Why AI Keeps Guessing About Your Business

You ask AI for something. The answer arrives in seconds and it is fine — competent, well organised, nothing actually wrong with it.

Then you spend twenty minutes making it sound like you.

That twenty minutes is the subject of this post, because it is now the most reliably measured thing about AI at work, and almost nobody talks about what it actually is.

The hours arrive, and then they leave again

In research published by Foxit, senior executives estimated AI saved them 4.6 hours a week. They also reported spending 4 hours and 20 minutes a week validating what it gave them.

Net gain, once the checking is counted: sixteen minutes. The people below them did worse, landing about fourteen minutes a week behind where they started.

Workday's research found the same shape from another angle — roughly 37% of the time AI saves is given back to correcting and rewriting low-quality output. For every ten hours gained, nearly four go straight back out.

Sage, surveying over 2,000 senior finance leaders, found 26% of expected gains lost to reverse-engineering AI output so it could be explained to someone else — and 22% said it swallowed more than half of everything AI saved them.

Almost everyone in these studies reports feeling more productive. The stopwatch disagrees.

The tool is not broken

The obvious conclusion is that the models are not good enough yet. It is worth being precise about why that is wrong.

Ask for a welcome email for new clients and you will get a welcome email for new clients. It opens warmly, says you are glad they are here, gestures at next steps, closes politely. There is nothing wrong with it.

It is a perfectly good welcome email for the average of everyone who has ever needed one.

That average is exactly what you asked for — not explicitly, but by omission. You did not say who your clients are, what worried them before they signed, what you promised them, what you never say, or how you actually sound. So the model did the only thing available to it and answered for the median business, writing to the median client.

A generic answer is not a failure. It is the correct answer to a question with nobody in it.

What the rewriting actually is

Watch yourself do the twenty minutes and notice what you are doing. You are not fixing errors. You are supplying facts.

You cut the line about being excited, because you do not talk like that. You add the detail about the two-week onboarding, because that is how your business works. You swap "clients" for "members", because that is the word you use. You delete a promise you would never make.

None of that is correction. It is context — arriving late, one message at a time, typed by hand, and thrown away the moment you close the window. Tomorrow you will supply it again.

That is the real tax. Not the model getting things wrong, but you paying in instalments for something you could have said once.

Why better prompts do not fix it

The usual advice is to write better prompts. It helps, and it does not solve this.

A prompt describes the task. The gap is in the briefing. You can write an extraordinarily good prompt and still be answering for the median business, because nothing in it says who is asking.

Prompt libraries have the same problem twice over — they encode somebody else's business, audience and voice, which is why output from a shared prompt pack tends to read like everyone else using that pack.

And waiting for better models will not do it either. A more capable model with no information about you produces a more fluent version of the average.

The four things worth writing down

The fix is unglamorous: write down what AI needs to know about you, once, and keep it where the work happens.

1. Who you are — what you do, for whom, and the opinion that separates you from the average person in your field. "I'm a business coach" tells a model nothing. "I coach solo consultants who are earning well and working too many hours, and I think most of them have a pricing problem dressed up as a time problem" tells it everything.
2. Who you are talking to — their situation, and crucially what they already know. This is what stops AI padding every draft with introductions your reader outgrew years ago.
3. How you sound — in observable patterns, not adjectives. "Professional yet friendly" is not an instruction; every business on earth would pick those words. "Short sentences, no exclamation marks, I say 'people' not 'individuals', I never use 'unlock' or 'leverage'" is something a stranger could actually follow.
4. What you will not say — the guardrails. Almost nobody writes these down, and they may be the most useful part, because models drift toward the confident and promotional unless told otherwise.

Add two or three samples of your own work, a sentence on why each one is good, and one deliberate counter-example — something plausible that is not you. Contrast teaches faster than praise.

The test that settles it

There is only one way to know whether any of this worked, and it takes half an hour.

Pick a task you do often. Do it the way you do it now, and time how long you spend editing the output. Then do the same task again, same prompt, with your context in front of the model, and time the editing again.

The difference is the whole argument. And every correction you still have to make is a line your context is missing — which is the most useful thing the exercise gives you.

Where to start

We have written this up properly as a course: Give AI the Context It's Missing. Seven short chapters, and the first two are free with no card — the rework numbers, and the reframe that explains them.

The rest walks through building the pack itself: the four components with worked examples, how to choose samples, where to store context so it applies automatically, what to leave out (too much context makes answers rigid and samey), and how to keep it true as your business changes.

It takes an afternoon. It is the difference between an assistant that knows where it works and one that meets you for the first time, every single time.