AI Content Creation for Professionals: Turning Expertise Into Posts

Feed the AI your own material rather than a topic: a two-minute voice note about a question a client asked this week, with the client's details removed. Ask it to draft from your words only, edit the draft into your voice, then run a profession-specific compliance check before anything is published. About an hour a week is enough.

Generic posts come from generic input. The model has never advised a client, so a prompt that says "write a LinkedIn post about common business mistakes" returns the average of everything it has read, which is what your readers have already scrolled past. The expertise is yours, and so is the compliance risk. The AI is a fast first editor of both, as long as you give it the right material and a clear list of what it must not say.

Follow me on Instagram@sagnikteaches

Why topic prompts give generic posts, and what to feed instead

Here is what a topic prompt produces for an accountant ("Write a LinkedIn post about cash flow for small business owners"), abridged and illustrative: "Cash flow is the lifeblood of every business! Here are 5 tips to keep yours healthy: 1) Invoice promptly 2) Chase late payers 3) Forecast regularly... Need help? Get in touch today! #CashFlow #SmallBusiness". Nothing in it is wrong, and nothing in it could only have come from this accountant. It is also the shape of post that tends to slide into advice, because "tips every owner should follow" speaks to nobody in particular and everybody at once.

Connect on LinkedInSagnik Bhattacharya

Now the input-led version. The accountant records a short voice note after a client call, strips out the identifying details, and pastes the transcript with a prompt that tells the model to use it and nothing else. Illustrative transcript: "A client asked why the business showed a profit but the bank balance kept falling. I told her profit isn't cash. She'd paid $14,000 for stock that hadn't sold, and two customers were 60 days late. Neither shows up as a cost. We built a 13-week cash view on one sheet." The draft that comes back reads: "A client asked me this week why her accounts showed a profit while her bank balance kept falling. Profit and cash are different things. She had paid $14,000 for stock that hadn't sold, and two customers were running 60 days late; neither is a cost in the profit figure. We built a 13-week cash view on a single sheet: what is due in, what is due out, and the gap by week. If your accounts look healthier than your bank balance, ask what you have spent that isn't a cost yet, and who owes you money."

Subscribe on YouTube@sagnikteaches

That post says one thing, uses a real example, and carries no instruction about what the reader should do with their own money. It is general information built from a specific moment, which is the pattern to repeat. Repeating it every week without mishaps takes a capture routine, a voice guide and a review step.

The five-minute capture: three questions answered aloud

Capture is the step that decides whether posts sound like you. Once a week, ideally straight after a client call, answer three questions into your phone for two or three minutes: what did someone ask me this week, what did they assume that was not true, and what did I tell them to do first? Speaking is faster than typing and keeps your own phrasing, which is exactly what the model lacks. Turn the recording into text with your phone's dictation or the transcription feature in your notes app, then remove names, places, sectors and exact amounts that could identify anyone before the text goes near an AI tool.

Notice what the third column contains: a first action, not a verdict. Each row can become a post that informs without deciding anything for the reader. The solicitor's "write it down and email it" is a habit worth sharing in any setting; "yes, it is binding" would depend on the facts and on the law that applies, and is exactly the sentence that should never appear in a post.

A 150-word voice guide the AI can follow

A voice guide is a short list of concrete rules drawn from three of your own emails or posts, not a paragraph of adjectives. Models follow "start with a number or a client question" better than "warm but authoritative". Write down your typical sentence length, words you use and words you never use, how you open, what you never do, and how you sign off. A filled-in example for a solicitor, which a colleague could check in a minute:

VOICE GUIDE (v3, reviewed by A. Partner)
- Sentences: mostly under 20 words. One short one per paragraph.
- Open with a specific moment or a plain statement of a common assumption. Never open with a greeting or a question to the reader.
- Say "we" for the firm, "I" only in the first person example.
- Use plain words: "ask" not "enquire", "end" not "terminate", "paper" not "documentation".
- Never use: leading, best, expert, specialist, guarantee, always, never (about outcomes), "you should".
- No lists of tips, no hashtags, no emojis, no rhetorical questions.
- Examples are hypothetical and say so ("a client in this position"). No real matters, no names, no amounts that could identify anyone.
- Close with one plain next step the reader could take alone, or one question they could ask a professional.
- Sign off with first name only. No sales line.

