drafted results narrative that confuses impressions, attention, engagement, and conversion or implies causal impact from an observational result. Strong evidence includes reconstructing the metric definitions, identifying the unsupported inference, and producing a narrower decision-ready statement.

Evidence matrix

| Capability claim | Direct evidence | Failure condition | Interpretation limit | | --- | --- | --- | --- | | Frames the advertising decision | Audience, objective, channel, and approval boundary stated | Optimizes output without defining the decision | One task does not show cross-channel mastery | | Grounds claims | Claim-to-source register and qualification checks | Repeats plausible unsupported copy | Legal adequacy needs appropriate review | | Orchestrates AI | Bounded instructions, comparison, and revision | Treats generated output as approved content | Prompt sophistication alone is insufficient | | Verifies work | Reperformed calculations, source checks, disclosure review | Checks tone but not material claims or metrics | Simulated checks do not certify a live process | | Exercises judgment | Tradeoff and escalation record | Hides uncertainty or chooses a metric for convenience | Judgment is context dependent | | Communicates clearly | Decision-ready artifact with limitations | Overstates evidence or causal impact | Polish does not repair weak evidence |

Worked artifact: fictional campaign review

A participant receives a brief for a fictional software campaign, an approved product-fact sheet, a small performance extract, brand guidance, and an AI-generated draft. The draft says the product “cuts reporting time by 50%,” although the source only describes a limited internal usability exercise. It also reports a 22% conversion improvement by comparing two channels with different attribution windows.

The participant starts a CLAIM record. Context limits the work to revising one business-to-business landing page and recommending the next test. The library lists the approved product facts, the study's scope, the channel definitions, and the disclosure policy. AI may propose copy and organize the analysis, but the participant remains responsible for claims and the recommendation.

During inspection, the participant rejects the 50% claim because the cited evidence does not support a general performance statement. They replace it with a descriptive claim that matches the approved fact sheet. They also rebuild the channel table using a common window, label the remaining coverage difference, and decline to call the result causal.

The final artifact contains a claim register, corrected copy, metric reconciliation, decision note, disclosure assessment, and escalation for legal or policy review. The measurement statement says the participant demonstrated source checking and metric discipline in this simulation. It does not say the advertisement is compliant or predict live campaign performance.

Measurement QA

Did the task create a realistic opportunity to verify claims and metrics?

  • Are authoritative sources distinguishable from generated or draft material?
  • Does the evidence show the participant's action rather than only the final copy?
  • Are disclosure, consent, accessibility, and jurisdiction issues routed appropriately?
  • Are campaign outcomes kept separate from assessment evidence?
  • Are missing files and ambiguous instructions recorded as evidence limitations?
  • Can a reviewer reproduce material calculations and claim decisions?
  • Does the report stay within the assessed task and conditions?

Two Prune boundary

Two Prune can assess how a participant handles a bounded advertising-style simulation, including evidence navigation, AI use, claim verification, measurement reasoning, judgment, and communication. It does not provide advertising or legal advice, approve campaigns, certify compliance, attribute campaign outcomes, or make employment decisions.

Sources

FTC: Policy Statement Regarding Advertising Substantiation

  • IAB: AI Transparency and Disclosure Framework V2
  • IAB and MRC: Attention Measurement Guidelines

Sources: [1] [2] [3]

Sources

  1. 1.FTC substantiation policy · Federal Trade Commission
  2. 2.IAB AI transparency framework · www.iab.com
  3. 3.IAB/MRC attention guidelines · www.iab.com