Every client engagement at Choco Media starts with a brief. Not a vague one-pager someone filled out in ten minutes, but a structured intake document that captures the real problem — the business context, the audience, the constraints, and the outcome the client actually needs. For a small AI-first agency, the brief is where the work either lands well or goes sideways. And over the past year, our ai agency workflow for brief-building has changed more than almost anything else we do.
This post is a behind-the-scenes look at how we use AI at the brief stage — what it handles, what we deliberately keep human, and the quality checks we run before anything leaves our desk. If you work with clients and use AI in your process, some of this will be familiar. But the specific decisions about where to put the human layer might be worth your time.
This is written for agency people, in-house teams managing external partners, and anyone who’s tried to use AI to accelerate briefing and ended up with output that felt slightly off. We’ll walk through the full intake-to-first-draft process — what happens at each stage, which tools we use, and where we’ve learned to slow down.
Why the brief matters more than the output
There’s a temptation, especially once you’ve seen how fast AI can produce a draft, to treat the brief as a formality. You get a few inputs, you prompt, you get a first draft. The problem is that AI is very good at producing fluent, confident text from underspecified inputs. It will give you something that looks like the answer even when the question wasn’t precise enough.
In client work, this shows up in a specific way: the first draft reads well, but it answers the wrong problem. The tone is slightly off, the framing doesn’t match how the client talks about their product, or it’s aimed at the wrong audience. You send it, the client says it’s not quite right, and you spend three rounds of revisions recovering ground you lost in the briefing.
- A weak brief produces plausible output, not correct output
- AI amplifies briefing quality in both directions — good brief, better output; weak brief, faster wrong output
- The revision cost of a bad brief almost always exceeds the time saved by skipping it
The brief isn’t a pre-step to the work. It is the work. Getting it right is where you earn the client relationship, not where you spend time before you can start earning it.
Our intake structure: what we capture before AI touches anything
Before we use any AI tool, we do a human intake. This usually takes 20–40 minutes and covers eight areas. We run this as a structured conversation rather than a form, because clients answer differently when they’re talking than when they’re filling out a document.
The eight intake areas
- Business context — what does the client sell, who do they sell it to, and what’s the competitive environment they’re operating in
- Campaign or content objective — what does success look like in concrete terms, not just “more awareness”
- Audience — the specific person this needs to reach, with enough specificity to write to them, not at them
- Tone and voice — how the client talks about their brand, and what they want to avoid saying
- Existing assets — what copy, positioning, or creative already exists that this needs to be consistent with
- Constraints — channel, format, length, legal or brand restrictions
- The problem behind the problem — what the client is really trying to fix, which is often different from what they initially described
- Definition of done — what needs to be true for the client to approve this without revisions
The last two are the ones that save the most revision time. They’re also the ones most likely to get skipped when intake feels like admin rather than strategy.
Where AI enters the process
Once we have clean intake notes, AI gets involved in three places: structuring the brief document, generating the first draft, and running a consistency check on output before it leaves our hands.
Structuring the brief
Our intake notes are messy — they’re written in real time during a conversation. We use AI to turn them into a structured brief document. This is where it earns its keep most clearly. Given good raw notes, it can produce a well-organized brief in under two minutes that would have taken 20 to write manually.
The prompt we use is explicit about what goes in each section and asks the model to flag anything that looks underspecified rather than fill in the gaps. That last instruction matters — a model left to its own devices will invent plausible-sounding detail rather than surface a gap, and invented detail is worse than a gap because you don’t know it’s there.
Generating the first draft
With the structured brief as input, we generate a first draft. We use Claude for most of this, with a system prompt that carries the client’s tone and voice parameters. For clients we’ve worked with before, we have accumulated examples of approved copy that we include as reference. The model is instructed to produce output that a human editor could approve with minimal changes, not a raw draft that needs heavy reworking.
The goal isn’t to produce a final draft with AI. The goal is to produce a draft good enough that the human editor’s job is judgment, not construction.
In practice, a first draft from a well-constructed brief gets approved with light edits about 60–70% of the time for content work, and about 40% of the time for brand copy where tone variance matters more.
What we don’t let AI touch
This is the part of the process that’s probably changed the most as we’ve learned what breaks. There are three areas where we’ve decided the human layer isn’t optional.
