One line saying AI may be used with approval leaves every real question open. Here's what to write instead, before the first draft exists.
Settle AI in the brief, not in week three. Write seven lines: approved tools, what you will and won’t hand over, what stays human, named sign-off on both sides, what counts as a revision, people and rights, and proof. Use the sample brief as a model, score yours out of 14, and send it before the work starts.
A creative brief has a place for the audience, the tone, the deliverables, the deadline and the budget. It rarely has a place for the question that now starts a lot of the arguments: where AI is allowed to touch the work, with what of yours, and who looks at the result. The usual stand-in is one line, “AI may be used with prior approval,” and that line leaves every real question open.
Which tool, and on what plan? Can the unreleased pricing go into it? Who checks the output before you see it, and does a regenerated image count as a revision? Each takes a minute to settle on day one. In week three, with a draft on the table and an invoice attached, the same question takes a meeting.
This is a guide to how to brief an agency on AI, or a freelancer, before the first draft exists. My argument is that the AI terms of a creative job are cheapest to settle in the brief, as seven specific decisions, and that writing them down protects a good supplier as much as it protects you. The seven are:
If you’re partway through a relationship and want to find out what your agency already does, start with what to ask your marketing agency about AI. That’s the renewal conversation. This is the document you write at the start of a job, and it runs both ways, because you may be using AI on your side too.
The seat you brief from changes the advice. An in-house marketer owns the brief and the approved-tool list, so the seven decisions are largely theirs. A founder is the marketer, the buyer and the legal department at once, so I’d keep all seven lines, hold each to a sentence, and pay a lawyer to read the rights line once. Procurement can turn the tools, inputs and rights lines into a standard schedule that every brief inherits. The supplier changes it too: a freelancer is both the maker and the checker, while an agency has roles, subcontractors and an AI policy of its own.
Start with inputs, because they’re the part you fully control. Whatever you hand a supplier can end up in an AI tool, whether anyone intended that or not: a brand guide pasted into a chat to get the tone right, a customer list used to draft a segment email, a roadmap summarised for the copywriter. Some of that is fine and some isn’t.
A designer, the day before a deadline, pastes the unreleased feature names into a chat tool on a personal account to tidy a caption. The caption is better. Nobody told them the names were off limits, because nobody had written the line. Three buckets are a workable start, and your own AI acceptable use policy is where the examples should come from if you have one.
| Bucket | Examples | What the brief says |
|---|---|---|
| Green | Published website copy, public decks, your brand guidelines, anything already in the market | May go into any approved tool |
| Amber | Unreleased feature names, draft messaging, internal positioning, early creative | Approved tools only, on the plan named in the brief, with the supplier confirming the training setting |
| Red | Customer lists and personal data, pricing or terms under NDA, other clients’ material, unedited interview recordings | Not to be put in any AI tool. You say if a file moves buckets |
A Future Factors working view, based on our own judgment, not a standard. Move examples between buckets to fit your risk.
The tool matters as much as the file, because the same assistant can sit on different plans with different data terms. OpenAI’s own privacy page says data from its business, enterprise and API products isn’t used to train its models unless a customer opts in, and lists data from its individual versions of ChatGPT as a training source.[1] That’s a vendor describing its own product at one point in time, so I’d treat it as a reason to name the plan in the brief and ask for confirmation in writing, not as a guarantee. If you need a way to decide which tools belong on your list, our AI tool evaluation checklist covers the eight questions.
Decide what goes in before you decide what comes out.
Say what you’ll do on your side too. If you plan to feed a supplier’s drafts into your own AI tool to summarise or critique them, name the tool.
Next, decide what stays with a person, deliverable by deliverable. Write it as a list of things AI doesn’t do, not as a feeling about authenticity. Claims, customer quotes and voices are the usual candidates, because a wrong one costs more than a slow one. For the rest, pair each AI-assisted item with a named person on their side and a named person on yours, and put the stages in the brief: concept, draft, final, with a sign-off at each.
