One Piece of Know-How, a Week of Articles and Posts: How an AI Content Engine Works for Expert Firms

Most expert firms I talk to have the same problem. The partners know more than anyone in their market. But that knowledge sits in Word files, PDFs, a directory profile and a blog that stopped years ago. Nobody has time to write every week, so buyers never see it.

That matters because buyers now ask AI assistants first. In G2’s 2026 survey of B2B software buyers, 51% said they start their research in an AI chatbot rather than a search engine. Those were software buyers, not your clients, but it shows where research is heading. An assistant can only quote what it can read on the open web.

Below I explain how an AI content and distribution engine turns one piece of your know-how into checked articles and a week of posts. You’ll see what your firm still approves, what it costs, and who shouldn’t buy one. The real example is a German law firm I call Corvane Legal (the name is changed).

The short version

• An engine rewrites your firm’s existing writing into articles on your own website. Each article then becomes a week of posts.
• Every draft is checked against the original. Anything the original doesn’t support is held for a person.
• Your firm approves: articles with one click from a daily email, posts from a phone. At the start, nothing publishes without that click.
• Building it is a one-off project from from €4,500from $4,950. Running it is mostly AI fees, paid to the providers directly, under a monthly cap you set.
• Without deep expertise, or someone who can say yes within days, it’s the wrong purchase.

What is an AI content engine, and why would my firm want one?

It is a system that moves your expertise from your files onto your website and channels every week, without anyone writing from scratch. It drafts and checks; a person at your firm decides what goes out.

A writing tool like ChatGPT waits for someone to type a request, and writes from what it remembers. An engine runs on its own schedule, works only from your material, and puts every draft through the same checks.

Most firms already have the material. In my experience it sits in four places:

• Documents: client memos, briefings and PDFs that were never published.
• Directory profiles: careful articles on legal, medical or industry directories, building someone else’s website.
• Old articles: a blog that stopped, or a newsletter archive nobody can search.
• People’s heads: the answers a senior partner gives on calls every week. The engine works from written material, so these need writing down first.

How does one piece of know-how become a week of articles and posts?

One document goes in; checked articles and weeks of posts come out. A week at an imaginary firm shows how.

Picture a 12-person tax advisory firm in Munich. A partner, call her Petra, wrote a six-page client memo last year on preparing for a tax audit. It sits in a shared folder. In a case like this, the week could run as follows.

Monday: the memo goes in. The engine splits it along the memo’s own headings, say into four topics. It keeps the original on file, because every draft is checked against it.

Tuesday: drafts arrive. Petra’s daily email holds four draft articles. Each answers one question clients ask, such as “Which documents does an auditor ask for first?” Each comes from her memo, not from the AI’s general knowledge.

Tuesday: one draft is held. It contains a sentence her memo doesn’t support, perhaps a deadline the memo never mentions. The check flags that sentence. Petra publishes three drafts with one click each, fixes the fourth and releases it. Her part is a few minutes of reading per article, not writing.

Wednesday: the articles go live on the firm’s website, with Petra’s name on them. The engine drafts a week of posts from each.

Thursday: the posts are approved. The managing partner, who owns the firm’s voice, reads the week on a phone. One tap approves it; a short note sends a post back.

The week after: the posts go out. One article becomes a week like this. It’s a format example, not output from a client’s account:

Day	Channel	What the post does
Monday	Company LinkedIn page	The article’s main question and its short answer
Tuesday	Petra’s LinkedIn profile	One point, in her own voice
Wednesday	Newsletter	The key point, in a short section
Thursday	X	One claim in a line or two
Friday	Facebook and Instagram	A short checklist from the article

Each post links back to the article on your website. Posts may not add facts, figures or quotes the article doesn’t contain. Four articles, a week each: one memo gives the firm a month of posts.

Who notices? Say a finance director asks an AI assistant what an auditor asks for first. If the assistant reads Petra’s article, it can cite it; nobody can promise that it will. If she then writes in, the firm sees which article her visit started on.

Steady beats loud. In the 2025 Edelman and LinkedIn report, most “hidden buyers” said consistent, high-quality thought leadership makes them more likely to back a vendor. These are the people in finance, legal and procurement who shape a decision without meeting vendors. One good article a week is a pace a small firm can keep for years.

