Know when AI is the wrong choice.
AI can prepare useful work and still produce an answer that sounds convincing but is wrong. The right question is whether the task gives you a reliable way to notice and correct that error before it matters.
When the source of truth is unclear.
If the team cannot agree which price list, policy or booking record is current, adding an assistant will not settle the disagreement. It can turn an outdated answer into fluent text.
First decide where the current information lives, who maintains it and how changes reach the people using it. A short, maintained reference may solve more than a larger tool.
When the output commits the business.
A reply can promise a date, accept work or change a price. Those are business decisions even when they appear inside an ordinary email. A draft may be useful; an unchecked send can cross a boundary nobody intended.
Separate preparing information from deciding and acting. Name the person who checks consequential details, and preserve their ability to stop the action.
When checking takes longer than doing.
Illustrative example: an assistant turns three short appointment notes into a summary, but a person must open every note, correct the dates and remove an invented detail. The new process has added work rather than removed it.
Try several representative examples. Count preparation, checking and corrections together. If the benefit is better formatting or consistency, describe that benefit honestly instead of calling it time saved.
When a simpler method is more dependable.
Some tasks are exact: route a message with a known subject, total agreed values or remind an owner that a date has arrived. A clear rule can be easier to inspect and maintain.
- Exact, stable condition
- Consider a rule or software setting.
- Interpretation with review
- Consider an AI draft supported by the original information.
- Unclear or consequential judgment
- Keep a person responsible for the decision.
- Rare task with many exceptions
- A checklist or manual process may be enough.
When you cannot explain the fallback.
Before using an assistant in live work, ask what happens if it cannot find the answer, returns conflicting information or stops working. Someone should be able to pause it and continue the task another way.
Start smaller if there is no clear answer. A useful assistant has a defined job and a visible limit; it does not need to handle every request the business might receive.
