Generative AI tools can produce a plausible draft of almost anything — an email, a summary, a spreadsheet formula, a policy — in seconds. The effect on everyday work is real. It also changes where the skill lies. When the first draft costs almost nothing, the value moves to two things: asking for the right output, and deciding whether what comes back is fit to use.
Specify the task
The quality of an output depends heavily on the quality of the request. A useful prompt gives the tool what a capable colleague would need:
- the task, stated precisely — "summarise the three main risks for a board audience in under two hundred words", not "summarise this";
- the context — who the output is for, what they already know, and any material the tool should work from;
- the constraints — length, tone, format, what must be included and what must not;
- the form of the answer — a table, a list, a paragraph, a formula with an explanation.
Requests like this take a minute longer to write and save much more than that in correction.
Treat every output as a draft
Generative AI produces text that is fluent whether or not it is accurate. It can state an incorrect figure with confidence, invent a reference that does not exist, apply a rule that does not hold in the situation, or quietly drop part of the instructions in a long task. None of this is signalled in the output.
So every output that will be relied on needs checking, and the depth of the check should match the consequence of an error:
- Facts and figures checked against the source, not against the tool's own restatement of it.
- Calculations and formulae tested with known inputs, or recalculated independently.
- References, citations and quotations confirmed to exist and to say what the output claims.
- Reasoning read critically: does the conclusion actually follow, and does it fit what you know of the situation?
The person who uses the output is responsible for it. "The AI wrote it" is not a defence in front of a client, a regulator or a manager.
Protect what should not leave
Anything entered into an external AI service may be stored, processed or reviewed outside the organisation. Personal data, confidential financial information and client material should only be used with tools the organisation has approved for that purpose, under terms it has reviewed. When in doubt, remove identifying details or do not use the tool.
Know when not to use it
Some tasks are poor fits: those where an error would be serious and hard to detect, those requiring information the tool cannot have, and those where the value lies in the thinking rather than the text. Drafting a difficult judgement can help clarify it, but the judgement itself remains a human responsibility.
The skill that lasts
Tools will keep changing. The underlying competence — defining a task clearly, verifying what is produced, protecting sensitive information and taking responsibility for the result — applies whichever tool is in use. It is the part of working with AI that makes the output something others can rely on.