The table below is how I’d lay it out for the campaign in the next section.
| Deliverable | What AI does | What the human still owns | How it gets checked |
|---|---|---|---|
| Landing page copy | Drafts options from approved messaging | Every product claim, the voice, the final wording | Your product marketing lead matches each claim to the feature list before design starts |
| Social post variants | Generates headline and body variants from approved concepts | Choosing which run, and rejecting off-brand ones | Your strategist compares what goes live with approved messaging |
| 30-second explainer video | Captions and clean-up of audio a person recorded | The script’s claims and the voiceover, which a person records under a signed release | You review the cut against the script, and ask to see the release |
| Hero images | Produces concept options | Choosing one, then reworking it by hand; checking logos, text inside the image and faces that resemble real people | Your brand manager reviews against the guidelines, and you ask for the working file |
| Customer quotes | Nothing | A person writes them from real input, with the customer’s written approval | You ask to see the approval for any quote you didn’t hear yourself |
An illustrative split for a launch campaign. The right split depends on the deliverable and the risk, so treat the rows as starting points.
Sign-off lines tend to be lopsided. They list who signs for the agency and leave the client’s side blank, and the delay shows up on your desk instead. Name your own approver for each stage, and say how quickly they’ll respond.
Whoever signs off on your side should read AI-assisted work with some suspicion, because models can state wrong things in the same tone as right ones. Our explainer on why AI hallucinates shows why, and it’s why the checking column names a source to check against.
Here’s the AI section filled in for a launch campaign. The company, the project and the dates are illustrative. The first row is the ordinary brief. The next seven rows are the AI section, and they’re the part that usually goes missing.
| Line | What the brief says |
|---|---|
| Project | Launch campaign for the new reporting feature: landing page copy, six social posts, one 30-second explainer video, three hero images. Drafts due end of week two, final end of week four. |
| 1. Tools | Writing and ideation: a general-purpose chat assistant on a paid business plan with training on our data switched off, confirmed to us in writing. Images: one generator, named in the kickoff note. Anything else needs our yes by email first. Voice cloning and face-swap tools are not approved. |
| 2. Inputs | Green: published site, public decks, brand guidelines. Amber: unreleased feature names and draft messaging, approved tools only. Red: customer lists, pricing under NDA, other clients’ material, interview recordings. We’ll tell you if a file changes bucket. |
| 3. Human only | Product claims, customer quotes and the voiceover. A person writes or records them. AI may caption and clean up audio a person recorded. |
| 4. Sign-off | Concept: your creative lead and our brand manager. Draft: your editor and our product marketing lead, who checks claims. Final: our marketing director. Each stage is approved by name in the project thread, and we reply within two working days. |
| 5. Revisions | Two rounds of consolidated feedback per deliverable. Regenerating options is not a round. An error of fact or a miss against this brief is not a round. More options than agreed is by agreement before work starts. |
| 6. People and rights | A written release for any real person’s face, voice or name. No synthetic people without our yes. No work “in the style of” a named living artist. You flag anything that needs a platform AI label, and we apply it on upload. Rights in the final work pass to us on payment, and our lawyer will review the contract wording. With each final file, you send a short record of what a person created, chose or changed. |
| 7. Proof | With each delivery: a source for every product claim, the tools used, and the working or layered file for images. We spot-check one item per delivery, picked at random, and ask for the sources behind it. We won’t rely on AI detectors. |
| Our side | We use an approved AI tool to summarise your drafts and prepare feedback. We don’t upload your unfinished work to a tool that trains on it. |
Illustrative brief for a fictional launch, not a template to paste unchanged. Dates, deliverables and approvers change from job to job; the seven lines stay.
A freelancer’s version is shorter. Line 4 can name one checker, and line 7 becomes a sentence about what they verify before sending. For an agency, add a line making the brief apply to anyone they bring in. The lines that tend to come back blank are 2 and 7, so fill them in before you send, not after.
Revision language is where AI changes the economics without anyone noticing. A draft takes minutes now, so a round is no longer a unit of effort. A brand manager gets six hero options on day two, two of them close, and a week of “can we try it warmer” and “can we try it with a different person” follows. Each request is answered within the hour, and by the end nobody can say how many rounds have been used.
