AI Content Workflow for Agencies: From Brief to Approved Draft

An agency AI content workflow runs through six gates: a structured brief, a research pack, an outline the client approves, a draft written inside a shared AI project, an editor's pass with a claims log, and one client approval round. AI speeds up research, outlining and first drafts. People own the brief, the facts and the final sign-off.

Agencies usually get the drafting part working within a week. What goes wrong is the handover between steps: every writer prompting from their own chat history, clients seeing a full draft before agreeing the angle, and nobody knowing which statistics came from the client and which came from the model. The biggest single fix is approving the outline before anyone drafts, because most rewrite rounds are really arguments about the angle, discovered too late.

Follow me on Instagram@sagnikteaches

The six gates, who owns each and what AI does there

Timings are illustrative for a 1,200-word article; adjust them after your first month. The editor in gate 5 should never be the person who wrote the draft, for the same reason nobody proofreads their own work well.

Connect on LinkedInSagnik Bhattacharya

Gate 1: a brief the model can't misread

Vague briefs produce vague drafts, and AI makes that worse because it fills every gap with plausible filler. Use one brief form for every piece. Here it is filled in for an illustrative wine merchant client:

Subscribe on YouTube@codingliquids

CLIENT: independent wine merchant, two shops plus online
PIECE: guide article, 1,200 words, for the website and the monthly email
WORKING TITLE: How to store wine at home without a cellar
READER: customers buying 6-12 bottles at a time, no special storage
ONE THING THE READER SHOULD DO AFTERWARDS: book the free "cellar
  planning" chat in-shop, or buy the 12-bottle rack
MUST INCLUDE: the shop's own storage advice sheet (attached); the rack
  range with current prices (attached); the in-shop service
MUST AVOID: health claims of any kind; suggesting wine improves
  indefinitely; competitor names
PROOF WE CAN USE: storage advice sheet; head buyer available for one
  quote (to be supplied by the client, not written by us)
VOICE: see client voice card v3
DEADLINE: draft 14 Oct, publish 21 Oct

Paste the brief into the client's project and ask the assistant one question before anyone starts: "What is missing or ambiguous in this brief? List questions for the client, most important first." In the example, a useful reply would ask whether the article covers fridge storage for everyday wine, and whether the rack prices include delivery. Five minutes spent answering those saves a rewrite later.

The "must avoid" line matters more for some clients than others. Alcohol, finance, health and food producers all work under advertising rules that vary by market, so the brief should carry the client's own compliance notes rather than leaving the model to guess.

Gates 2 and 3: research, then an outline the client signs off

Research is where models are both fast and dangerous. Use them to summarise sources you've chosen, pull facts from the client's own documents and turn notes into a list of claims, each tagged with where it came from. Don't let them supply facts from memory; a claim with no source is a question, not a fact.

A research-pack request that keeps sources attached looks like this: "From the attached storage advice sheet, rack price list and the two articles I've pasted, list every factual point relevant to the brief. Tag each one [CLIENT DOC], [ARTICLE: title] or [NO SOURCE]. Then list questions the client needs to answer." An illustrative extract of the reply:

• Keep bottles away from direct sunlight and strong light. [CLIENT DOC, point 3]
• Avoid storing wine on top of the fridge because of heat and vibration. [CLIENT DOC, point 4]
• The 12-bottle rack is sold flat-packed and fits under most stairs. [NO SOURCE]
• Question for client: which wines on the current list do you recommend keeping for more than five years?

The third point sounds harmless, and it may even be true, but nothing supports "fits under most stairs". It goes to the client as a question. That tagging habit, set up at gate 2, is what makes the claims log at gate 5 quick to fill in rather than a hunt through the draft.

Then outline. Ask for options, not one answer, so the writer is choosing rather than accepting:

Using the brief and research notes in this project, propose three
different outlines for the article. For each: the angle in one sentence,
H2 headings, and which facts from the notes each section uses. Mark any
section that would need a fact we don't have yet.

