Checklist for Reviewing AI-Generated Text Before You Send It

Checklist for Reviewing AI-Generated Text Before You Send It

To review AI-generated text, check in this order: facts and claims, numbers and names, tone, confidentiality, and finally who signs and sends it. The most common mistake is reading only for fluency, because AI text almost always sounds correct even when it is not. A useful review treats the draft as work from a fast writer who has no guaranteed access to the truth about your company.

Key takeaways

• Fluent text is not true text: review the claims, not just the writing.
• Every number, date, proper name, deadline and commercial condition needs a source you can point to.
• The tone should sound like your company and fit the channel, whether that is email, WhatsApp or a public page.
• Personal data and internal information should not appear in the final text, and should not have been pasted into the prompt without a rule.
• A named person with responsibility approves the send. The tool is never the final author.

Why AI text needs a different kind of review

When a colleague writes something, you tend to trust what they know and doubt what they do not. With AI this flips: the model writes with the same confidence about what is right and what it made up. There is no hesitation in the text to act as a warning.

This shows up in the risk documentation from people who study the topic. OWASP lists overreliance on model output and disclosure of sensitive information among the risks of applications built on language models. NIST, in its AI risk management framework, treats human review and oversight as part of responsible use, not as an optional step. Those are official sources. What follows is a working method, not a standard.

The practical consequence is simple. The review has to be done against a list, because a quick read tends to approve whatever sounds good. A short list, applied every time, is worth more than a careful but improvised review.

What to check first: facts and claims

Start by underlining every sentence that states something verifiable. You do not need to check transition sentences or a polite opening. What matters are the sentences that cause damage if they are wrong: what the product does, what is included, what the law requires, what the company promises.

For each claim, ask three questions:

• Do I know this is true, or does it just seem plausible?
• Is there a document, contract, official page or responsible person who can confirm it?
• If the customer quoted this sentence back to us, would the company stand behind it?

If the answer is "it seems plausible", the sentence is removed or rewritten from a source. A useful rule: the model can only state what you put in its context. If a fact was not in the material you provided, any extra detail came from the model's statistical memory and deserves double suspicion.

Pay special attention to promises. Delivery deadlines, warranties, coverage and cancellation terms are often "rounded" by AI to sound more helpful. That creates commitments nobody in the company made. To understand who answers when this goes wrong, read who is liable when company AI makes a mistake, which covers Brazil's consumer code (CDC) and data protection law (LGPD).

How to check numbers, dates, names and links

Numbers and names are where AI fails most often without warning. A swapped value, a date in the wrong year, a job title attached to the wrong person or a link that does not exist all go unnoticed because the format is perfect.

Use this short routine for the check:

• List every number, date, person name, company name and link in the text.
• Compare each item with the original source, not with another answer from the same AI.
• Open each link and confirm that it exists and says what the text claims.
• Confirm the spelling of proper names, legal company names, acronyms and product names.
• Review units and formats: currency, time zone, decimal separator and date format for the reader's country (for Brazilian readers, day before month and a comma as the decimal separator).
• If an item has no source, remove it or replace it with wording that does not need a number.

For public text there is one more criterion. Google's guidance on helpful, reliable content emphasizes originality, demonstrable experience and transparency about who wrote the content and how. An article with data of unclear origin weakens exactly that. If your goal is to appear in search, the supporting reading is the article on AI content SEO and whether Google penalizes it.

An honest caveat: asking the same AI to check its own text reduces some errors but does not replace the source. The model can confidently confirm something it invented. Use a second AI pass to find internal inconsistencies, not as proof of accuracy.

Is the tone right for this reader and this channel?

After the facts, review the voice. A text can be correct and still damage the relationship: too formal for a long-standing customer, too upbeat for a complaint, too generic for someone expecting a specific answer.

Questions for the tone review:

• Does the text answer what the person actually asked, or does it fill space with generic courtesy?
• Does the level of formality fit the channel? An email to a director should not read like a WhatsApp message.
• Are there expressions our company would never use, such as exaggerated enthusiasm, catchphrases or manual-style sentences?
• Does the text acknowledge the customer's problem when there is one, or does it just apologize vaguely?
• Would it sound natural read out loud?

A common sign of automatic text is excess symmetry: three benefits, three sentences of the same length, a standard invitation at the end. Break that pattern. Cut what does not inform and shorten sentences where the subject is simple.

