In small teams, social media content is now largely produced with the help of a language model. That is not something to confess; it is a reasonable productivity decision. The problem lies not in using the tool but in the absence of any structure around how it is used.
When no structure is built, the result is familiar: all the posts start to look alike, the sentences are correct but belong to nobody, and the business's real voice gets lost. The copy is not bad — worse, it has no character. Any other business in the sector could publish the same text.
As far as we measured, Turkish search results for "yapay zeka ile sosyal medya içeriği ve marka sesi" (AI social media content and brand voice) do not return an editorial guide; every result is either a software product page or a video. The text below does not recommend a tool. What it covers is how to build a brand voice guide, what to give the model, where to verify the output, and which jobs are never handed to the model under any circumstances.
"Content for AI" and "content with AI" are not the same thing
When these two expressions get mixed up, the discussion gets mixed up too.
Content for AI is about building text so that an assistant can read and cite it: where the answer to the question sits, how the structure is put together, how the claims are supported. We collected the mistakes made on this side, and how to fix them, in our article on the mistakes made when writing content for AI.
Content with AI, on the other hand, is a production operation: who drafts, who edits, who approves, which information gets verified. This article is about the second.
Drawing the distinction has a practical effect: the first is a question of content quality, the second a question of workflow. If the second is sped up before the first is solved, what you get is an archive that was produced quickly and does nothing.
What a brand voice guide looks like: one page, not five
The document that comes to mind when people say brand voice guide is usually long, abstract and unused: "friendly but professional", "trustworthy", "innovative". These adjectives tell a model nothing, and they tell a new employee nothing either, because every business describes itself this way.
A guide that works is one page long, and it contains examples and rules, not adjectives.
| Section | What goes in it |
|---|---|
| Who we write for | The reader's real situation, in one paragraph |
| How we address the reader | Formal or informal "you" (siz or sen in Turkish), one decision |
| Words we use | The terms the business actually uses |
| Words we do not use | A list of banned expressions |
| Three good examples | From our own archive, real copy |
| Three bad examples | Why each one is bad, in one sentence |
The two most valuable rows are the last two. Telling a model to "write professionally" does not work; telling it "the three texts below are our voice, make it sound like them" does. The examples should come from your own archive — other brands' copy does not teach the model your voice, it teaches theirs.
The banned-words list has to be concrete as well. "Avoid exaggeration" is not a rule; "do not use these expressions" is a rule, and it can be applied directly.
The input you give the model: real customer sentences, not the product
In text generation, what decides the quality of the output is not the length of the instruction but how real the material behind it is.
The usual approach goes like this: the name of the service is typed in and the model is asked to "write a post about this". What comes out is generic, familiar and characterless — because from a single input the model can only produce something generic.
The input that works is the business's own raw material: the question a customer actually asked on the phone, the sentence they typed in a message, the objection that keeps coming up, the explanation that comes from inside the business. If this material is already being gathered — we wrote in detail about how one topic list can feed two channels in our article on mapping blog content to social media — the input problem solves itself.
A warning: when you put real customer sentences into the input, anything that identifies the person has to be stripped out. Names, phone numbers, addresses, health information or details that would make someone recognisable should not be entered into a third-party tool. We covered separately why this is not merely a matter of courtesy, and where the risks concentrate in free tools, in our article on free AI tools for small businesses.
What the model does well, and what it cannot do at all
Putting this distinction in writing ends most of the arguments about expectations inside the team.
What it does well. Adapting an existing text to different lengths, pulling short pieces out of a long text, turning scattered notes into an orderly structure, producing several different phrasings of the same idea, correcting spelling and flow, and drafting the headings and subheadings for a text.
What it cannot do. Know the truth inside the business. It cannot know your pricing policy, which jobs you take on and which you turn down, how your team works, or what happened last month. If this information is not in the input, the model makes it up — and made-up information usually reads as fluent and convincing.
The distinction fits into one sentence: the model produces the form, you supply the content. Leaving the form to the model is efficient; leaving the content to it turns every published text into a risk.
How to recognise template language
What gives model-generated text away is not grammatical errors but recurring patterns. The six below are the most common, and they are easy to fix.
- The empty opening sentence. An introduction that says nothing, such as "With the advance of technology in our day and age…". If the text loses nothing when it is deleted, it should be deleted.
- The triple adjective. Adjective triplets such as "fast, reliable and effective", none of which can be measured.
- The symmetrical list. Lists in which every item has exactly the same length and structure. Real information is not distributed symmetrically.
- The contrast pattern. The "not X, but Y" structure repeated several times in a single paragraph.
- The closing call. A generic close such as "Take the first step today" that is tied to nothing.
- The unquantified claim. Unsourced generalisations such as "many businesses", "most experts", "research shows".
The sixth is not only a style problem but an accuracy problem, and it is the subject of the next section.
The verification layer: numbers, claims, regulation and price
In text produced by the model, four kinds of information must always be checked by a person before publication. These four are the most expensive to undo when something goes wrong.
Numbers and ratios. To complete a sentence, the model can produce a number that looks plausible. No number whose source cannot be shown should be published; if the number is not needed it should come out of the sentence, and if it is needed it should be written together with its source.
Claims. Sentences of the form "this method delivers that" should be sentences the business can genuinely stand behind. A promise made in a post is a promise made to the customer.
Regulation and sector rules. In regulated fields — health, law, finance — text produced by the model cannot be used as a source. In these fields the text has to be tied to the regulation in force itself and to the approval of an authorised person.
