Lesson objective

Build a claims register in which intake automation extracts express and implied claims, records the level of support each implies, and leaves the substantiation decision to a named reviewer.

Plain-language explanation

Why this matters now

Lesson 1 locked drafting until a claim is approved. This lesson designs the record that makes that lock work: the claims register, filled by intake automation and decided by people. Level 4 built a substantiation pack for the agency's own AI capability claims. Here the same discipline is applied to the client's product claims as they arrive, one row at a time, before any of them becomes copy.

The FTC's statements set out what the register has to capture.

  • Professional guidance (FTC, Policy Statement Regarding Advertising Substantiation, 23 November 1984; describes the FTC's enforcement position under section 5): what counts as a reasonable basis depends on factors including the type of claim, the product, the consequences of a false claim, the benefits of a truthful claim, the cost of developing substantiation, and the amount of substantiation experts in the field believe is reasonable. Where an ad states a level of support, such as a statement that tests prove a result, "the Commission expects the firm to have at least the advertised level of substantiation." The statement names ad agencies alongside advertisers.
  • Professional guidance (FTC staff, Advertising FAQ's: A Guide for Small Business, April 2001, edited January 2025; not binding): the guide distinguishes express claims from implied claims and says advertisers must be able to support both. Claims need a reasonable basis, "objective evidence that supports the claim", and health or safety claims need "competent and reliable scientific evidence".

Two points follow for automation. First, the register has to hold implied claims, which are exactly the claims an extraction tool is most likely to miss. Second, the wording of a claim sets its own evidence bar. "Clinically tested" and "proven" are not decoration; they promise a level of support.

One more point is an inference, not a quotation. The substantiation policy predates AI and says nothing about it. But it follows from the doctrine that an AI tool's output is not itself substantiation for a factual claim. A model that says a product "is proven to" do something has proven nothing.

A question to sit with

If your intake tool extracted every sentence a client wrote about its product, how many of the claims the eventual ad makes would still be missing from the list?

What goes wrong

Only express claims are captured. The tool pulls "reduces bloating" from the product sheet. It does not capture the implication of a before-and-after image, a lab coat in the mood board or a product name that implies a result.

The support level is lost. The claim is logged as "reduces bloating", while the client's draft headline says "clinically proven to reduce bloating". The register's row no longer shows the level of support the ad would promise.

Evidence is matched by file name. The workflow links a claim to a file titled "study.pdf" and shows a green tick. Nobody has read whether the study tested this product, this dose, this population or this outcome.

The AI answers the evidence question. Someone asks an assistant whether the evidence supports the claim, and the answer "yes, the study demonstrates..." is pasted into the reviewer's notes.

Claims mutate after approval. A claim approved as "may support restful sleep" reappears in a later draft as "helps you sleep". The register shows it approved; the copy says something else.

The practical answer

Design the claims register. The fields below are this programme's template, an AI Coalition Network recommendation. Automation fills the left-hand fields; people fill the decision fields.

Field Filled by What it holds
Claim ID and client Automation A unique ID per claim, per client
Claim text as received Automation The exact words, with the source file and page
Express or implied Automation proposes, reviewer confirms Whether the claim is stated or conveyed by imagery, name or context
Support level implied Automation flags wording, reviewer confirms Words such as "clinically", "tested", "proven", "studies", "doctor", numbers and comparisons
Claim category Automation proposes, reviewer confirms General, comparative, health or safety, financial, environmental, other licensed field
Evidence linked Automation Each evidence file, with what the file says it covers
Reviewer finding Reviewer Whether the evidence supports the claim at the implied level, and why
Approved wording Reviewer The exact wording the evidence supports, which may be narrower
Decision Reviewer Approved to draft, approved with narrower wording, returned for evidence, rejected, or referred to counsel
Reviewer and date Reviewer Named person and date
Version Automation Changes to the claim or evidence since the decision

Let the automation flag, generously. Extraction should over-capture. It is cheap for a reviewer to dismiss a flag and expensive to miss an implied health claim. Useful flags include support words ("clinically", "proven", "tested", "studies", "recommended by"), numbers, percentages and time periods, comparisons ("better", "faster", "number one"), health, safety, financial and environmental terms, and imagery or product names that suggest a result.

Keep the reviewer's finding human. The reviewer answers the questions the policy's factors raise: what kind of claim is this, what product is it about, what would happen if it were false, and what evidence would experts in the field think reasonable? Health and safety claims need competent and reliable scientific evidence. Where a claim sits in a licensed or specialist field, the reviewer records a referral to counsel or to the client's qualified expert rather than a finding. This programme does not teach the substantive rules for those fields.

Use AI to prepare, not to conclude. An assistant can summarise what an evidence file says it tested, list its stated population and outcome, and point out mismatches with the claim. The reviewer then reads the evidence itself. The register never records "AI confirmed support".

