You bought the AI. Why is nothing faster?
Plugging in AI doesn't make a business faster. Here's the four-part harness that does, and the four questions that tell you whether a job is ready to hand over.
You paid for the tools. Everyone has a login. And your week looks about the same as it did in March.
That isn't a model problem. Any serious AI can draft a quote. What it can't do is know your prices, your rules, and who reads the thing before it reaches a customer.
The model is the engine. The harness is everything around it that turns motion into finished work. Almost nobody writes the harness down, which is why almost nobody gets their week back.
The test that makes this obvious
Picture the sharpest graduate in the country. You hire them on Friday. Can they start on Monday?
Not really. Not because they aren't capable, but because nobody has told them what the business does, which files they may open, what they must never send without asking, and who carries it when something goes wrong.
You'd never expect a new hire to work without that. AI is the same, with one difference: it won't tell you it's confused. It'll produce something confident and wrong.
Even the people selling the engines say this part is yours
At the G20 Innovation Ministerial in Chapel Hill in September 2026, Nvidia's Jensen Huang made the case that AI safety should sit with companies rather than public officials. His words: "Every single company has to take it upon themselves to develop the technology safely." (Bloomberg, 2 September 2026)
He was talking about frontier labs and national policy. The same sentence lands on a five-person business in Bentleigh. Nobody writes your rules, your access list or your sign-off for you. That part is the harness, and it's yours whichever model you rent.
What a harness actually contains
Four parts. None of them are technical, and all four are things only you can write.
1. The goal
One job, named plainly. Not "help with admin", but "turn site photos and job notes into a draft quote in our format". A vague goal produces vague output every time.
2. The rules
What it may say, what it must never say, and the numbers it has to use. This is the part almost nobody writes down, and it's the part that decides whether the output sounds like your business or like everyone else's.
3. The access
Which files, inboxes and systems it can read, and which it can't touch. Narrow on purpose. An assistant that can see everything is a risk, not a feature.
4. The sign-off
Who reads it before it leaves, and what happens when that person is away. If nobody would notice a wrong answer going out, the job isn't ready to hand over yet.
Which is why "which AI is best" is the wrong question
A new model tops a chart every few weeks and the internet argues about it for a day.
For the work most small businesses actually want handled, the gap between models is smaller than the gap between having your price list written down and not having it. A cheap fast model with the rules beats an expensive slow one guessing.
And why the thing you keep isn't the AI
You rent the model by the month. It gets more expensive, and it goes down on the day you need it.
What stays yours is the harness: the written rules, the workflow, the examples of how you actually write, and the mistakes you already paid for. That's the asset, and it moves with you to whatever model comes next.
Four questions before you hand a job over
Does it happen more than once a week? Could you explain the rules to a new hire in five minutes? Would you notice if it came out wrong? And does someone sign off before it leaves?
Fail the last two and automating the job makes the problem faster, not smaller. That's the part people skip.
We've written the harness out job by job for the seven a small Australian business actually runs: quoting, enquiries, marketing, the numbers, rosters, documents and onboarding. See what AI can actually do in a business your size.
Where to start
Not with the tool. With the job that eats your week.
Write down the four parts for that one job. If you can't fill in the rules or the sign-off, that's the finding, and it's worth more than another subscription.
That's what a diagnosis day is: we map the job with you, build one workflow, and check the result together. Sometimes the honest answer is that a job isn't worth automating at all, and we'd rather say so on the day.