Keep the guide in a place the assistant reads every time, so you do not paste it each week. A project in ChatGPT or Claude does this (both offer projects on free plans), and so does a Gem in Gemini (becoming a skill from November 2026). Google says the switch reaches personal accounts first and Workspace business accounts in 2027, and that it will recreate existing Gems as skills automatically; confirm yours still behaves as you expect after the change. Update the guide whenever a post needs an edit you have made twice, because the edit you keep repeating is a missing rule. The tutorial on Claude Projects for marketing: brand voice, briefs and a content calendar shows the same set-up for a general business.

From notes to draft: a prompt, its output and the edit that matters

Take the solicitor's row from the capture table. The transcript, after the solicitor has removed anything identifying, reads (illustrative): "A client asked on Monday whether a handshake deal with a supplier counts. She assumed that without a signature there was nothing to worry about. I said whether it counts depends on the facts and I wasn't going to guess on the phone. What I told her to do first was write down what was agreed, who, what, how much and by when, and email it to the supplier that day." The prompt that turns it into a post does three jobs besides drafting: it blocks new facts, forces a hypothetical framing, and makes the model confess what it added.

Draft a LinkedIn post of 150 to 200 words from the transcript below, following the voice guide in this project.
Use my words and my example. Do not add facts, rules, statistics or legal conclusions that are not in the transcript.
Present the example as hypothetical ("a client in this position").
Do not tell the reader what they should do about their own situation.
After the post, list in square brackets everything you added that was not in my transcript.

A realistic draft (illustrative):

Picture a client who has agreed terms with a supplier on a handshake. She assumes that without a signature there is nothing to worry about.
Whether the agreement counts depends on the facts, and it is not a question to guess at over the phone. What a client in this position can do first is write down what was agreed (who, what, how much, by when) and email it to the other side the same day. The reply is often the most useful part, because it shows whether both people remember the same deal.
If you are unsure about an agreement of your own, a short conversation with a solicitor now is usually cheaper than a long one later.
[Added: "the reply is often the most useful part"; "usually cheaper than a long one later"]

The bracketed list is the most valuable line in the output, because it turns the model's embellishments into a checklist. Here both additions are claims the solicitor never made. "Usually cheaper than a long one later" is an unsupported generalisation and a gentle sales line; cut it. "Often the most useful part" is a plausible observation but still the model's, so either keep it because it matches your own experience, or reduce it to "it shows whether both people remember the same deal". Two further edits are typical. Change "can do first" back to "I told her to do first" so the post carries your own action, not generic guidance. And read the closing line against your voice guide: "sign off with first name only, no sales line" means the last sentence about speaking to a solicitor goes, or becomes a plain question the reader can put to any professional.

Wording limits for accountants, solicitors and consultants

Each profession has its own tripwires, and your own regulator or professional body has the final word on all of them. What follows are the patterns that recur, written as rules the AI can check and you can apply by eye. The common thread is that a post may inform in general terms but may not advise one reader, promise a result or reveal a client.

One more rule for all of them: a disclaimer does not repair a post that advises. "This is not legal advice" under a paragraph that tells the reader what to do about their own contract protects nobody. The disclaimer is a second layer; the first is writing a post that does not need one. If the model ever writes the words "you should", treat it as a flag. The model vendors draw a related line. From 20 November 2026 OpenAI's usage policies also bar presenting the AI itself as holding professional credentials, which is one more reason to keep a named professional behind every post.

A ten-minute review before anything is published

Review has two readers. The first is the AI, used as a quick filter: give it your wording rules and ask it to quote what breaks them, without rewriting anything. The second is a person, who is never the author for regulated advice. A good filter prompt is short:

Check the post below against my rules list. Quote every phrase that (1) tells the reader what to do about their own situation, (2) promises or implies a result, (3) compares us with others or claims we are best or expert, (4) states a rule, rate or deadline as current, or (5) could identify a client or a matter. Quote the phrase and say which rule it breaks. Do not rewrite the post and do not say that it is compliant.

On the accountant's cash-flow draft from earlier, a reply looked like this (illustrative): "Line 6: 'ask what you have spent that isn't a cost yet' is a question to the reader, not an instruction; acceptable. Line 3: '$14,000' and 'two customers' could identify the client if she recognises the story; consider rounding or generalising. No other phrases found under rules 1 to 4." The reply is useful and incomplete, and you should expect both. The filter caught the amount, and it missed that the post says nothing about rules changing, which is fine here but would not be for a post about a deadline. It also cannot know your regulator's rules, only the list you supplied. That is why the human read happens anyway.

