Skip to main content
Insights

May 6, 2026

A five-minute check before you send AI output to a customer

A short habit that catches wrong numbers, fake citations, and tone problems—without slowing you down much.

This is the same image social apps use for link previews (generated from the title and description, not a separate photo asset).

Generative tools are fast. That speed is the risk: a wrong fact or a promise you cannot keep can reach a customer in seconds. You do not need a twenty-step review for every draft. You need a small checklist your team can remember—and a culture where skipping it is the exception, not the rule.

When to use this

Use the checklist for anything customer-facing: email, chat, proposals, invoices, policy explanations, and public posts. For internal drafts, shorten it—but still catch numbers and confidentiality.

1. Numbers and names

Skim for anything that looks like a statistic, date, price, or proper noun. If the text includes numbers you did not provide, verify them or remove them. Models often sound precise when they are guessing.

If the draft mentions a law, standard, or product version, confirm it—or soften the language (“as of our last review…”).

2. “Sources” that do not exist

If the draft cites a report, article, or regulation, open the source or delete the claim. Fake citations are common; they are also one of the fastest ways to lose trust.

If you keep a citation, make sure the link works and the quote matches the page.

3. Promises and guarantees

Highlight anything that sounds like a legal commitment: timelines, SLAs, refunds, compliance. Those belong to your real policies—not to a model’s confident sentence.

Watch for “we will always” or “we guarantee”—unless legal approved that language.

4. Tone and audience

Read the last paragraph out loud. Does it sound like your firm talking to this customer, or like generic marketing? Adjust phrasing until it matches how you actually speak.

If the customer is upset, check that the draft acknowledges that before jumping to solutions.

5. Sensitive data

Confirm you did not paste passwords, private IDs, or confidential terms into a tool your company has not approved for that data class. When in doubt, use redacted examples.

If the draft includes another customer’s details, stop—rewrite from scratch.

6. Handoff clarity

If more than one person touches the thread, state who owns the next step. Ambiguous ownership creates delays that look like AI’s fault when they are not.

Teaching the habit

Post the checklist where replies get sent. Pair it with one real example of a catch your team made—people remember stories.


For a deeper take on when automation is a bad fit, read when not to use AI. For definitions of terms like “hallucination,” see the glossary.