AI for business · Practical guide
Draft from available facts and review every commitment.
Create a customer-service draft without sending it. A fictional order-status query helps distinguish what you can state from what you need to check first.

01Supply a small, explicit context
Fictional query: “When will order P-42 arrive?” Confirmed information: the order is recorded but has no dispatch date. The carrier has not been consulted. No delivery date is confirmed.
Separate the customer’s question from internal facts. A wish to receive the order tomorrow must not become evidence of a planned delivery tomorrow.
02Request a draft with specific boundaries
Reusable instruction: “Prepare a short, polite reply using only confirmed facts. Promise no dates, compensation or unagreed actions. Separately identify missing information and what a person must confirm before sending.”
Keep personal data out of the exercise. For real cases, use the minimum context authorised for the tool and retain the sending controls agreed by your team.
03Compare with a fact-limited reply
Possible draft: “Order P-42 is recorded. We do not currently have a confirmed dispatch date, so we cannot provide a delivery date.” The team can add a next step once it has agreed who will take it and under what conditions.
“It will arrive tomorrow” and “We have contacted the carrier” are unsupported. “We will update you within an hour” also introduces an absent commitment. Review verbs and deadlines, not just tone.
Reference [1]: NIST · AI 600-1
04Record an approval criterion
Review record: each fact has support; no new promises; open questions stay outside the final message; recipient and reference are correct; a person authorises sending according to the team’s process.
Repeat with a confirmed delivery date and with conflicting references. The draft should change when evidence changes rather than always producing the same reassuring phrase.
Sources and further reading
These references expand on the concepts indicated. The examples and exercises are original editorial material.
[1] NIST · AI 600-1
Generative AI Profile ↗
Context on unsupported content.
Back to the related section
Frequently asked questions
Can these drafts be sent automatically?
This exercise does not validate automatic sending. Start by reviewing facts, recipients and commitments. Connecting actions requires separate definitions and tests for permissions, errors and oversight.
Keep reading
Learning AI
How to check an AI assistant’s answers before using it ↗
Build a simple test with questions, evidence and acceptance criteria. Use a worked exercise to find unsupported answers before relying on an assistant.
AI for business
Rules or AI? How to choose for an automation task ↗
Break down a business task, identify explicit rules and test where AI may help. Work through a request-routing example with a practical decision checklist.
AI for business
How to document an automation so someone else can maintain it ↗
Write an operating note covering inputs, rules, owners, failures and recovery. Test the handover with a practical exercise that reveals missing information.