The human read follows a short checklist, which the reviewer ticks and initials:

1. Could a reader take this as advice about their own situation?
2. Does it promise, guarantee or imply a result?
3. Does it use a superlative, a comparison or a claim of expertise?
4. Is every figure, rule and deadline either dated or removed?
5. Could any detail identify a client, a matter or a person?
6. Does it name another firm, product or regulator unfairly or inaccurately?
7. Is it in the firm's voice, with a named person behind it?
8. Has the reviewer, not the author, approved it, with the date recorded?

Keep the approvals in a simple log: date, post title, author, reviewer, any changes. When the platform or your regulator asks how posts are approved, the log answers it. Schedule only approved posts, and never connect the drafting tool directly to the publishing tool; the tutorial on an approval queue for automated social posting shows how to keep a human gate in the middle.

One week for a three-partner accountancy

An illustration with assumed minutes. Three partners share one weekly post, each writing one every third week, with another partner as reviewer. A typical week for the author looks like this:

Say a post used to take two and a half hours from a blank page, which is 150 minutes. Over 46 posting weeks, allowing six weeks off, that was 6,900 minutes, or 115 hours across the firm. At 65 minutes it is 2,990 minutes, or about 50 hours. The saving is about 65 hours a year. Each partner writes about 15 posts and reviews about 15 more, 55 minutes of authoring and 10 of reviewing each time, so the load per partner is a little over 16 hours a year. The sum has a floor worth respecting: the 25 minutes of editing and the 15 of review are the parts that keep the post yours and safe, so when time is short, post less rather than review less.

Habits that make a post sound like everyone else's

Models write the left-hand column below by default, because it is the average of what is already online. Your capture notes are the only source of the right-hand column. If a draft is mostly left-hand material, the input was too thin, and the fix is another minute of your own words, not a better prompt.

What goes in instead is always one of four things: a number from your practice, an anonymised question, an opinion someone could reasonably disagree with, or the first step you would take. None of them needs the model to know anything. All of them need you to have said it aloud once. A final read for rhythm helps too: if every sentence is the same length, or every paragraph ends on a neat summing-up line, read it aloud and break the pattern.

Where posts come from after week four

After a month the problem changes from "what do I say?" to "am I repeating myself?". Keep a questions log in a notes app, one line per client question with the dates used. Three entries a week outnumber one post a week, so you pick the best rather than the latest. Add three recurring series that cost nothing to plan: "A question this week" (the capture routine), "A common mistake" (always hypothetical, never a real matter), and "What changed" (only when a rule or deadline has changed and you have confirmed it from the official source, not from the model).

Calendar items need extra care. Put posts about deadlines and year ends in the diary six weeks ahead, and verify every date the week before publishing, because a model's memory of a deadline is not a source. Once a quarter, search your own posts for dates, rates and rules and update or take down anything stale. And when you reuse a post as a newsletter paragraph or an answer on a service page, run the review again; wording that was fine in a post can be a problem on a page that sells a service. The plan for a regulated trade is laid out month by month in the tutorial on AI marketing for accountants, and the general version of the voice problem is covered in writing LinkedIn posts with AI without sounding like AI. For the firm-level rules on what may be said about a law practice, see AI marketing for small law firms: content, reviews and the rules.

Questions professionals ask about AI-assisted posts

Should we say a post was written with AI?

Check your professional body's guidance and any platform rules first. Where nothing requires a label, the better discipline is that a named professional stands behind every post and has read it. Decide one policy and apply it consistently, and never describe the AI as if it were a qualified person.

Can the AI publish posts for us automatically?

I would not allow it. An unreviewed post is exactly what the review step exists to stop, and a scheduled post can go out after a rule or deadline has changed. Schedule only approved posts, record who approved each one, and keep one person able to pull any post quickly.

How do we use client stories without identifying anyone?

Prefer composites labelled as hypothetical, such as a client in a position like this one. Strip anything identifying, including sector, place and amounts, and get written permission whenever a real matter could be recognised. Confidentiality duties can cover anonymised stories, so ask your compliance contact before you rely on one.

How many posts a week are worth it?

One you have checked beats three you have not. Start with one a week for eight weeks, then look at what you learned and how long review took. If review starts slipping, cut the frequency, not the checks.

Further reads

• AI Case Study Writer for Freelancers: Draft It and Get Client Approval — Turn a finished job into a case study with client approval.
• AI Marketing for Financial Advisers: Content That Clears Compliance — A fuller review routine for regulated advice firms.
• Mortgage Broker Social Media With AI: A Month of Compliant Posts — A month of compliant posts in another regulated trade.
• Can You Use AI to Write Testimonials? Where the Legal Line Is — Where the line sits on AI-written testimonials.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI usage policies (revision effective 20 November 2026). The worked examples and time figures are illustrative assumptions, not measurements.