In sensitive situations, such as public complaints, tone matters as much as content. The article on handling consumer complaint sites with AI explains how to use history and drafts without losing human review.

What must not leak: confidentiality and LGPD

The confidentiality review has two sides: what is in the final text and what was sent to the tool to generate the draft.

In the final text, look for:

• Third-party personal data the recipient should not see, such as name, phone number, tax ID, address or another customer's history.
• Internal information: margins, strategy, supplier disputes, team problems, contract details from another customer.
• Passages copied from third-party or confidential documents.
• References to earlier conversations the recipient was not part of.

When using the tool, ask before pasting any material: can this data leave the company? Does the tool have the right settings and contract for this kind of information? That decision belongs to internal policy, not to each person in the moment. If your company does not have that rule yet, the guide on a company policy for employee ChatGPT use is a good starting point.

On LGPD, it is worth stating only what is safe to say: it applies to the processing of personal data, including when done through AI tools, and ANPD is the authority that publishes official guidance. For specific questions about legal basis or data transfer, consult the ANPD website and your legal counsel instead of relying on an AI-generated summary.

The final checklist, in order

This is the version to print or paste into your workflow. It takes a few minutes for a short text and can be expanded for publications.

• Purpose: does the text do what was requested and answer the reader's real question?
• Claims: does every verifiable sentence have a source, or has it been removed?
• Numbers and dates: checked against the original document, in the right format for the reader?
• Names and links: correct spelling, links opened and working?
• Promises: nothing the company has not decided to commit to (deadline, warranty, discount, coverage)?
• Tone: does it fit the brand, the channel and the customer's moment?
• Confidentiality: no improper personal data, no internal information, no third-party passages?
• Signs of automatic text: repetition, excess symmetry, generic sentences and standard closings cut?
• Owner: is there a named person who approves and answers for the send?
• Record: is it saved which version was sent and who approved it?

If any item fails, the text goes back for revision. If it fails repeatedly at the same point, the problem is not the review but the prompt or the context material, and that is where it should be fixed.

An illustrative scenario and how to apply the method

Illustrative scenario: a small distributor uses AI to draft reply emails to customers about late deliveries. The draft arrives well written, with an apology and a new date. Applying the checklist, the person responsible notices that the delivery date was invented, because the model was given only the order number. They also notice the text mentions another customer in the same neighborhood as an example of "grouped deliveries". Both problems would have passed a quick read. With the date confirmed in the system and the mention removed, the email goes out.

Juan Carlo's working method for this kind of flow has four steps. Map: understand which texts the company produces, who approves them today and where the sources of truth are. Limit: define which data enters the tool, which claims the text may make and which subjects always stay with a person. Pilot: choose one real type of text, record what the review finds and agree on a success metric. Hand over: give the team the checklist and the record, and decide whether to expand, adjust or stop.

It is also worth defining what stays out of automatic drafting. The article on what not to delegate to AI in business helps separate what can be drafted from what should start human.

Frequently asked questions

How long does it take to review AI text with this checklist?

It depends on the length and risk of the text. A short customer service message can be reviewed in a few minutes, while a public article or a proposal takes longer, mainly because of source checking. Review time is part of the real cost of using AI and should count when you decide what is worth automating.

Can I ask the AI itself to review the text it wrote?

Yes, as support for finding contradictions, repetition and clarity problems. It does not prove that the facts are right, because it can firmly confirm something it invented. Checking facts and numbers against the original source remains a task for a person.

Do I need to tell readers that AI helped with the text?

There is no single answer for every case, and the decision depends on context, channel and industry rules. What matters in any scenario is that the company takes authorship and responsibility for the content. For publications, follow Google's guidance on transparency about how content was produced, and consult legal counsel when regulation is in doubt.

What should I do when the same error appears in many texts?

Treat it as a process failure, not a failure of each text. Check whether the prompt asks for something the model cannot know, whether reliable context material is missing, or whether the desired tone was never described with examples. Fix the instruction or the source, redo a few test texts and only then return to using the workflow day to day.

Sources

• OWASP: Top 10 for LLM Applications
• NIST: AI Risk Management Framework
• Google Search Central: conteúdo útil, confiável e voltado para pessoas
• ANPD: Autoridade Nacional de Proteção de Dados