What does my firm still approve?

Everything that carries your name, at the start. Later you can let some kinds of article publish on their own; held drafts and posts still wait for a person.

Decision	Who makes it	How often
What counts as a source	You	Once, at the start
Topics that stay out, such as client matters	You	Once, changed any time
Each article	One named approver, from the daily email	Every article at first; later only kinds you keep on manual, plus every held draft
Each week of posts	Whoever owns your voice, on a phone	Weekly
The monthly AI budget	You	Once, changed any time

Every engine starts in pilot mode, where nothing goes live on its own. Once you trust the output, you can let some kinds of article publish automatically when they pass the check. Each page that went live then appears in the next morning’s email, with a link to take it offline or edit it.

Can I trust AI to write about my expertise?

Not on its own. You can trust a system that checks it, and a person who reads what the check flags.

For the law firm, I let an AI model suggest links between 230+ articles. It made 1,175 suggestions. A plain rule-based check rejected 72 as invented: wording the model made up, or pages that don’t exist. A second AI check, asked only whether each link would help a reader, rejected 179 more.

AI is a good suggester and an unreliable judge of its own work. Had I published its first answer, every rejected link would have gone live.

The article check works the same way. A second AI call reads each draft with one question: does the original say this? That narrows the room for error without closing it, which is why a person approves.

Two shortcuts I don’t take: letting the AI write from memory instead of your source, and letting a failed draft go live without a person deciding. Both are cheaper to build, and both fail quietly.

Is anyone running one today?

Yes. A German law firm runs the article side in its own account. The posting side is my own product, Cuckoo, still in pilot.

Corvane Legal: a law firm’s articles brought home

Corvane Legal’s lawyers had written expert articles for years, but published them on a lawyer directory. Each one built the directory’s standing, not the firm’s. The engine captures those articles, rewrites each into a page that answers a client’s question first, and checks every claim against the original. In pilot mode, a lawyer releases each page with one click.

As observed on 5 October 2026, 1,100+ pages are live on the firm’s own domain. The engine tracks 16 AI crawlers by name (crawlers are the programs AI companies send to read websites). When a prospect writes in, the firm sees which article the visit started on.

AI spend runs at ~€33/mo, held there by a daily limit on rewrites, under a hard €100 monthly stop. In Sept 2026 I moved the whole system into the firm’s own Cloudflare account. I don’t publish traffic or citation numbers yet; they need a full quarter behind them. More in the law-firm AI content engine case study.

Cuckoo: the posting side, in pilot

Cuckoo is in pilot, with platform approvals pending. It drafts a week of posts per channel in a brand’s voice, with 4 AI providers behind it. The client approves the week, or sends posts back with a note, from a phone.

6 publishing connections are built, each waiting on that platform’s approval. Until then, Cuckoo doesn’t post there itself; approved posts go out by hand. LinkedIn, for example, reviews every app that posts to company pages, under its app review rules. That’s the right order: the platform approves the tool, your expert approves the words. More in the Cuckoo pilot case.

Will Google treat this as AI spam?

Not if each page is useful, checked and rooted in real expertise. Google targets pages made mainly to game search results, however they were made.

Its spam policies call that “scaled content abuse” and name AI as one way to do it. Its guidance on AI content says AI-written pages should be fact-checked and reviewed by a person before they go live. Here is how the engine answers it:

• A source you can point to. Every article comes from your own material, not from what the AI remembers.
• A check, then a person. Every draft is checked against its source. In pilot mode a person reads every one; after that, held drafts still wait for a person, with the doubtful sentences marked.
• A named expert on each page. Google’s guidance on helpful content asks whether it’s clear who wrote a page and how AI was used. A name and a short note answer both.
• Real questions only. Topics come from your material and your team, not from keyword lists.

None of this promises a ranking, and I don’t sell it as if it did.

Couldn’t we just use ChatGPT, or hire an agency?

