How to make AI-assisted work trustworthy.

You’re reading The AI Directive — practical AI intelligence for business leaders. Subscribe here if someone forwarded this.

I use AI extensively. I give it source material, explain the history, correct what it has misunderstood, challenge weak conclusions and keep going until the output faithfully represents what I think.

It is still dramatically faster than beginning with a blank page. That is the point.

But the fact that I have worked this way may be invisible to the person receiving the final result. They see the document or the slides, not the hours of context, challenge and judgement behind them. If the presentation also has the familiar look of AI, they may reasonably wonder whether any of that thinking happened at all.

That has been the thread through the last two articles. #024 looked at the suspicion created by recognisably AI-generated presentations. #025 looked at the human work that should happen before AI turns the thinking into polished output.

So how do we bring those two things together? How do we use AI for the productivity it offers without asking colleagues, customers or senior leaders to take the quality of the work on faith?

I think the answer is surprisingly simple. For any material piece of AI-assisted work, I want five questions answered.

The quality of AI-assisted work starts before the first draft.

What data did you provide? Which documents? What organisational history, commercial reality or customer context did you explain? What was the work actually meant to help someone decide?

If the answer is “I wrote a prompt”, that is not enough. A well-written prompt can produce a polished response, but it cannot create knowledge that was never supplied.

This is one of the biggest differences between casual AI use and serious AI-assisted work. The person has to bring the ingredients that make the answer relevant to the real organisation. AI can structure and synthesise them at extraordinary speed. It cannot know which awkward fact everybody inside the business already understands but nobody thought to mention.

I rarely accept AI’s first interpretation of a substantial problem. It usually has part of the picture, but not all of it.

The useful work happens in the exchange that follows. I tell it what it has missed. I push back when the recommendation is generic. I ask it to test the opposite view. I add experience it could not have inferred and make it revisit conclusions that feel too neat.

That is not editing. It is where the human judgement enters the work.

If I am reviewing somebody else’s AI-assisted recommendation, I do not particularly care how many of the final words they typed. I care what changed because they were involved. Which assumption did they reject? Which piece of context changed the answer? Which conclusion did they choose—and why?

Those questions reveal far more about authorship than asking whether AI wrote the draft.

This is where the presentation problem returns.

Good thinking can be undermined by generic packaging. Familiar AI imagery, repeated slide patterns and frictionless language can make serious work look as though it came from a single prompt.

The answer is not to roughen it up so it passes as human-made. That is theatre. The answer is to make it genuinely specific to the organisation, the audience and the decision.

Use the real template. Use language people in the business would recognise. Replace generic examples with relevant ones. Remove the slides that add polish but not meaning. Make sure the person presenting the work can explain every conclusion without hiding behind the words on the screen.

That is not concealing AI. It is stopping the production method from overwhelming the point of view.

AI can make a weak answer sound confident. Senior leaders should be particularly wary of this because confidence and fluency are exactly what make material easy to consume at speed.

The checks should follow the consequence. I do not need an audit trail for a brainstorm or a first-pass meeting summary. I do want the numbers recalculated, the important claims traced to their sources and the assumptions tested before an AI-assisted recommendation influences a customer, a board decision or a significant investment.

This does not need to become a compliance industry. It is the same judgement we already apply to human-produced work: the greater the consequence of being wrong, the stronger the evidence should be.

The difference with AI is that an answer can arrive looking finished long before it is ready to be trusted.

This is the final and most important question.

Someone must be prepared to say: I understand the evidence, I challenged the reasoning, I know what remains uncertain and this is the conclusion I recommend.

“A human reviewed it” does not meet that standard. Review can mean anything from a careful interrogation to a quick glance before the meeting. Ownership means being able to defend the work when the conversation moves beyond the slide.

It also means being candid about the limits. Trust is not created by pretending the answer is certain. It is created when the owner can explain where the evidence ends, what has been assumed and what would cause them to change their view.

I would not require every prompt or draft to be stored. Nobody needs a forensic archive of the conversation with AI.

For consequential work, I would expect a short record of six things: the purpose, the key sources, the material human challenge, the checks performed, the remaining limits and the person who owns the conclusion.

That can fit on one page. In many cases, it can fit in the final document itself.

The point is not to prove that AI was used responsibly according to a grand policy. It is to make the work understandable to somebody who was not present while it was created.

That is how I think AI-assisted work earns trust. Not by looking handmade. Not by minimising how much AI contributed. And not by attaching a senior name to the final page.

It earns trust when a person can show what went in, explain what they changed, demonstrate what they checked and stand behind what came out.

AI can do a great deal of the production. The human contribution moves upstream into context, challenge and judgement—and downstream into accountability.

That is not a dilution of authorship. Used properly, it is a better use of it.

The next time a material piece of AI-assisted work reaches you, ask five questions:

If the answers are clear, the fact that AI helped produce the work should not count against it.

This is the final article in a three-part series on earning trust in AI-assisted work.

The standard is not “human-made” or “AI-made”. It is whether the work carries enough human context, judgement and ownership to be relied upon.

See you Tuesday.
— Toby

These are tools I use personally. Affiliate links marked - I earn a small commission if you sign up, at no extra cost to you.

Some links in this issue are affiliate links. I only recommend tools I actually use.