The worst business brief for artificial intelligence has five words: ‘Write us an article about AI.’ The model will comply, but you may receive polished copy that anyone could have published, contains inaccuracies and leads to no commercial decision.
Text remains a good first AI experiment for small and medium-sized businesses. The important condition is to begin not with generation, but with source material and accountability.
AI is not the source of your experience
A model can assemble, shorten, compare and propose alternatives. It has not attended a meeting with your customer, seen your production process or learned why you abandoned an approach. Without such information, it fills gaps with plausible generalities.
A sound division of work looks like this:
- A person provides: intent, facts, experience, priorities, boundaries and final accountability.
- AI helps with: an outline, alternatives, counterarguments, shortening, changing format and editorial review.
- A person decides: what is true, what fits the brand and what may be published.
A workflow for copy your company can sign
1. Define one reader and one decision
Define a specific reader and the decision they face, rather than writing merely ‘for businesses’. Also state what the reader should know or do after reading. Without that, the model mainly optimises fluency.
2. Prepare a pack of facts
Include an approved service description, expert answers, concrete examples, limitations and links to primary sources. Mark what is a company claim, what is a cited fact and what must not be guessed. Anonymise confidential data or use approved tools and settings.
3. Show the voice, not a list of adjectives
Instructions such as ‘write professionally and humanly’ are not enough. Three to five approved examples work better, with an explanation of how the company begins, uses first person, avoids exaggerated results and translates technical details into business impact.
4. Let the model ask questions first
Before an outline exists, have the model list missing materials and disputed claims. This stage often saves more time than faster drafting because it reveals a missing audience, price, condition or proof.
5. Create the draft in parts
For a longer text, approve the angle, structure and main conclusion first. Only then expand the chapters. This reduces the risk of polishing several pages built on the wrong premise.
6. Switch the model into the role of critic
After the first draft, do not ask only for improvement. Ask for unsupported claims, generic passages, repetitions, missing examples and questions the reader still has unanswered. The criticism must be specific, and a person must decide what to accept.
7. Verify facts outside the model
Check current features, prices, legislation and technical recommendations in primary sources. A model can state a non-existent feature and produce a convincing-looking link. Record the source and review date for important claims.
8. Carry out a final human edit
Read the text as a customer. Does it answer quickly enough? Does it say for whom the advice does not apply? Does it lead naturally to a service or useful further content? Does it sound experienced rather than compiled from internet maxims?
Where AI creates the most value
- alternatives for subject lines, headlines and calls to action from an approved offer,
- turning a webinar or interview into an article outline,
- shortening a long expert text without changing facts,
- preparing questions for an expert,
- finding objections and missing parts,
- adapting one source document for a newsletter, social network and sales follow-up,
- language and structural review before human editing.
For these tasks, the input is traceable and the result can be compared. That is why they can be introduced in a smaller company without a large AI project.
Where I would set stricter controls
Health, legal and financial advice, contractual terms, crisis communication, public company results and copy based on non-public customer data carry higher risk. AI can help with structure, but a named expert must approve the result.
Be equally careful with automatic translation into a language nobody on the team can check. Fluency is not proof of correctness, and a small shift in meaning can cost money in an offer.
How Google views AI content
In its guidance on generative content, Google does not say that text created with AI assistance is automatically bad. Accuracy, quality and value for users matter. Mass-producing pages without value to influence rankings may breach its policy against scaled content abuse, whether made by people, automation or both.
Do not use AI simply to fill a calendar. Use it to produce better-supported content at lower cost. I discuss how evidence and experience affect search in How to become a source for Google and AI answers.
How to measure whether the workflow saves work
Before a pilot, measure the time from brief to approval, revision rounds, factual errors and the commercial outcome. Compare the same figures after several outputs. A draft created in a minute is no saving if a senior person spends two hours correcting it.
Also measure repeatability. A one-off good result is not a process. You need a briefing template, saved examples of voice, prohibited claims, an owner for facts and clear approval.
Start with one text, not a whole-company transformation
Choose an article or newsletter for which you already have good source material. List the facts, create a draft using this workflow and record time and errors. Then decide which part can be standardised.
For a specific tool, see my overview of seven AI tools for working with text. If you need to create a brand voice and a safe editorial process first, write to me through contact.