Define a round by the feedback, not the production: one consolidated set of comments from you, answered by one revised delivery. Then say what doesn’t count.
| What you ask for | Counts as | Why |
|---|---|---|
| A change to the message, claim or direction | A round | You’re changing the brief |
| A fix to a factual error or a miss against the brief | Not a round | The supplier’s check should have caught it |
| More options than agreed | By agreement, before work starts | Producing options costs little now, but reviewing them costs time on both sides |
| A re-run in a different tool | Needs your yes | It changes line 1, and possibly where your files went |
Our suggested wording logic, not a standard. Set the round count to the size of the job.
Likeness is where I’d stop negotiating and write a flat line. A real person’s face, voice or name goes into the work only with a written release, and a synthetic person only with your yes.
If a real person’s face, voice or name is in the work, there’s a written yes behind it.
That’s my judgment about where the risk sits, not a statutory rule, and the law on voice and image varies by place, so it’s a line for your counsel to firm up. On the ethics, see our piece on deepfakes at work.
Disclosure to your audience is a platform question before it’s a brief question. YouTube says it requires creators to disclose AI-generated or altered content that seems realistic, and it lists what doesn’t need a label, including production help such as scripts, outlines and thumbnails, captions, upscaling, and minor cosmetic edits. Creators who repeatedly skip disclosure can have a label applied or face penalties including removal or suspension from the YouTube Partner Program.[2] Other platforms word theirs differently and change them, so the brief shouldn’t copy any one. It should say who checks the current rules for each channel and who applies the label at upload, usually whoever publishes. For the wider list of what to check before anything goes live, use our AI marketing compliance checklist.
Ownership is the line where a short clause hides a lot of uncertainty. In January 2025 the US Copyright Office concluded that, based on how current generally available technology functions, prompts alone don’t provide sufficient control to make a person the author of what comes back. It also said human authors are entitled to copyright in their own expression perceptible in AI outputs, and in creative selection, arrangement or modification of them, with sufficiency decided case by case.[3] That’s a US view of US law, and I’m not a lawyer, so this is a way to write a brief, not legal advice.
In practice it changes what you ask for. A clause saying all rights in the final work pass to you is fine as far as it goes, and it can only pass on rights that exist. So ask for the thing that helps you later: a short record, for each deliverable, of what a person created, chose or changed. The Office’s 2023 registration guidance asks anyone registering a work to disclose AI-generated content that is more than de minimis and to describe the human author’s contribution,[4] and a record made while the work is being produced is far easier than one rebuilt a year later.
Ask for the record of human contribution while the work is made, not after a dispute starts.
What to ask for depends on the asset. The logo and the signature visual are the ones I’d be most careful about, because they’re what you’ll most want to protect.
| Asset | Ask for | Why |
|---|---|---|
| Logo, mascot, signature visual | A designer’s hand on the file, with the working file and a note of what they drew or reworked | An image that is essentially raw generator output is the case most likely to raise an authorship question |
| Hero and social images | Layered file, tool used, and what was changed by hand | It shows the human contribution and makes later edits possible |
| Copy | The final text with a source for each claim, and who edited it | It supports the claim check and shows who shaped the wording |
| Video and audio | Project file, original recordings and releases for everyone who appears or speaks | It ties the work to real consent and real performers |
Our judgment about useful records, not a legal requirement. Have counsel review the contract wording that backs these up.
The tempting check is a detector. I wouldn’t build a clause on one. OpenAI released a classifier for AI-written text and withdrew it on July 20, 2023, citing its low rate of accuracy. In its own evaluation it correctly flagged about 26% of AI-written text and wrongly flagged about 9% of human-written text.[5] That’s one tool from 2023, so I wouldn’t generalise the numbers. I’d generalise the lesson, which is to check the work itself: claims against their sources, figures recomputed, images against the guidelines, and a random spot-check so that nobody learns when you’re looking.
