The smartest bloke you could hire turns up tomorrow morning and has no idea who you are. Doesn't know your prices, never met your regulars. Couldn't tell you which supplier stiffed you back in March. He's sharp as anything. Reads a situation fast and gives you a clean answer to whatever you put in front of him. Then he clocks off and forgets the lot. Next day he's back, brilliant again, and its his first day all over again.
Thats generic AI. The best temp on the books.
You can brief him. You spend twenty minutes every morning explaining the job and the customer and the way you like things done, and he nods and gets on with it and does a decent job. But you're paying for those twenty minutes every single day, forever, because nothing sticks. He carries none of yesterday into today.
Now picture the foreman who's been with you twenty years. You say three words and he's already moving. He knows the quote you'd write before you write it, and which job on the board is about to go sideways. Knows old mate on Rundle Street pays late but always pays. He doesn't need briefing. The whole business is in his head, and every job he does adds another line to what he knows.
The gap between those two comes down to memory. Both are clever. What the foreman's got is your own data sitting in one place where the thing can actually reach it.
Erik Brynjolfsson ran a study on an AI assistant put in front of thousands of workers, published in 2025, and the finding was blunt. The tool worked by handing every worker access to the accumulated know-how of the best workers. Context is what made it useful. Generic context, the stuff any model already knows, lifts everyone the same little bit. Your context, your files and your history, lifts you specifically. One's a floor. The other's a moat.
And the memory does more than make the answers faster. It makes them true. There's a stack of research now on grounding a model in an organisation's own documents instead of letting it answer from whatever it half-remembers. In one clinical study the wrong answers dropped from 8 percent to zero when the model was made to answer from the source material in front of it rather than reaching into its own head. Eight percent to nothing. Same model. The only change was giving it the real documents to stand on.
Every business owner has felt that 8 percent. The confident answer thats just wrong. The quote that's off because the person didn't know about the deal you struck last year. The foreman doesn't make that mistake, because he's not guessing. He's reading off twenty years of the actual business.
So that's what an embedded brain actually is. You take the temp, the clever one, and you pour your whole business into him. Every job, every price, every customer note, every lesson you learned the hard way. He stops guessing and starts remembering. And unlike the temp, he keeps it. Tomorrow he's still got today.
The part that matters most is what happens over time. Your competitor can buy the same model off the shelf this afternoon. Same brains, same temp. What he can't buy is your twenty years of jobs, quotes, customers and hard lessons, because that only exists in your shed. The pooled memory of your business is the one thing nobody else can download, and it gets deeper every week you run.
Anyone can hire the temp. Only you can build the foreman.