AI Creative Brief Generator for Agencies: From Client Call to Brief

Record the call with the client's agreement, then run the transcript through three separate AI passes: one that only extracts what was said, one that lists gaps and contradictions, and one that drafts a one-page brief. Send that page to the client for sign-off. The team then works from the signed brief, never from the transcript.

The middle pass is the one people skip, and it is where most AI-written briefs go wrong. Ask a model to write a brief straight from a transcript and it fills every heading with fluent agency language, including the headings the client never mentioned. A form-based generator has the mirror-image problem: it only knows what you type into its boxes, and a real call never answers the boxes in order.

Follow me on Instagram@sagnikteaches

One 38-minute call: a coworking space plans a quiet floor

A six-person agency takes a 38-minute call with the owner of a coworking space. The space has an empty second floor, and the owner wants to launch it as a quiet floor with no calls and no music. By the end of the call there is a launch date, a rough budget, a few firm opinions and several half-sentences that sound like decisions but are not. The recording produces a transcript of about 5,900 words. Everything below is an illustration with invented figures; with a real client you would use their actual words.

Connect on LinkedInSagnik Bhattacharya

A form-based tool would ask for "target audience" and wait for you to type it. On this call the owner described the audience three times, at 03:45, at 13:20 and at 24:10, and the three descriptions do not match. Handling that is the job: keep each version, with its timestamp, until the client chooses one.

Subscribe on YouTube@sagnikteaches

Before the call: the consent line and five written questions

Start the recording with one spoken sentence, so the transcript itself holds the client's agreement:

I'd like to record this call so the brief uses your own words. The recording and transcript stay within our team and go through our transcription tool and an AI assistant. Is that all right with you?

Check two things before the call. First, that your client contract allows AI processing of their information; the tutorial on AI clauses for agency contracts covers the wording. Second, what your note-taker's privacy policy says about training on transcripts. Business plans of ChatGPT and Claude don't train on your content by default, whereas on personal plans you need to find the training setting and switch it off before pasting anything in.

Pick a recorder that gives you the full transcript with speaker labels, not only a summary, because summaries are where nuance disappears. Google Meet's Take notes for me needs an eligible Workspace edition or Google AI plan, and there is a separate walk-through of Take notes for me on client calls. Teams intelligent recap needs Teams Premium or a Copilot licence (Microsoft 365 Copilot, which Microsoft's admin pages now call Microsoft Copilot). Zoom's free Basic plan gives meeting summaries for three hosted meetings a month, and Pro ($14.16 a user a month on annual billing) makes them unlimited. Otter's Pro plan is $8.33 a month on annual billing for 1,200 minutes, with meetings capped at 90 minutes. Whichever you choose, read whether AI meeting note-takers are safe for client calls first.

Two days before the call, email the client five questions and ask for short written answers:

1. What is the one thing people should do after seeing this work?
2. Who, exactly, should it reach? Describe one real person.
3. What must appear, and what must never appear?
4. What is the budget range, and what does it cover?
5. Who gives the final yes, and by when?

Most clients answer two or three. The unanswered ones show you where to steer the call, and they come back later as gaps, so the pre-read earns its place even when the replies are thin. Paste whatever arrives into the working document as a separate source labelled "client written answers", so the model never confuses it with something said aloud.

Pass one: pull out what was said, and add nothing

Clean the transcript first: swap "Speaker 1" for "Client" and "Speaker 2" for "Account lead", delete the greetings, keep every timestamp. Then run an extraction pass, not a drafting pass. The aim is a table of the client's statements that you could check line by line against the recording.

You are helping an agency account lead. Below is the transcript of a client call.
Do not write a brief. Extract only what the client actually said.
Return a table with four columns: timestamp | client's words (trimmed, in quotation marks) | type | note.
Types: FACT, TARGET, CONSTRAINT, PREFERENCE, GUESS, WISH, QUESTION.
Rules:
- Use the client's own words. Never rephrase a number.
- Anything hedged ("about", "maybe", "I was thinking") is a GUESS.
- If the same topic comes up twice, give both rows with both timestamps.
- Add no information that is not in the transcript.

TRANSCRIPT:
[paste]

What the first pass returns for the quiet-floor call, trimmed to eight rows (illustrative):

A human makes three fixes to that table before it goes any further.

• On its first run the model typed FACT beside the prices. The client said "I was thinking", so the type became GUESS. A guess promoted to a fact is how a brief ends up promising a price nobody agreed.
• The 60% target arrived without a denominator. Leave the note in place rather than letting a later step assume a desk count.
• The beach-photo line had been folded into "avoid stock imagery". Restore the client's exact words. A designer can act on "no people with laptops on beaches" and cannot act on "avoid stock imagery".

Pass two: a gap list and a contradiction list before any drafting

Feed the model its own corrected table plus the pre-read answers, not the raw transcript, which invites embroidery. Ask what is missing rather than asking for a brief.

