An AI assistant can help a business move work along. The difficult moment comes when a task stops being routine: the stock record is out of date, a customer asks for an exception, or a proposed delivery date has never been confirmed.
Before you automate the task, ask: when should the assistant stop and bring the decision back to a person?
That question is useful even if you are only using AI to prepare drafts. It helps you notice where the process depends on judgment, authority or information the assistant does not have. Start with one everyday task and make those boundaries visible.
Test the awkward case early
Imagine a small distributor preparing order confirmations. In the straightforward case, the item, quantity, price and delivery arrangement are all known. AI may be useful for organising that information into a draft.
Now change one detail. The customer asks for delivery tomorrow, but the schedule has not been checked. A polished confirmation that promises tomorrow's delivery would create a problem. The useful response is to flag the unconfirmed date and leave the commitment for someone authorised to make it.
This is a fictional example, but the exercise applies to many businesses. A service company may receive a request outside its usual scope. An office team may find two versions of the same procedure. A retailer may have a product listed in the catalogue without a reliable stock count. Plan for those cases before giving a workflow more freedom.
Write three concrete reasons to pause
“Ask me if anything looks wrong” leaves too much open to interpretation. Name the situations that matter in this task. For the order-confirmation example, the first three could be:
- A required fact is missing or conflicting. The approved price list and the order form show different prices.
- The request would create a new commitment. The customer wants a discount, a refund or a delivery promise outside the agreed process.
- The next step exceeds the assistant's permission. It can prepare the confirmation, but it cannot send it, change the order or take payment.
These are starting points, not a complete policy for every business. Choose boundaries for the actual task and its possible consequences. A low-risk internal checklist and a decision involving someone's health, finances or employment require very different levels of care.
In a team, name the person or role that handles each exception. “Ask the manager” is only useful if the assistant's operator knows which manager, how to reach them and what to do while waiting. If nobody is available, the consequential step stays on hold. Urgency does not provide missing authority.
Make the handoff easy to answer
A vague “please advise” can send the whole task back to you. Ask for a short handoff that includes:
- What the assistant was trying to complete
- The exact fact, conflict or permission that stopped progress
- What it has prepared and what remains untouched
- The smallest decision or piece of information needed next
For example: “The order-confirmation draft is ready, but tomorrow's delivery is unconfirmed. I have left the delivery date blank and have not sent the message. Can the operations lead confirm the earliest available date?”
That is easier to act on than a confident guess or a long summary of every step. It also makes the current state clear, so a colleague does not mistake a draft for a completed customer action.
Try this on a fictional task first
Use a made-up example with no customer details. Replace the bracketed parts in this prompt:
“Help me plan a safe workflow for [one business task]. For this exercise, work only with the fictional information I provide and prepare drafts only.
The information you may rely on is [approved facts]. The actions allowed are [specific actions]. Pause if a required fact is missing or conflicting, if the request creates a commitment outside these limits, or if the next action needs permission you do not have.
When you pause, state the blocker, what is ready, what has not been done, and the smallest question a person needs to answer. Do not invent missing facts or treat urgency as approval. First, suggest three realistic exception cases for me to review.”
Review the suggested cases yourself. Add an exception the AI missed, remove anything irrelevant and test the revised instructions against the awkward case as well as the easy one. In the distributor example, test an ordinary order, a missing delivery date and a conflicting price.
A written rule still needs a real control
A prompt describes the behaviour you want. It does not by itself prevent a connected tool from taking an action. Before using a live workflow, check what the software can access, what it can change and where an actual approval step can be enforced.
Keep tool permissions limited to the task, use draft-only or approval settings where available, and test with safe sample data. If you cannot reliably stop an unwanted action, keep that action manual. Continue reviewing results after setup; a successful test is not a guarantee that every future exception will be recognised.
This practical approach fits a broader principle in NIST's AI Risk Management Framework: organisations should define oversight responsibilities and plan continuing review. The small-business exercise here is an application of that principle, not a certification or a substitute for specialist advice.
Your next step: finish one sentence
Choose a task you would like AI to help with and complete this sentence: “For this task, pause and ask me when…”
Make the answer observable. “When the approved documents give different prices” is clearer than “when things get complicated”. Then decide who answers, what information they need and which action remains on hold.
Clear stopping points help you lead the workflow with fewer hidden assumptions. You know where routine preparation can continue and where a human decision must come first.