Price and campaign information. Only the business knows its current pricing policy. If this information was not given in the input the model guesses; once published, it is the hardest kind of error to correct.
The verification layer should not be a habit but a mandatory step in the workflow: a text does not move into the publishing queue until these four headings have been checked.
Image and video generation: no substitute for your real premises and team
Image generation carries a different risk from text generation, because an image is read as a claim.
Presenting a generated image as the business's real premises, team or work creates a false expectation in the viewer. This is not only an ethical question — a customer whose expectation is not met once they see the service is the most expensive customer there is.
The distinction is simple: in abstract and explanatory images a generation tool can be used — an illustration that explains a concept, a background, a diagram. In images that represent reality it cannot — the premises, the team, the product, the result of the work. For this second group the only right source is the business's own photography.
There is an in-between area that needs care: editing a real photo with a generation tool. Correcting light and framing is one thing; an intervention that changes the content is another. The second may additionally be prohibited in regulated sectors; in those fields the text of the applicable regulation is what decides.
Approval workflow: who produces, who edits, who publishes
The time the tool saves is lost again through disorder. The workflow is built on three roles, and in setups where all three are the same person the error rate goes up, because nobody can read their own text critically enough.
The producer creates the draft: prepares the input, uses the model, builds the first version. The aim at this step is not a perfect text but a draft that can be worked on.
The editor pulls the text into the business's voice and strips out template language. At this step the brand voice guide is used like a checklist.
The approver applies the verification layer and makes the publishing decision. In regulated fields this role has to lie with an authorised person.
Having the workflow in writing and keeping a record of each step ends any later "who published this" argument. Connecting workflows like this with tools and automating the repetitive steps is a separate job; we described its scope on our agentic AI systems page.
Measurement: does it really speed things up
The answer to this question should not be assumed; it should be measured — and measuring it is not hard.
Two things are recorded: a piece's preparation time and its number of revisions. The same two numbers are kept for a few pieces before the tool comes into use and for a few pieces after.
The pattern seen in most teams is this: the time spent on the first draft falls and the time spent on editing rises. Whether the total time actually falls depends on the team's editing discipline. When the editing step is not taken seriously, the time drops sharply but the published content loses its character; that is not a measured gain but a hidden loss.
The second item to measure is the output itself: after the tool came into use, did the demand each content type brings in change? If you gain speed while losing response, the gain is negative.
Red lines: what we never leave to the model
This list should be short and not open to negotiation.
- Personal data belonging to customers. Names, contact details, health information or details that would make someone recognisable are not entered into third-party tools.
- Complaint replies. The reply to a negative review is one of the most-read texts a business publishes, and it is written by a person.
- Crisis communication. The text published when something goes wrong is put together by the person who takes responsibility.
- Regulated medical, legal and financial statements. In these fields the text is tied to the regulation in force and to the approval of an authorised person.
- Price and commitment sentences. Promises the business will stand behind are written by the business.
- Images that represent reality. Images of the premises, the team, the product and the result of the work are not generated.
What these six items have in common: in every one of them the cost of a mistake is greater than the time saved. The right place for the tool is where that cost is low.
If you would like to build your content workflow around these distinctions and draw your brand voice guide out of your own archive, take a look at our social media management service or get in touch with us.
Frequently Asked Questions
Do platforms penalise AI-generated content?
Because platforms' rule sets keep changing, it would not be right to give a permanent answer; but the common framework points towards judging the quality of the content rather than the tool itself. Content that is misleading, repetitive, spammy or impersonates someone else causes problems whichever tool produced it; an informative, original text is likewise assessed independently of the tool behind it. The practical approach is not to try to keep up with the rule set but to keep the content defensible on its own terms: if the text rests on real information and says things the business can stand behind, the production method stays secondary.
Should we disclose that we use AI?
It would not be right to speak of a general obligation; expectations on this vary from platform to platform and from sector to sector, and in some fields they may be separately regulated. The distinction can be drawn like this: the fact that a text's draft was prepared with a tool generally does not call for a disclosure, as long as the end product remains the business's responsibility. By contrast, if a generated image or voice is presented as a real recording, disclosing it is not just a preference but a matter of trust. In regulated fields the answer should come not from guesswork but from the text in force in that field.
How do we teach the model our brand voice?
With examples, not adjectives. Definitions such as "friendly but professional" do not work, because they produce the same thing for every business. The method that works is to pick three texts from your own archive that you genuinely like and three that you do not, add them to the input every time, and also provide a concrete list of expressions you do not use. These six examples and the banned list are the core of a one-page brand voice guide, and they get updated over time as the archive grows. If the examples are taken from other brands, the voice being taught will not be yours.
Doesn't the same prompt produce the same content for everyone?
When the input is generic, largely yes — and that is the real source of template language. The same instruction given with the same service name produces similar texts for everyone in the sector. What makes the difference is not how clever the instruction is but whether the input is specific to the business: real customer questions, explanations from inside the business, example texts chosen from your own archive. This material does not exist anywhere else, so the text produced from it is not one anyone else could produce either.
For which content types should AI never be used?
In practice six headings can be treated as red lines: anything that contains personal data belonging to customers, replies to negative reviews and complaints, crisis communication texts, regulated medical and legal statements, price and commitment sentences, and images that represent the business's real premises, team or work results. What they share is that when a mistake is made, the cost is greater than the time saved. For the rest of the work — drafting, adapting length, editing, producing headline alternatives — using the tool is a reasonable productivity decision.