AI training for marketing teams should teach employees to turn approved source material into useful campaign assets, verify claims and hand work through the normal approval process. Prompt writing is one skill. Recognizing when a persuasive sentence exceeds the available evidence is another.
A practical workshop can center on one campaign brief and a controlled source pack. Participants produce a draft, inspect its claims and revise it into an asset a reviewer can assess. The plan below is a suggested workshop design, not an advertised LuMay course schedule or a documented customer result.
Define the job participants must complete
Choose a task from the team's actual work: a product announcement, an email sequence, a landing-page outline or a customer education post. Identify the audience, intended action and approval owner before opening an AI tool.
Limit the first workshop to one format. A participant who can verify a short campaign asset has a clearer next step than someone who has generated ten unrelated drafts without reviewing any of them.
Prepare an approved source pack
Include current product facts, permitted claims, positioning guidance, audience notes and approved examples. Label effective dates and owners. Remove customer information and unpublished material that the chosen environment is not approved to process.
Separate verified facts from proposed messages. A positioning idea can be explored creatively, but it should not become a factual performance claim merely because the model repeats it fluently.
The FTC's advertising substantiation policy explains the need for a reasonable basis for objective advertising claims before dissemination. Use that principle to motivate evidence checking, with qualified review for the team's specific claims and markets. The workshop is not a substitute for that review.
Run a proposed two-hour practice session
| Time | Exercise | Evidence of learning |
|---|---|---|
| 0–15 minutes | Inspect the brief, tool boundary and source pack | Participant identifies permitted inputs and missing facts |
| 15–35 minutes | Generate and compare two campaign angles | Selection is tied to audience and task |
| 35–60 minutes | Draft one asset with source references | Claims link to actual supporting material |
| 60–85 minutes | Challenge unsupported claims and vague promises | Participant rejects or qualifies misleading language |
| 85–105 minutes | Revise for channel, tone and clarity | Asset meets the brief without changing factual meaning |
| 105–120 minutes | Submit a review package and explain decisions | Reviewer can reproduce the claim checks |
Adjust the proposed timing to the team's baseline and accessibility needs. Provide the source pack and task instructions in readable formats, and let employees demonstrate judgment without requiring fast typing or public performance.
Teach a prompt that exposes missing evidence
Illustrative prompt template:
Using only the approved source pack, draft [asset] for [audience]. The reader should understand [message] and consider [next action]. For every objective product or performance claim, list its source ID. If evidence is missing, omit the claim and list the missing fact separately. Do not invent customer results, certifications, quotations or comparison data. Keep the requested tone and channel limits. Return the draft and a separate reviewer claim table.
Model-provided source references still require verification. A reference can exist while failing to support the sentence. Participants should open the source and check scope, date, units and qualifications.
Worked example: an unsupported speed claim
Illustrative example: the source pack says a workflow routes documents for review. The generated ad says it “cuts review time by 80%.” No approved study in the pack supports that number.
The participant removes the percentage and writes a narrower statement describing the documented routing function. The reviewer record identifies the rejected claim and why it exceeded the evidence.
Next, the trainer changes the exercise: a valid study is supplied, but it covers a different workflow and a limited sample. The participant must explain why a true number can still mislead when applied to a broader claim. This tests reasoning beyond spotting obviously invented statistics.
Score the work, not tool enthusiasm
Use a proposed rubric with separate checks for factual accuracy, source support, audience relevance, clarity and appropriate escalation. Treat an invented customer quotation or unsupported material claim as requiring revision even when the writing is strong.
Compare a baseline task with an equivalent post-training task. Capture reviewer corrections and total time including verification, rather than measuring generation speed alone. Keep the assessment developmental and consistent with company policy.
Transfer the lesson into everyday production
Give the team a reusable claim table: sentence, source, scope, reviewer and disposition. Assign owners to source-pack updates so yesterday's approved capability does not become tomorrow's outdated promise.
The NIST AI RMF Playbook recommends suitable training and clear accountability across roles. This workshop applies those ideas to campaign production; it does not imply certification or regulatory approval.
Use LuMay's corporate AI training scorecard to plan follow-up measurement. Begin with one approved campaign task and ask whether participants can defend the final wording with evidence.
