ai:content
The AI content workflow for agencies: brief to draft to QA.
By Loan Laux · September 3, 2026 · 8 min read
An AI content workflow is the production line that turns a brief into a shippable piece: structured intake, retrieval-grounded drafting, a human editorial pass, and a QA gate that checks facts and voice before anything gets published. The point is not to let AI write your content. The point is to let AI do the assembly (research digests, first drafts, reformatting, repurposing) so your editors spend their hours on the judgment clients actually hired you for.
This is the workflow we build inside agencies when our AI engineers embed with a content team. Everything below runs on tools you likely already pay for; the hard part is the discipline, not the software.
Why 'let AI write it' fails.
The naive version of this workflow is a one-line prompt and a publish button, and it fails twice. It fails editorially, because a blank-prompt model produces the median of the internet: fluent, plausible, and interchangeable with what every competitor's tool wrote the same week. And it fails commercially, because clients pay agencies precisely for the parts the blank prompt skips: a point of view, real sourcing, and a voice that is recognizably theirs.
Search does not save you either. Google's guidance is that it rewards high-quality content however it is produced, and its scaled content abuse policy targets mass-produced pages that add little value, whatever tool made them. That cuts both ways: AI-assisted content is explicitly fine, and undifferentiated AI volume plays are exactly what the spam policies describe.
So the workflow worth building is different in kind, not in degree. AI compresses the production; a smaller amount of concentrated human attention holds the standard. The rest of this post is that line, stage by stage.
The AI content workflow, stage by stage.
01:
Start from a structured brief
A fixed form, not a Slack message: audience, angle, the argument in one sentence, claims to make, sources to use, target keyword if there is one, and what the piece should cause the reader to do. Ten minutes from the strategist here saves an hour of editing later.
02:
Assemble the source pack
Gather the raw material the draft will be grounded in: interview notes, the client's data, past pieces on the topic, the voice guide. If a piece has no proprietary input at all, question whether it should exist; that piece is the one generic AI blogspam already covers.
03:
Draft with retrieval, not a blank prompt
The model writes from the brief plus the source pack, in the client's voice profile, section by section. One long grounded prompt beats twenty short clever ones, and it makes the output reviewable because every claim has a visible source.
04:
Editorial pass by a human who can reject it
An editor rewrites the argument where it is soft, cuts anything a generic tool could have written, and adds the specifics only your team knows. The editor must have the authority to kill the piece; a pass that can only approve is a rubber stamp.
05:
Run the QA gate
A fixed checklist before anything ships: every factual claim traced to a source, quotes verified, voice checked against the guide, links tested. Detailed below, because it is the stage agencies skip first.
06:
Publish, then repurpose
Once a piece passes QA it is trusted material, so derivatives (newsletter cut, social posts, a deck slide) can be generated cheaply from it. Repurpose after the gate, never before, or you multiply errors instead of assets.
Encode the editorial standard so the model can follow it.
Most agencies keep their editorial standard in their editors' heads, which is exactly where a model cannot reach it. Writing it down is the highest-value work in this whole system, and it is also the piece that survives staff turnover.
Concretely: a voice guide per client (words they use, words they would never use, three annotated examples of 'sounds like them'), a house style file, and the QA checklist as an actual document. If your team runs on Claude, Claude Skills are the natural container, because the voice guide loads automatically whenever someone drafts for that client instead of living in a doc nobody remembers to paste in. The same material doubles as an SOP for onboarding human writers, which is a good tell that you have written a real standard rather than a prompt hack.
What AI does and what the editor owns.
The line gets drawn stage by stage, not as a blanket 'AI drafts, humans review'. Our default:
| Stage | AI does | Editor owns |
|---|---|---|
| Brief | Suggest angles from past performance and the source pack | The angle, the argument, and whether the piece is worth making |
| Research | Digest sources, extract claims, flag gaps | Which sources are trustworthy enough to build on |
| Draft | Full first draft, grounded in the source pack and voice guide | Nothing yet; let it draft |
| Edit | Apply the mechanical fixes the editor requests | The rewrite, the cuts, and the specifics that make it yours |
| QA | First-pass checks on links, claims, and style | Final sign-off, and every fact published under a client's name |
| Repurpose | Cut-downs and reformats of the approved piece | Which channels actually deserve it |
The QA gate.
This is the stage that makes the rest safe to speed up. Ours has four checks, run in order:
- Claims check. Every concrete factual claim gets traced to a source the editor would cite by name. Models fabricate confidently, and a fabricated statistic under a client's logo is their problem first and yours forever. Anything unverifiable gets cut, not softened.
- Voice check. Read three paragraphs against the client's voice guide with the logo covered. If the editor cannot tell whose piece it is, it goes back.
- Sameness check. Search the piece's core argument. If the first page of results already says it, the piece needs the client's data, experience, or a contrary position added, or it does not ship.
- Mechanical check. Links resolve, names are spelled right, numbers add up, image rights are cleared. The model can run this pass first; a human signs it off.
Where this breaks.
The failure modes we see most often, all avoidable:
- Skipping the brief. 'Write a post about X' produces content-shaped filler, and the editorial pass balloons into a rewrite. The workflow's speed comes from the front of the line, not the back.
- A thin source pack. Retrieval from nothing is a blank prompt with extra steps. If the pack is empty, do the interview or pull the data before anyone drafts.
- Volume as the goal. The workflow makes ten posts a week possible, which is exactly the trap. Publish at the rate your QA gate genuinely runs, and spend the surplus capacity on depth.
- The editor as formality. If the editorial pass has never killed a piece, it is not a pass. Track the rejection rate; zero means the gate is standing open.
Do it yourself, or bring us in.
Everything above runs on a normal agency tool stack: a team-plan assistant that can retrieve from your own documents, a form for the brief, and a written editorial standard. The build is roughly two weeks of a capable person's attention. The maintenance is editorial discipline, which no tool supplies.
Where agencies bring us in: our AI engineers embed with the content team, wire up the retrieval and the per-client voice profiles, and run the first production cycles alongside your editors until the gate holds without us. The deliverable is the working line and the documented standard, not a dependency on out:grow.
Frequently asked questions.
Will Google penalize AI-assisted content?
Not for being AI-assisted. Google's published guidance says it rewards high-quality content however it is produced, and its scaled content abuse policy targets mass-produced pages that add little value regardless of how they were made. The risk is not the tool, it is publishing undifferentiated volume. A workflow with real sourcing and an editorial gate sits on the right side of both.
How much time does this actually save?
In our builds, a piece that took a writer two days lands at roughly a half-day of concentrated human time: the brief, the editorial rewrite, and QA. Drafting and reformatting stop consuming anyone. The honest caveat is that the first few weeks are slower, because you are writing voice guides and tuning retrieval while still shipping.
What tools do we need for this?
Fewer than the market suggests. A team-plan assistant that can ground drafts in your own documents, a form tool for briefs, and your existing publishing setup cover it; specialized AI writing suites mostly resell the workflow this post describes. Our stack guide has the specifics.
How do we keep ten clients' voices from blurring together?
One voice profile per client, stored where the model loads it automatically rather than pasted in ad hoc. Claude Skills or equivalent project instructions both work. The blur happens when writers share one generic 'professional but friendly' prompt across accounts. The voice check in QA is the backstop: logo off, guess the client.
Should we tell clients AI is involved?
Yes, framed accurately: AI drafts and assembles, your editors own every word that ships, and the QA gate is documented. Most clients already use these tools themselves and mainly want to know their content is not unreviewed model output. The awkward conversation is the one where they find out you hid it.