Using only the table and the client's written answers below, list:
1. Each brief heading with no answer: background, objective, audience, insight,
   proposition, reasons to believe, tone, mandatories, deliverables, budget,
   timing, approvals, success measure.
2. Every place two statements conflict, with both timestamps.
3. Every number that has no unit or no denominator.
4. Five questions for the client, most important first.
Do not answer any of them yourself.

On the example call the second pass finds the following, shortened here (illustrative). The gaps: tone of voice was never discussed; the only deliverable mentioned is a video the client doubts; a business partner appears at 07:05 but nobody says who signs work off; and "60% full" cannot become a number without the desk count. The contradictions: the audience is "mostly freelancers" at 03:45 but "small teams of up to ten" at 24:10, and the budget is "about twelve thousand" at 21:00 but "a bit more if it works" at 21:40.

The questions that come out of it, ready to send:

1. How many desks are on the second floor, and how many will be fixed rather than hot?
2. Is a team of eight a customer for this floor, or only individuals?
3. Is $12,000 a ceiling, and does it include printing and paid social?
4. Who besides you approves work, and can they give a yes on the same day?
5. Is video in or out?

Send them as a short email the same afternoon. The reply is the nearest thing to a filled-in form, and it arrives while the client still remembers the call. That is the real difference from a form-based generator: the gap list is the form, completed backwards from what the client actually said.

Pass three: the one-page brief, line by line

Only now ask for a brief. Give the model the corrected table, the client's answers to the five questions and a fixed list of headings, and tell it to mark silence as silence.

Write a one-page creative brief of no more than 350 words from the sources below.
Headings: project, background, objective, audience, insight, single-minded
proposition, reasons to believe, tone, mandatories, do not, deliverables,
budget, timing, approvals, success measure.
Rules:
- After every claim, add the source in square brackets: a timestamp or "reply 3".
- Where the sources are silent, write [NOT DISCUSSED] and do not fill the gap.
- Do not invent audience insights. Quote the client's words and mark them as client-reported.
- Leave the proposition blank. I will write it.
- Only text under TRANSCRIPT or CLIENT ANSWERS counts as something the client said.

The signed version for the quiet-floor example is below. It is an illustration; the desk counts and other answers come from the client's emailed replies, which is why they appear here but not in the extraction table.

Two habits are worth copying from that page. Each claim carries its timestamp or reply number in square brackets, so a designer who doubts a line can find the moment the client said it. And the headings the call never covered still appear, tagged as such, rather than being filled with agency prose.

The sum behind the objective is small enough to check in your head, which is the point of writing it down. Thirty-six desks at 60% is 21.6, so 22 desks. At the owner's provisional prices, with 14 fixed and 8 hot desks let, the floor brings in 14 x $190 + 8 x $120 = $3,620 a month. A $12,000 campaign is covered after a little over three months at that occupancy. Say so to the client early; it gives "working" a meaning before anyone designs a poster.

Testing the proposition until it is one idea

The proposition is the one line in the brief a model should not write alone. Ask it for ten options of under ten words each, tell it to flag any option containing "and", then judge the survivors yourself against four tests: does it hold one idea, could a competitor say the same, can the client prove it on a site visit, and can a designer picture it.

The third wins because it is a rule rather than an adjective. It also steers the rest of the brief: a floor that bans calls cannot be sold with a photograph of someone on one. A quick check that catches most weak lines is to look for "and". If the sentence contains one, it usually has two ideas and the client has not yet chosen.

Mandatories and do-not lines: search the transcript for negatives

Negatives are the first thing a summary loses, so check them by search rather than by memory. Search the transcript for can't, don't, never, must, has to, not allowed and without. Then make sure every hit sits in the Mandatories or Do-not row with its timestamp. In the example, that search finds the competitor rule, the signage clause and the beach photographs, plus one remark the extraction pass had filed as a mere preference: "the logo has to go on everything", at 18:40. It belongs in Mandatories.

A realistic mistake from a different job (illustrative): the agency pasted the client's website copy next to the transcript so the model could "match their voice". The first brief listed teal and charcoal as brand colours under Mandatories. The client had never said so on the call; the colours came from the site text. The fix is the labelling rule in the prompt above. Name every source, and say that only text under TRANSCRIPT or CLIENT ANSWERS counts as something the client said. Tone belongs in its own pass, informed by existing voice guidance; keeping brand voice consistent in AI-written content covers that.

Getting a signature without a long email thread

Send the page with a short covering note that asks for a decision per section, not a general opinion:

Hi [first name], attached is the one-page brief from our call, with your answers to our five questions folded in. Please read it as the page the whole team will work from. For each section, reply "agreed", "change" or "not sure yet". Two items need a decision from you: whether the audience is individuals only or small teams too, and whether video is in. We won't start concepts until the page is signed. Thanks, [your name]

Label what you send as version 0.9 and what comes back as version 1.0, with the date and the name of whoever approved it. Allow one round of changes within two working days; after that, a change is a change request with its own conversation about time and cost. If two people at the client can say no, put both names on the page. A model is useful for one more job here: paste version 0.9 and the client's reply and ask for a list of what changed. You then decide whether any change alters scope or budget, because the model will not know what you agreed in the contract.