An illustrative reply for the wine merchant piece: option A is a room-by-room guide (kitchen, hallway cupboard, under the stairs); option B is built around the four enemies of stored wine (heat, light, movement, dry air); option C is a decision guide by how long you plan to keep bottles. Option C flags two missing facts: which wines on the shop's list are meant for keeping, and for roughly how long. That flag is exactly what the account manager takes to the client.

What you'd fix: option B reads well but repeats advice every storage article already gives. Option C is the distinctive one and ties naturally to the in-shop service. The client approves option C with one edit, and only then does anyone draft. When outline approval became a fixed step in this kind of workflow, the typical before-and-after is two or three rewrite rounds becoming one light round, because the argument about the angle has already happened.

Gate 4: drafting inside a shared project, not a personal chat

Every client should have one shared project in the agency's business AI plan, holding the brief, the voice card, the product or service facts, approved past pieces and the research pack. ChatGPT Business and Claude Team both offer shareable projects; Gemini has Gems (which become skills from November 2026). Don't build custom GPTs for this: OpenAI is retiring custom GPTs, and existing ones stop running on 11 December 2026, with a migration route into plugins and skills.

The drafting prompt is short because the project holds the context:

Write the article from the approved outline (option C), following the
voice card. Use only facts from the research notes and attached client
documents. Where you need a fact we don't have, write [CHECK: what's
needed] rather than filling it in. Leave the head buyer's quote as
[QUOTE FROM CLIENT].

Keeping everything in the project means any writer can pick up any client, the voice stays consistent, and you can see what the model was given if a claim is ever questioned. Keeping voice steady across a team is its own discipline, covered in how agencies keep brand voice consistent.

Check that the placeholders survived. Models dislike leaving gaps, and the usual failure is a missing fact papered over with vague authority instead of a [CHECK] marker. An illustrative line from a first draft:

"Experts generally agree that most everyday wines are best enjoyed within two to three years, while fine wines can improve for decades."

Nothing in the research pack says that. "Experts generally agree" is the tell: it's how a model sounds when it's supplying a fact from memory. The editor swaps it for the marker the prompt asked for, "[CHECK: which wines on our list are for keeping, and for roughly how long]", and it goes to the client with the other questions. Searching each draft for "experts", "studies" and "research shows" catches most of these before the claims log does.

Gate 5: the editor's pass and the claims log

The editor does two jobs: makes the piece good, and makes it defensible. The second job uses a claims log, a simple table of every factual statement in the draft, where it came from and its status. A filled-in extract for the wine piece:

That last row is the realistic mistake worth planning for. Models like to add a credibility line, and a precise-sounding number about the client's own team is the most tempting kind. Nobody at the merchant said it. Without the log, it would have gone to the client, who might well have approved it without noticing. The general method for this kind of checking is in fact-checking AI marketing copy.

The editor also reads for voice and rhythm: cut the stock phrases, vary sentence length, and make sure the one action in the brief is clear by the end.

Gate 6: client approval in one round

Send the draft with the claims marked "client to confirm" highlighted and a short covering note that says what you need: confirmation of the flagged claims, the supplied quote, and any changes. When feedback comes back scattered across email and comments, paste it into the project and ask for a change list:

Turn this client feedback into a numbered change list. For each item:
the exact change, where in the draft, and whether it conflicts with the
approved brief or outline. Flag conflicts for the account manager.

An illustrative change list built from one client email and three document comments:

1. Intro: replace "cellar" with "wine rack" in the first line
   (comment 1). No conflict.
2. Section 2: add that the shop stores customers' cases free
   of charge (email). CHECK: current service? Not in the
   research pack.
3. New section on sparkling wine (email). CONFLICT: not in
   the approved outline.
4. Head buyer's quote supplied (email). Insert at
   [QUOTE FROM CLIENT].

Item 2 is the one to catch. A new factual claim arriving through feedback skips gates 2 and 5 unless someone logs it, so it goes into the claims log with the client's email as its source.

Conflict flags are the useful part. If the client now wants a section on sparkling wine that wasn't in the approved outline, the account manager can decide whether that's a tweak or a new piece, rather than the writer quietly absorbing it. Record what the client approved and when; the structure for that record is set out in a draft, check and sign-off workflow. Contract wording about AI use and approvals belongs in your agency's contract clauses, not in each email.

