Responsible AI at Work: A Pre-Use Checklist

An AI-assisted draft can look ready long before it is ready to use. The language may be clear, the structure convincing and the recommendation confident. Yet a date may be wrong, an important condition missing or a promise outside your authority.

Responsible AI at work means treating that draft as something to review, not something that approves itself. Before sending it to a customer or using it to support a decision, ask: Is this an appropriate task? Was the information suitable to enter? Which claims have actually been checked? Who can approve the result?

This guide offers a practical review routine. It does not replace your organization's policies, specialist advice or the permissions required for a particular task. If you are still developing your approach to prompting and questioning, begin with AI as a thinking partner.

1. Decide whether the task is suitable

Start with the consequence of getting the answer wrong. Exploring possible headings for an internal presentation is different from drafting a contractual commitment or advising someone on a consequential decision.

Ask what the draft will be used for, who will rely on it and what review the task requires. If you lack the expertise or authority to check it, identify the appropriate reviewer before proceeding. An AI-generated suggestion does not remove that requirement.

For example, you might use an approved tool to explore alternative ways to explain a routine process. You should not treat its answer as permission to change the process or promise a customer a new service.

2. Check the information before entering it

Review the input, not just the output. Does the task require customer details, confidential documents or commercially sensitive information? Do your organization's policy and the approved tool's arrangements allow that information to be used for this purpose?

If permission is unclear, pause. A fictional example may be sufficient for exploring tone or structure. Removing a person's name alone should not be treated as proof that information is safe to share.

Keep the initial exercise simple: describe an invented customer asking for an update, without copying a real message or account history. This lets you explore a reply structure without introducing actual customer information.

3. Verify claims and missing context

Read the draft sentence by sentence. Mark statements that need checking: dates, prices, availability, technical details, quotations and descriptions of policy. Confirm them against the relevant current source or responsible person.

NIST's Generative AI Profile, section 2.2, identifies the risk of confidently presented false content, including misleading citations. A fluent explanation is therefore not enough to establish that a claim is correct.

Also ask what has been left out. Does the customer need to know about a condition? Does the draft imply certainty where the team has not made a decision? Asking AI to review its own answer can generate useful questions, but it is not independent verification.

4. Use the responsible AI at work checklist

Use these six checks before relying on a draft:

• Task suitability: Is AI assistance appropriate for this task and its consequences?
• Information permission: Was the input allowed for this purpose and tool?
• Factual accuracy: Which claims have been checked, against what source?
• Missing context: Are relevant conditions, uncertainties or limits visible?
• Tone and commitments: Is the wording suitable, and are promises authorized?
• Responsible reviewer: Who has the knowledge and authority to approve use?

Copy this short record for each check:

• Check:
• Status: checked / unresolved / not applicable, with a reason
• Evidence or explanation:
• Next action and owner:

An unresolved item is a reason to investigate, revise or stop—not a box to tick for completeness. Record enough detail that another reviewer can understand what was checked.

Illustrative example: a customer update

This is a fictional exercise, not a Tibyan client case. An AI-assisted email says: “Your order will arrive on Thursday, and we will refund the delivery charge.” Neither statement has been confirmed.

The factual check flags the arrival date; the commitments check flags the refund. The reviewer removes both promises and prepares a holding reply: “We are checking the delivery status and will provide an update once it is confirmed.” Before sending, the reviewer still checks that this wording follows the team's actual process.

The lesson is not that one alternative sentence is always correct. It is that useful review separates facts, assumptions and authorized commitments.

5. Decide what happens next

A review can end in approval, revision, specialist review or a decision not to use the output. Keep that choice explicit. If a material fact remains uncertain, do not hide it behind smoother wording.

Start with one appropriate, low-risk drafting task using an approved tool and permitted information. Apply the checklist, record what needed correction and consider whether AI assistance was useful for that task. For broader learning, explore AI for Work & Business.

If your team wants to practise prompts, output review and human judgment together, explore Practical AI at Work and discuss your learning needs with Tibyan. Scope and delivery are agreed before a program begins.