Lesson objective

Design an onboarding workflow in which automation gathers and organises client material while named people decide client acceptance, scope and whether each claim is supported before drafting.

Plain-language explanation

Why this matters now

Onboarding is where an agency's exposure is set. What the agency accepts in the first two weeks (the client, the scope, the claims the client wants made, the lists the client hands over) shapes everything that follows. It is also the most tempting place to automate. Intake is repetitive: forms, folders, brand files, product sheets, evidence packs, logins, customer lists. A workflow can chase missing items, file them, summarise them and draft a first brief in a fraction of the time a coordinator needs.

The risk is that the workflow quietly makes decisions while it collects. The sources below explain why that matters for an agency and not only for its client.

  • Professional guidance (FTC, Policy Statement Regarding Advertising Substantiation, 23 November 1984; a policy statement describing how the FTC applies section 5, not itself binding): the statement reaffirms that "advertisers and ad agencies have a reasonable basis for advertising claims before they are disseminated." It names ad agencies alongside advertisers. The evidence must exist before the claim goes out; it cannot be gathered afterwards.
  • Professional guidance (FTC staff, Advertising FAQ's: A Guide for Small Business, April 2001, last edited January 2025; plain-language guidance, not binding): on agency liability, the FTC looks at "the extent of the agency's participation in the preparation of the challenged ad" and "whether the agency knew or should have known that the ad included false or deceptive claims." The guide says "the advertising agency also may be held legally responsible for misleading claims in ads." It was written in 2001 and does not address AI-generated content.

Read together, these make onboarding a control point. An agency that receives a claim at intake, files it and starts drafting around it has participated, and what arrived in the intake file is part of what the agency knew or should have known.

  • Professional guidance (NIST AI RMF 1.0, NIST AI 100-1, January 2023; voluntary): Appendix C says human roles in decision making and oversight "need to be clearly defined and differentiated," and that human-AI configurations can span "from fully autonomous to fully manual." NIST has announced a revision of AI RMF 1.0; this lesson refers to version 1.0.

A question to sit with

If an intake workflow filed a client's evidence pack, extracted the claims and handed a copywriter a finished brief, at what point did anyone at the agency decide the claims were supported?

What goes wrong

Collection becomes acceptance. The workflow sends the welcome pack, creates the project, grants platform access and books the kick-off as soon as the contract is countersigned. Nobody decided whether the agency should take this client's category of work, or whether it has the people to do it safely.

The brief arrives as a fact. An AI summary of the client's material becomes the brief. Claims in it read as settled because they are formatted neatly, and the copywriter drafts around them.

Scope expands by upload. The client uploads a customer list "for the email programme" and an influencer contract "just so you have it". The workflow files both and triggers the email and creator workstreams, and the agency is now doing work nobody scoped.

The evidence pack is never opened. The workflow confirms that a file called "clinical study" exists. It does not know whether the study supports the claim, and nobody checks.

Automation by the client's clock. Intake is designed for speed because the client wants a launch date. The human decisions are the part that slows it, so they are the part that gets squeezed.

The practical answer

Separate collecting from deciding. Use the three levels from Module 01. Collecting, chasing, filing, summarising and flagging are level 1 and level 2 work, inside the agency. Accepting a client, agreeing scope and judging whether a claim is supported are human decisions. The orchestration below is an AI Coalition Network recommendation.

Stage What automation does The human decision Who decides
1. Enquiry Logs the enquiry, gathers category, channels and markets Whether to proceed to proposal Principal or new-business lead
2. Acceptance Assembles a pre-acceptance summary: category, regulated claims likely, data involved, capacity Whether to accept the client for this work Principal
3. Scope Drafts a scope from the proposal and flags items the client mentioned that are not in it What is and is not in scope Account lead with the principal
4. Brand and asset intake Chases, files and indexes brand files, logos, tone guides, prior campaigns Whether the materials are the current approved versions Account lead
5. Claims and evidence intake Extracts claims from briefs and product sheets, links each to uploaded evidence, flags gaps Whether each claim is supported at the level it implies, before drafting Named claims reviewer; counsel where needed
6. List and data intake Receives lists into a quarantine area and runs the checks in Lesson 3 Whether the list may be used, and for what Privacy owner with the account lead
7. Access set-up Prepares access requests for the client's platforms Which accounts, at what permission, for which agents Technology owner (Module 08)

Gate the drafting workstream on the claims decision. The simplest, strongest control in this module is a deterministic one: the drafting workstream for a claim cannot start until the claims register (Lesson 2) shows that claim as approved to draft, with a reviewer's name and date. A copywriter can still write exploratory work on brand tone. They cannot draft product or performance copy around a claim that is still "in the file".

Make the acceptance decision a real one. The pre-acceptance summary should put in front of the principal the facts that change the agency's exposure: claims in health, finance or other licensed fields (where this programme does not teach the substantive rules and counsel must be involved), data about children or health, political or issue content, and any service that would involve the agency building or selling audiences (Lesson 3). An automation can collect those facts reliably. Whether to accept the work is a judgement.