Before you send the brief, have it read cold. A colleague works, and so does an AI tool on your approved list, as long as you strip out anything amber or red first.
| Act as an experienced freelancer who has just received this creative brief. Read only the AI section. List every question you would need answered before starting work, grouped under: tools, inputs, what stays human, sign-off, revisions, people and rights, proof. Then point to any two lines that could be read in two different ways, and say how. Do not rewrite the brief and do not invent details I haven’t given you. If a line is clear, say so. [PASTE THE AI SECTION OF THE BRIEF HERE, WITH NO CONFIDENTIAL DETAILS] |
Answer the questions it raises before you send the brief, and skip any your contract already covers.
Then score the brief. Give each line 0 (missing), 1 (there but vague) or 2 (specific enough that a supplier could act on it without asking).
| Line | A 2 looks like |
|---|---|
| 1. Tools | Named tools or tool types, the plan, and who approves additions |
| 2. Inputs | Files sorted into buckets, with at least one red example |
| 3. Human only | A list by deliverable of what a person does |
| 4. Sign-off | Named approvers for both sides at each stage, and a response time |
| 5. Revisions | A definition of a round and what doesn’t count as one |
| 6. People and rights | Releases, labels, ownership and a record of human contribution |
| 7. Proof | What the supplier sends with each delivery, and how you’ll spot-check |
Maximum 14. Score the brief as it will be read by a supplier who has never worked with you.
| Score | What it means | What to do |
|---|---|---|
| 12 to 14 | Ready to send | Send it and ask the supplier which line they’d change |
| 8 to 11 | A start | Fix the lowest-scoring lines before you send, beginning with inputs and sign-off |
| 0 to 7 | Not yet an AI section | Write the seven lines first. The one-line “AI may be used with approval” scores 2 or 3 on a good day |
Bands are a Future Factors working guide based on our own judgment, not a benchmark.
To see whether the brief is working, use the lens we apply to any rollout. Use: does every brief that goes out carry the section? Persistence: is it still there on the third project, without a reminder? Impact: did it prevent a late-caught error, a dispute or an uncounted round? Count those, not the number of AI tools in play.
A team can start with the seven lines on its own. Training for the people who write and sign off briefs shortens the learning curve, builds the same standard across the team so the lines don’t change with whoever is busiest, and gives people enough hands-on use to judge what comes back. A good agency will have views on every line, which is the point of sending the brief first. For your next one:
I don’t know of a general rule, and your contract may say something. I’d say so anyway when you’ll feed their drafts or files into a tool, because it affects what they send you and what they’re comfortable sharing. Name the tool and what you won’t upload.
Yes, and the human-only line is the place to do it. Customer quotes, voiceovers and anything with a real person’s likeness are the usual candidates for no AI. Set it per deliverable instead of for the whole project.
Seven lines, about half a page. For a freelancer, one sentence per line is enough. A longer section tends to go unread. Put extra detail in a contract schedule and keep the brief to the seven lines.
Make the brief apply to anyone they bring in, and ask them to confirm that in writing before work starts. Subcontractors may use tools you’ve never heard of, so ask for their tool list too.
No. The brief governs the job, while the contract carries rights, liability and payment. Ask your lawyer to turn the tools, inputs and rights lines into contract wording or a standard schedule, so the two documents agree. This is a judgment about how to organise them, not legal advice.
Sources were read on October 15, 2026: OpenAI’s enterprise privacy page (last updated January 8, 2026), YouTube’s Help page on disclosing altered or synthetic content, the US Copyright Office’s January 2025 report on copyrightability, the Federal Register notice of the Office’s March 2023 registration guidance, and OpenAI’s post on its AI text classifier, including the July 2023 discontinuation note. OpenAI’s privacy statement is a vendor description of its own products and other vendors differ. The Copyright Office material is a US view and copyright questions are decided case by case. Platform labelling rules vary and change, and we checked only YouTube’s. The seven-line AI section, the input buckets, the completeness score and its bands are Future Factors working methods based on our own judgment, not research-validated models or legal standards. The scenes and the sample brief in this guide are illustrations, not client stories. We left out adoption and pricing statistics for agencies using AI because we could not confirm any against a primary source. This is not legal advice.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.