AI content operations: human-led, AI-assisted, no slop
AI makes research faster, structure easier and first drafts cheap. It does not know what is true or what your customers need. We build a workflow where machines do the tedious work and experts make every decision that matters.
- 2-week setup
- Expert edit on every page
- Every claim sourced
Seven steps, with a gate at each hand-off
What has to be true before a page moves on
| Gate | Check | Who decides |
|---|---|---|
| Brief approved | Real reader question, clear angle, a reason this page should exist rather than a competitor's | Content lead |
| Research complete | Sources logged with URLs and dates; internal documents and experts consulted; gaps listed | Writer |
| Draft usable | Follows the brief and the voice guide; style checker shows no generic patterns; structure is answer-first | Writer, with tooling |
| Expert edit done | A named expert has added judgement, examples and opinion, and cut anything they would not say to a client | Subject-matter expert |
| Facts verified | Every number, claim and quote traced to a source; unverifiable claims removed; dates current | Fact checker |
| Ready to publish | Structured data, internal links, visible author and date, meta title and description, images with alt text | Editor |
| Refresh due | Scheduled by topic volatility; triggered early by lost citations or rankings in the visibility monitor | Content lead |
Small tools around the model, not a content factory
Grounded research assistant
A retrieval system over your documentation, past proposals, support tickets and product data. Writers ask it questions and get answers with sources, not guesses.
Style and voice checker
Flags generic openers, hedging, filler phrases and sentences over 25 words before a human reads the draft. Tuned to your voice guide.
Claim extractor
Lists every factual claim in a draft with its source, or marks it unsourced. The fact checker works from the list. Built on the same evaluation ideas as our evals practice.
Refresh scheduler
Each page has a review date based on how fast its topic changes. Pricing pages monthly, methodology pages yearly. Lost visibility moves the date forward.
Thin, interchangeable pages that no engine wants to cite. Statistics the model invented, published under your name. Three articles on the same topic from three briefs nobody reconciled. A brand voice that drifts into the model's default register. And eventually, a quality update that removes the lot. Every one of these is a process failure, not a model failure.
Original data from your own systems. Opinions a competitor would not dare print. Worked examples with real numbers. A named author who would defend the page in a meeting. That is the content AI engines cite, and it is the content buyers trust.
Two weeks to a working pipeline
Voice and rules
Interview your experts, write the voice guide, define the gates and who owns each. Days 1–4.
Tooling
Research assistant on your documents, style checker, claim extractor, refresh calendar. Days 3–8.
First cycle together
Four to six pages through the full workflow with your writers and experts, reviewing each gate. Days 8–14.
Hand over or continue
Your team runs it, or we run it monthly with your experts in the loop.
Content is one input. These are the others.
GEO programme
Structure and entity work that makes content citable.
Start GEOProgrammatic SEO
When the value is in data rather than prose.
Scale responsiblyHallucination
Why the fact-check gate exists.
Read the definitionAI Agents & Automation
The engineering behind the tooling we build for content teams.
See the serviceFrequently asked questions
Will search engines penalise AI-assisted content?
Search engines say they judge content on quality and usefulness, not on how it was produced, and they do act against scaled, low-value content however it was made. Our workflow is designed so that every page has an expert's judgement and verified facts in it. That is what makes it rank and makes it worth citing.
What does 'no AI slop' mean in practice?
No page is published without a named subject-matter expert editing it. Every statistic, claim and quote is traced to a source. Generic filler is cut. If a page does not say something a competitor could not say, it does not go out. We would rather publish four good pages a month than forty forgettable ones.
Which models and tools do you use?
Anthropic Claude, OpenAI and Google Gemini models through their APIs, chosen per task. The tooling around them matters more: a research assistant grounded in your documents, a style checker, a claim extractor for fact checking and a refresh scheduler. We build these as small internal tools rather than buying a suite.
Can our own team run the workflow?
Yes, that is the goal. We set up the pipeline, train your writers and reviewers, run the first two or three cycles with you and hand over. Many clients keep us for the monthly refresh review only.
How do you keep the brand voice consistent?
A written voice guide with examples of what to do and what not to do, used both by the humans and as context for the drafting model. A style checker flags deviations before the editor sees the draft. The editor has the final say.