You can, and for some firms that’s the right call. The difference is in three places:

What differs	Content agency	ChatGPT and a scheduler	Content engine
Where the facts come from	Interviews and the writer’s research	What the AI remembers	Your own documents and articles
Who checks them	An editor, if budgeted	You, by hand	Every draft against its source; failures held for a person
Who owns it	The agency runs the process	There is no system	You, after handover

What does it cost to build and run?

Building is a one-off project, priced per module. My published prices, excluding VAT:

Module	What you get	Timeline	Price
Edge Hub	Your back catalogue as a knowledge hub beside your current site.	2–3 weeks	from €4,500from $4,950
Full Engine	Continuous capture → rework → fact-check → approval → publish.	6–9 weeks	from €9,500from $10,500
Distribution module	A week of on-brand posts per channel, approved from a phone.	3–4 weeks	from €3,500from $3,850
AI Department (optional)	I run the engine, watch the budget and make changes. 20 hours/month. I take on at most 6 retainer clients.	monthly · 3-month minimum	from €2,900/mofrom $3,200/mo

The Edge Hub is a knowledge section on your own web address, beside your current website, which stays as it is. Typical projects: Edge Hub €4.5k–7k$5k–7.5k, full engine €9.5k–18k$10.5k–20k, distribution module €3.5k–6k$4k–6.5k. Where you land depends on how scattered your material is and how many channels you want.

Running costs are ~€30–150/mo, paid directly to the providers~$35–165/mo, paid directly to the providers. That is mostly AI fees, plus hosting. A budget breaker, a hard monthly limit you set, stops all AI calls when it’s reached, so there is no surprise bill. The law firm above spends ~€33/mo on AI. There is also your time: one approver who opens the daily email, and someone who reads the week of posts.

I usually suggest starting with the Edge Hub. It ships in weeks and shows whether your material is rich enough to feed an engine. If you aren’t sure where to start, the AI Visibility Deep Audit at €1,500$1,650 is the smaller first step. Fixed price. What you pay in cash is credited in full to any build within 60 days. All prices are on the pricing page.

Is this right for my firm?

It fits if you have deep expertise written down somewhere and one person who can say yes within a few days. It’s the wrong purchase in four cases:

• Thin expertise. With no deep, specific knowledge to capture, the engine has nothing true to say. I won’t ask an AI to make it up.
• Pages for rankings alone. Thousands of near-identical pages to catch search traffic is the spam pattern above. I don’t build it.
• No approver. If nobody can approve within a few days, drafts pile up. An engine without an approver is just a queue.
• One article a month is enough. Then a good writer and a calendar cost you less.

What you can do this week

About two hours in all, and none of it needs me:

• Check whether AI can read your site (2 minutes). Many sites show their text only after a script runs. Some AI crawlers don’t run scripts. My own old site showed them 0 words. Use the free scan below.
• Ask an assistant what your clients ask (15 minutes). Put three typical client questions to ChatGPT or Perplexity. Note whose articles it cites.
• List where your know-how lives (30 minutes). Shared folders, directory profiles, old blog posts, newsletter archives.
• Pick one document (15 minutes). One memo or guide you’d be happy to publish from.
• Write down your no-go topics (20 minutes). Client matters, anything privileged, claims your compliance team won’t sign.
• Name an approver and a backup (10 minutes). Someone with authority who will actually open a daily email.
• Check who holds the keys (20 minutes). Who has admin rights on your domain, your LinkedIn page and your newsletter tool?

See what an AI crawler gets from your site today.Free AI-readiness audit · about 60 seconds.

Your website address Scan my site →

For your web developer

• The hub runs on Cloudflare Workers, bound to one path on your domain. The existing site is untouched; if the Worker fails, requests pass through to it.
• Capture is deterministic: a model may propose where articles split, but the text itself never goes through the model.
• The pipeline runs every 5 min. The claim check marks each claim supported, partial or unsupported; JSON-LD is built in code, never by a model.
• Content must be in the server’s HTML, and bot settings must let AI crawlers in (Cloudflare’s can block them).
• Attribution carries the article in the URL to the contact form, without cookies. Everything runs in your own Cloudflare and AI-provider accounts.

If you have the material and someone who can say yes, start a project with a short letter and tell me where your expertise lives today. I’ll reply with which module I’d build first, and why.