Induct AI like a new hire, not a search box

If you hired someone, told them nothing about how the business works, gave them access to no files, and introduced yourself again every morning, you would not blame them for the work. That is close to how most small businesses use AI, and it explains the generic output far better than any theory about model quality. The fix is not a better prompt. It is an induction: what good looks like, what it can reach, and what it keeps between sessions.

Why does AI keep giving me generic answers?

Because generic is the only honest answer to a question with no context in it.

Ask a capable stranger to "write a quote for a bathroom renovation" and they will produce something plausible and wrong: wrong inclusions, wrong tone, wrong assumptions about what you do and do not cover. Not because they are incapable, but because you did not tell them anything. They filled the gaps with the average of everything they have ever seen, which is exactly what generic means.

So the owner tries a better prompt. Gets a slightly better generic answer. Tries again, adds more detail, gets closer, and eventually gets tired of typing the same context every time and stops.

Prompting harder is the wrong lever. It is just the only one visible inside a chat box, which is why everyone reaches for it.

The three induction questions

Here is the reframe. Stop treating it as a search box and start treating it as somebody's first week. Not because AI is a person, it plainly is not, but because the induction checklist you already have happens to be the right checklist.

1. What does good look like, and what does bad look like?

This is the standard. It is the one everybody skips, and it is where nearly all of the value is.

Writing down what good looks like is genuinely hard, and nobody has ever made you do it, because the standard has always lived in your head and been applied by eye. You know a bad quote when you see one. You have never had to say why.

An induction forces the saying. What must always be in it. What must never be. What tone. What we do when the information is incomplete. Three or four real past examples of the good version and, if you can bear it, one of the bad version with a note about what went wrong.

This is also the most transferable thing you will produce. The page you write for the AI is the page you would hand a new employee, and it keeps paying off long after the AI conversation has moved on. We go deeper on the worked examples half of this in training AI with your own examples.

2. What can it reach, and what can it definitely not?

This is access, and it is where most of the practical benefit is unlocked and most of the risk lives.

An assistant that cannot see your price list cannot price anything. An assistant that can see your entire drive can see things it has no business seeing. The useful question is narrow: what is the smallest set of files and systems that lets this job be done properly?

Start there rather than at the other end. It is far easier to widen access later than to unpick it, and the answer to this question also decides whose login the thing runs under, which we treat properly in AI assistant access and permissions.

3. How does it pick up where it left off?

This is continuity, and it is the quiet one.

If every session starts from nothing, you are paying the setup cost repeatedly and never accumulating anything. The useful version keeps the standard, the examples and the recent decisions in a place it can read each time, so the second month is better than the first rather than identical to it.

In practice this is less exotic than it sounds. It is usually a document, in a folder, that the tool is pointed at. The sophistication is in what is written there, not in the mechanism. Giving AI business context covers what belongs in that document.

Which of the three matters most?

The first one, by a distance, and it is not close.

Access and continuity are plumbing. They are fiddly, they matter, and a competent person can set them up in an afternoon. The standard is the thing only you can supply, and it is the thing that determines whether the output is worth having.

Most of a first session at Handiwork is not technical at all. It is getting a standard out of an owner's head and onto one page, by asking about specific past jobs until the rules that were being applied by instinct become sayable. Owners often find this the least expected part of the engagement and the most useful, because they end up with something they needed anyway.

Once that page exists, the tool is the easy part. So is the next person you hire.

What this looks like in practice

Pick the single job you most want help with. Not the biggest, the most repeated.

Write one page. What a good version looks like, what it must never do, where two or three real past examples live. Half an hour, not a project.

Point the tool at that page instead of typing the job out fresh every time.

Then judge the output. If it is still wrong, you now have something far more useful than frustration: a specific gap between what you wrote down and what came back, which tells you what was still living in your head.

That loop is the work. Everything else is arranging tabs.

This is also the honest answer to a pattern we see constantly, which we wrote about in why AI projects fail. Most businesses that say AI did not work for them never set it up. They tried it. Those are different things, and only one of them is a test.

If you want the wider adoption picture, our pillar on the next step, not more tools covers where this fits, and our services page sets out what a first session actually involves.

Ready to find out where you stand?

Our free AI Readiness Check takes about five minutes and helps you work out which job in your business is worth writing the standard for first: start the AI Readiness Check.