A polished draft can still contain the wrong date, an invented detail or a promise your business cannot keep. When AI helps prepare work, a clear review process makes it easier to spot those problems before the output is used.
This is a known limitation of generative AI. The US National Institute of Standards and Technology describes how these systems can confidently produce incorrect information, including unreliable reasoning or citations. Fluent language alone is not evidence of accuracy. Read NIST's Generative AI Profile.
For a business owner, the practical question is straightforward: what needs checking, and who is able to check it?
1. Agree what the tool may do
Write a simple boundary before starting. For example: “AI may prepare a draft response using the approved information sheet. A team member must check it before it is sent.”
Be specific about actions that remain with a person. These might include confirming availability, setting prices, promising a refund, publishing a claim or changing a customer's record.
Choose the amount of review according to the consequences of a mistake. Brainstorming an internal meeting title needs less scrutiny than sending a customer a quotation. Work involving health, legal rights, finances or employment decisions needs specialist consideration and appropriate oversight.
2. Start with information you trust
Give the tool the source material relevant to the task. That might be an approved product sheet, a current service description or a cleaned set of business data you are authorised to use in that system.
Check the source itself. An out-of-date document can produce an answer that looks well supported but is still wrong for today's customer.
Use only the information necessary for the task. Before entering personal or confidential material, check that the tool, account settings and intended use are approved by your business. Use fictional examples while testing if you are unsure.
3. Check facts against the original source
Review the output alongside the source, rather than relying on memory. Pay particular attention to:
Names, dates, prices and quantities.
Product specifications and service limits.
Calculations, percentages and units.
Availability, delivery and refund statements.
Claims about customers, qualifications or results.
Open any cited sources and confirm they support the claim. Asking the tool whether its own answer is correct is not a substitute for checking the evidence.
If the source does not contain an answer, the draft should identify the gap. It should not invent a plausible detail to make the response feel complete.
4. Review meaning and business judgement
An answer can be factually accurate and still be unsuitable. It might sound dismissive, omit an important condition or imply a commitment you did not intend.
Read it from the recipient's point of view. Does it answer the actual question? Is the next step clear? Does it sound like your business? Would a reasonable customer interpret any wording as a promise?
The reviewer needs enough knowledge of the subject to challenge the result. Give them access to the source and permission to reject or rewrite the output.
A fictional example
Imagine an events business using AI to draft a reply about a venue booking. This is an illustrative example, not an Amazency client case study.
The tool has a general brochure and a customer's proposed date. It writes an attractive reply saying that the date is available and an evening extension is included.
Neither claim is supported by the brochure. A reviewer should remove them and check availability and the applicable package before replying. Better instructions would require the draft to list those unanswered questions instead.
The correction also improves the process: the team now knows that availability and package inclusions need a separate check.
5. Record recurring errors and retest
Keep a short note of mistakes worth learning from. Record the task, what went wrong and the correction needed. Use those examples to improve the source material, instructions or review checklist.
Retest when the tool, source information or workflow changes. Stop using the output when a critical check cannot be completed or the same consequential mistake keeps recurring.
NIST's AI Risk Management Framework supports clearly assigned responsibilities, ongoing monitoring and documented review. The checklist here is a practical starting point, not a claim of certification or complete compliance with that framework. Read the NIST framework's core guidance.
Make approval an explicit step
Before work leaves the business, someone should be able to say what they checked and why it is ready. That habit helps turn AI from an occasional experiment into a process your team can understand and manage.
Looking for a practical first project? Talk to Amazency about the work you want to improve and the checks that matter in your business.