Two jobs that don't fit the six gates neatly

The first is the client who sends their own AI-written draft to "just polish". It's tempting to start at gate 5 and edit. Start at gate 1 instead, in a fast version: write the one-page brief from the draft itself (who it's for, what the reader should do next) and send it back for a yes. Then run the claims log on their draft before touching the prose. A draft a café owner generated about a new brunch menu might say the sourdough is "slow-fermented for 48 hours" and the eggs come "from our own hens". If neither is true, polishing the sentences first just makes a false claim read better. The claims log takes about 20 minutes and tells you how much of the draft is usable.

The second is short-form work, which compresses the gates rather than skipping them. For twelve social posts a month for a dental practice, gate 1 becomes a monthly brief (themes, dates, the one offer, and what not to say, such as results claims the practice can't support). Gate 3 becomes a one-line angle per post, approved by the practice manager in a single reply. Gates 4 to 6 run on the whole batch at once, and the claims log shrinks to a "claim / source" column in the content calendar. The order doesn't change: angles approved before drafting, facts checked before the client sees the batch.

What the workflow costs and saves on a four-article retainer

An illustrative before-and-after for one client receiving four 1,200-word articles a month:

• Before: about 6 hours per article (research 1.5, drafting 3, editing 1, revisions 0.5 across two or three rounds), so 24 hours a month.
• After: about 3.5 hours per article (brief check 0.25, research pack 0.75, outline 0.5, draft 1, edit and claims log 0.75, one revision round 0.25), so 14 hours a month.

Ten hours back per client per month is plausible, but notice where the hours went: drafting shrank, while the editor's pass and the claims log got more time, not less. An agency that cuts the edit to bank the savings ends up publishing the invented "10,000 wines" line sooner or later. Tool costs are modest: business AI seats run $25 per user per month on monthly billing for ChatGPT Business Standard or Claude Team Standard ($20 billed annually), so a four-person content team pays around $80 to $100 a month.

Two numbers show whether the workflow is working, and both come from records you'll already keep: revision rounds per article (from the approval record) and invented claims per month (from the claims logs). If a client's articles averaged two and a half rounds before and just over one after three months, the outline gate is doing its job. If invented claims stay at two or three a month, look at the drafting prompt and the project files, because the editor is catching what the set-up should have prevented.

Where agency content pipelines break

• Skipping gate 3 when busy. The week you skip outline approval is the week the client rejects the angle. Keep it even for short pieces; a five-line outline in an email is enough. In the retainer's own numbers, the outline takes half an hour, while a rejected angle costs the draft (an hour), the edit (45 minutes) and another round of client time before the second attempt even starts.
• Stale project files. Last season's price list in the project produces this season's wrong prices. Date every file and review each client project monthly.
• Writers drifting back to personal chats. It feels faster. It isn't, because the next writer can't see what the model was told. Make the shared project the only approved place for client work.
• The claims log becoming a tick-box. If every row says "supported", someone isn't reading. Spot-check one article per client per month against its sources.

Further reads

• How to Build a Brand Voice Guide That AI Can Follow — Write the voice guide every client project should hold.
• How Much Does Human Editing of AI Content Cost? — Price the editor's pass properly in your retainers.
• How to Write Website Articles With AI That Still Rank on Google — Make the drafts earn search traffic, not just approval.
• How to Check AI Content for Plagiarism Before You Publish — A pre-publish check to add after client sign-off.
• How to Write an AI Content Policy for Your Marketing — Put the agency's AI rules in writing for staff.
• Hire an AI Specialist or Upskill Your Agency Team? — Decide who on the team owns this workflow.
• AI Rollout Plan for a Ten-Person Marketing Agency — A ten-person agency's 12-week AI rollout, week by week: platform choice, client rules, two pilots, the numbers after a quarter and what went wrong.
• Which Agency Tasks AI Handles Well, and Which It Doesn't — Sort agency work by evidence, review effort and consequences, then test a reporting workflow with a clear human sign-off.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI's custom GPT retirement and migration FAQ; ChatGPT Business and Claude Team pricing pages; vendor documentation for shared Projects on business plans.