Tie copy to approved wording deterministically. Downstream, the drafting and approval tools compare product and performance statements in copy against the register's approved wording and flag any claim that is not there or is stronger than approved. That comparison is an ordinary software check; it does not decide, it routes to the reviewer. Module 07 designs the publication gate that consumes it.

Re-open on change. If new evidence arrives, the claim wording changes, or the product formulation changes, the automation marks the decision as needing re-review. An approval does not survive a change to what it approved.

Remember this

  • Capture express and implied claims, and the level of support each one's wording promises.
  • Where an ad states a level of support, the FTC expects at least that level of substantiation.
  • Health and safety claims need competent and reliable scientific evidence, per the FTC staff guide.
  • An AI tool's output is not substantiation; it can prepare the review but never conclude it.
  • Record the approved wording and re-open the decision whenever the claim, evidence or product changes.

Marketing agency example

Tallbridge & Rowe Communications, a fictional agency, onboards Brightkeel Home, a fictional maker of cleaning products. The intake agent reads the client's product sheets and brand deck and proposes 23 claim rows. The claims reviewer, Daniel Osei, starts with the flags. The agent has marked "kills 99.9% of germs" as a health or safety claim with a number, and "plant-powered clean" as a general claim. Daniel reclassifies "plant-powered" as an environmental claim, because the brand deck pairs it with leaves and a globe. He also adds a row the agent missed: the mood board shows a toddler crawling on a freshly cleaned floor, which implies a safety claim about residue around young children. For the 99.9% claim, the linked evidence is a laboratory report. The agent's summary says "confirms 99.9% kill rate". Daniel reads the report itself: it tested four named organisms on a hard, non-porous surface after a ten-minute contact time. He records the finding, approves narrower wording naming the surface and contact time, and notes that the claim cannot appear in social copy without the qualification. For the toddler image, no evidence addresses residue safety, so he records "returned for evidence" and refers the question to the client's regulatory contact and the agency's counsel. "Plant-powered" is returned for the client to state what share of ingredients the phrase refers to. Of the 24 rows, 15 are approved as received, 5 are approved with narrower wording, 3 are returned for evidence and 1 is rejected. Drafting unlocks for the 20 approved rows only, and the copy check compares every draft against the approved wording.

Responsible-use warning

Do not let an automation record a claim without the words that promise its level of support, and do not let an AI summary of an evidence file stand in for reading it. A green tick next to a file called "study" shows only that a file exists. The reviewer must read what was tested, on what, and with what result, and record the narrower wording the evidence actually supports. A register row without its support wording cannot tell anyone what bar the ad has set for itself.

Human-review requirement

A named claims reviewer confirms each claim's category, express or implied status and implied support level, reads the evidence, and records the finding, the approved wording and the decision with their name and date. The account lead confirms the register is complete before drafting unlocks. Claims in health, finance or other licensed fields are referred to counsel or the client's qualified expert. The client remains responsible for holding substantiation; the agency's reviewer owns the agency's decision to draft.

Practical exercise

Use fictional product material, or a past campaign with nothing confidential.

  1. Choose one fictional product sheet and one mood board or visual reference for it.
  2. List every express claim in the text, with the exact words and the page or section it came from.
  3. List every implied claim conveyed by imagery, the product name or context, and explain what each implies.
  4. For each claim, underline the words that promise a level of support, such as numbers, "tested" or "proven".
  5. Classify each claim as general, comparative, health or safety, financial, environmental or other licensed field.
  6. For two claims, write the reviewer's finding: what evidence would be reasonable, what the linked evidence actually covers, and the approved wording.
  7. Mark any claim that should be referred to counsel or a qualified expert and say why.
  8. Write the rule the downstream copy check will apply against the approved wording.
  9. List three changes that should re-open an approved claim, and who is told when each happens.
  10. Ask an AI assistant to summarise one fictional evidence file, then compare its summary line by line with the file itself and record every place the summary overstated, omitted or misreported what was tested.

Knowledge check

  1. What does the FTC expect when an ad states a level of support such as "tests prove"?

    Answer guidance: Under the 1984 substantiation policy, the Commission expects the firm to have at least the advertised level of substantiation, so a claim that tests prove something needs tests that prove it.

  2. Why must the claims register capture implied claims, and where do they come from?

    Answer guidance: Because the FTC staff guide says advertisers must be able to support both express and implied claims. Implied claims come from imagery, product names and context, which extraction tools are most likely to miss.

  3. What evidence does the FTC staff guide say health or safety claims need?

    Answer guidance: Competent and reliable scientific evidence. Other claims need a reasonable basis, which the guide describes as objective evidence that supports the claim.

  4. What role may AI play in reviewing a claim's evidence?

    Answer guidance: It may summarise what an evidence file says it tested and point out mismatches, but the reviewer reads the evidence and makes the finding. An AI output is not substantiation, and the register never records that AI confirmed support.

Sources and further reading