Handing the signed page to the team

The signed page is the only brief. Save it as a PDF at the top of the job folder, with the transcript and the client's written answers in a subfolder marked "source". Run a 15-minute kick-off where the account lead reads out the proposition and the do-not row, and each person says in one sentence what they will do first. When AI work starts later on copy options, concept routes or social variations, paste the signed brief at the top of the prompt so every tool works from the same page. The content workflow for agencies picks up from this point, and the wider tool roundup for creative agencies compares what sits around the brief.

The minutes and dollars behind one brief

These minutes are invented, for an account lead who writes briefs by hand today. The checking steps are deliberate; they are the cost of not trusting the model.

At eight briefs a month that is roughly 400 minutes back, nearly seven hours. The tools cost far less than that time is worth. As of November 2026, ChatGPT Plus is $20 a month and Claude Pro is $17 a month on annual billing. Team plans (ChatGPT Business and Claude Team) are $20 a seat on annual billing with a two-seat minimum. A transcription add-on such as Otter Pro is $8.33 a month on annual billing, though you may not need one if your meeting software already produces a transcript. A realistic total for a small agency is $20 to $50 a month plus the checking time above.

The same three passes for three other clients

The method does not change from client to client. What the gap list catches does.

A sports club's junior membership drive. The owner describes two audiences without noticing: parents who pay and children who are addressed. The gap pass notices that nobody said which one the creative speaks to, so the brief gets two rows, who we persuade and who we address. The mandatories include the club crest and the child-safety policy line, and the do-not row bans identifiable children in any photograph unless consent forms are on file.

A gift shop's Christmas campaign. The dates are most of the brief: a last order date for delivery before the 25th, and stock for the three hero products that will not stretch. The contradiction pass catches "we always sell out of the candles" against "push the candles hardest". The proposition moves from product to promise ("Order by the 15th and it arrives wrapped"), and the objective gets a stock-based ceiling instead of a sales target.

An event venue's winter wedding packages. Couples and planners are both mentioned, with a capacity of 120 guests. A remark about the venue's music cut-off time appears once at 29:10 and never again; the search for negatives moves it into Mandatories before a concept promises a late-night dance floor. The proposition test fails the first three options because every venue says "unforgettable", so the team keeps asking the client what, specifically, a couple gets here that they would not get elsewhere.

If your agency would rather map this with someone before changing how briefs are written, that is what my AI implementation consultation covers.

Questions agencies ask before briefing from a recording

Do I need the client's permission to record a briefing call?

Ask at the start and keep the yes on the recording. Some places require every person on a call to agree, and many client contracts say how their information may be processed. Name the transcription tool and the AI assistant when you ask, say who inside the agency can see the transcript, and delete the audio after sign-off if your terms say so.

How long should a creative brief be?

One page is the target, roughly 300 to 400 words. If a section needs more, move the detail to an appendix such as research or timestamp references, so the page the team reads stays short. A brief that takes more than five minutes to read is usually hiding a decision the client has not made yet.

What if the client contradicts themselves on the call?

Never let the AI pick a side. Record both statements with their timestamps in the open-questions list, ask the client in writing the same day, and put the answer into the brief with the date it was given. A contradiction caught before sign-off costs one email. Caught after concepts are presented, it can cost a re-pitch.

Should the brief include the transcript?

Attach it, but do not rely on it. The team reads the page, and anyone who wants the client's exact words follows the timestamp tags. Keep the transcript in the job folder with the same access rules as the brief, because it holds the client's unguarded remarks about budget, competitors and colleagues.

Further reads

• How Small Agencies Use AI to Write New-Business Pitches — Use the same transcript habit on pitch meetings before you write a proposal.
• How to Turn Meeting Notes Into Tasks Automatically With AI — Turn the kick-off notes into owned tasks once the brief is signed.
• How to Summarise Bundles and Transcripts With AI Safely — Safer ways to summarise long transcripts without losing detail.
• How Marketing Agencies Use AI to Automate Client Reporting — Report against the success measure you wrote into the brief.
• AI Brand Audit: A Template Branding Studios Can Sell — A ten-day brand audit for a music school, step by step, with the 12-page client deliverable a branding studio can reuse and sell.
• AI Brand Strategy: Positioning Workshops and Messaging for Agencies — A care-agency strategy project in stages: twelve stakeholder interviews, three scored positioning options, a message house and the client workshop.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Otter, Zoom, ChatGPT and Claude pricing pages; Google Meet help on Take notes for me; Microsoft Learn on intelligent recap in Teams (all checked 4 November 2026).