The “problem behind the problem”
When a client says they want a landing page to drive sign-ups, that’s the surface request. The actual problem might be that their current page converts at 1.2% when industry average is 3%, or that their sales team says leads from the page are consistently poorly qualified. These are different problems with different solutions. AI can process whatever you give it, but it can’t surface a problem the client didn’t articulate. That requires a human who knows enough to ask the right follow-up question.
Tone calibration for new clients
For the first piece of work with any new client, we don’t use AI for the draft. We write the first version by hand. This gives us a baseline we can use to build a tone reference for everything that comes after. The time cost is real, but the alternative — using AI on a new client with no reference material — produces output that’s technically competent but tonally generic. Clients can usually feel it even if they can’t articulate why.
Final approval before client delivery
Everything goes through a human read before it leaves our hands. Not a skim — a read for logic, tone, accuracy, and any claim that needs to be verifiable. AI doesn’t fact-check, and it doesn’t catch a sentence that’s technically correct but sounds odd for this client’s brand. A human read at the end is not optional. We budget 20–30 minutes for it on most content pieces.
The quality checks we run
We run three checks before delivery, in order.
1. Brief-to-output alignment
Does the draft actually answer the brief? This sounds obvious but it’s the most common failure mode. AI can drift — especially on longer pieces — toward what it finds interesting or what the training data suggests is typical, rather than what the brief specified. We check the objective, audience, tone, and constraints explicitly against the draft.
- Does paragraph one reach the right audience?
- Is the tone consistent throughout, or does it shift mid-piece?
- Are the constraints respected — length, channel format, any restrictions?
- Does the CTA match what the brief said success looks like?
2. Brand voice review
We keep a short list of things each client never says and things they always say. AI trained on general data will occasionally reach for generic marketing language — “end-to-end solution”, “world-class”, “seamless experience” — even when the brief said to avoid it. The brand voice review is a fast scan for these. For clients using our AI content creation retainer, we maintain a running list of flagged phrases that gets updated after every delivery.
3. The cold-read test
We ask a team member who wasn’t involved in the brief to read the draft cold and answer two questions: what does this say, and who is it for? If their answers match the brief, the piece is probably clear enough. If they’re off, we have a communication problem in the draft that a client will also notice.
The tools in our current brief-to-draft stack
We keep this deliberately simple. Complexity in tooling usually means someone bypasses part of the process when they’re in a rush.
- Intake notes — written in Notion during the client call, structured with a template that maps to the eight intake areas
- Brief structuring — Claude, with a fixed prompt we’ve refined over six months of client work
- First draft generation — Claude for most content types; GPT-4o for specific formats where we’ve found it performs better (shorter ad copy in particular)
- Brand voice check — a custom GPT trained on client-approved examples and flagged phrases, running as a Notion integration
- Final human review — no tool. One person, a full read, notes in the doc
The stack cost per client engagement at the brief-and-draft stage is under €5. The time saving versus a fully manual process is roughly 60–70% on first drafts. The quality ceiling is set by the human layer, not the tools.
What we’d change if we were building this from scratch
The biggest thing we underinvested in early was the tone reference system. We spent a lot of time on prompt design and not enough time building per-client voice documentation. A model given a well-structured brief and a strong tone reference produces dramatically better output than the same model given a brief with no reference. If we were starting over, we’d build the client voice library in week one of every engagement, not after the first round of revisions.
We’d also have been more disciplined earlier about the no-AI-on-first-draft rule for new clients. We bent it a few times under time pressure. It cost us revision rounds we didn’t have budget for. The rule exists for a reason.
- Build tone references from day one of every client relationship
- Don’t use AI on the first piece for a new client — write it by hand and use it as reference
- Build the quality checks into the workflow, not as an optional last step
- Keep the tooling simple enough that no one bypasses it
How this connects to our broader AI automation practice
The brief-to-draft workflow is one layer in a larger set of processes we’ve built around AI-assisted client work. Our AI automation work covers everything from content production to reporting to lead qualification — and in each case, the same principle applies: AI accelerates the work that’s already well-specified, and surfaces problems in everything that isn’t.
The brief is where specification happens. Get it right, and AI is a significant force multiplier. Get it wrong, and the same tools produce confident, fast, wrong output that you then have to manage with the client.
If you’re building a similar workflow and want to compare notes, or if you’re considering how to bring structured AI processes into your agency, the contact page is the right starting point. We’re a small team in Rovaniemi, and we’re reasonably open about how we work.