Quarantine anything that changes scope. Uploads that do not match a scoped workstream go to a holding area, and the account lead is told. They are not filed into a live workstream until the scope is amended.

Keep AI summaries in their place. A summary of the client's evidence is a reading aid, not a finding. An AI output is never proof: before anyone relies on it, the reviewer reads the underlying evidence, and any figure, study result or citation the summary reports is checked against the original document.

Remember this

  • Onboarding sets the agency's exposure: its participation and what it knew or should have known start here.
  • The substantiation policy names ad agencies alongside advertisers, and the evidence must exist before dissemination.
  • Automate collecting, chasing, filing, summarising and flagging; keep acceptance, scope and claim decisions with named people.
  • Lock the drafting workstream for a claim until a reviewer has approved it to draft.
  • Quarantine uploads that change scope until the scope is amended.

Marketing agency example

Kestrel Media Partners, a fictional performance agency, signs a new client, Fernvale Wellness, a fictional maker of sleep supplements. Its operations lead, Callum Price, has built an onboarding workflow in the agency's project tool. In its first version, countersigning the contract triggered everything: project creation, platform access requests, a brief generated by an AI tool from the client's uploads, and copywriting tasks for eight product pages. Within two days the copywriter had drafted pages around "clinically shown to help you fall asleep 30 minutes faster", a line taken from the client's product sheet. The claims reviewer, Aisha Mensah, opened the evidence pack for the first time on day three. The uploaded study summary described a small trial of a different formulation and did not report a 30-minute figure at all. Callum redesigns the workflow. Countersigning now produces a pre-acceptance summary for the principal listing health claims, the absence of children's data, and a request from the client for "lookalike audiences built from our buyers", which is flagged for the privacy owner. The intake agent extracts each claim from the product sheet into the claims register with its source file and marks it "in the file". Copywriting tasks for product pages are created locked. Aisha reviews each claim against the evidence and records a decision: two claims approved to draft with narrower wording, one returned to the client for evidence, and the 30-minute claim rejected. The two drafted pages built on the rejected claim are withdrawn before any client review, and the incident is recorded.

Responsible-use warning

Do not let a countersigned contract or a completed form trigger drafting. Automation that moves from intake straight to production turns every claim in the client's uploads into copy before anyone has judged it, and the agency has then participated in preparing it. The substantiation policy expects the reasonable basis to exist before a claim is disseminated; the agency's control is to lock drafting until a named person has approved each claim to draft.

Human-review requirement

The principal decides whether to accept the client for the work, using the pre-acceptance summary. The account lead owns scope and decides what happens to uploads outside it. The named claims reviewer decides, claim by claim, whether the evidence supports the claim at the level it implies, and records the decision before drafting starts. Counsel owns any question of what the law requires for claims in health, finance or other licensed fields, which this programme does not teach.

Practical exercise

Use a fictional client, or your own agency's onboarding with nothing confidential.

  1. List every step of your current onboarding, from first enquiry to the first draft reaching the client.
  2. Mark each step as collecting (chasing, filing, summarising, flagging) or deciding (accepting, scoping, approving).
  3. For each deciding step, name the person who decides and the record that shows they did.
  4. Identify every event that currently triggers a production task, such as a signed contract, a completed form or an upload.
  5. Redesign the flow so that product and performance copy cannot start until the claims register shows the claim approved to draft.
  6. Write the fields of a pre-acceptance summary that would tell a principal about regulated claims, sensitive data, political content and audience-building requests.
  7. Describe the quarantine rule for uploads that do not match a scoped workstream, and who is told.
  8. Choose one AI-generated summary used at intake and list the figures and citations in it that must be checked against the original documents.
  9. Estimate how much time the redesign adds and where, and state which decisions you would refuse to speed up.
  10. Name the deputy for each decider and describe what the workflow does while a decision is waiting, so that nothing proceeds by default.

Knowledge check

  1. What does the FTC look at when deciding whether an ad agency is liable for a misleading ad?

    Answer guidance: According to the FTC staff small-business guide, the extent of the agency's participation in preparing the challenged ad and whether the agency knew or should have known the ad included false or deceptive claims. The guide is not binding.

  2. Why must the substantiation decision happen before drafting rather than before publication?

    Answer guidance: Because the reasonable basis must exist before a claim is disseminated, and drafting around an unreviewed claim is already participation. Locking drafting until a reviewer approves each claim keeps unsupported claims out of the work entirely.

  3. Which onboarding decisions stay with named people in this lesson's orchestration?

    Answer guidance: Whether to proceed and accept the client, what is in scope, whether brand materials are current, whether each claim is supported before drafting, whether a list may be used, and which platform accounts and permissions agents receive.

  4. What is the status of an AI summary of a client's evidence pack?

    Answer guidance: It is a reading aid, not a finding. The reviewer reads the underlying evidence and checks every figure, result and citation in the summary against the original document before relying on it.

Sources and further reading