Guides · October 2, 2026
How AI works, explained in jobsite terms
Short answer: An AI model is like a skilled tradesperson who forgets yesterday unless it's written down. It's fast and capable, but it guesses when it's missing information. What turns it into a reliable coworker is the setup around it: the plans, the posted rules, a written scope, and an inspector before anything goes out.
The five steps from chat app to jobsite
- Know who you hired. The model is the tradesperson. A hosted model (most chat apps) is like a crew from a labor agency: their terms, billed by use. A local model is your in-house crew. Nothing leaves the shop, but the crew is smaller.
- Hand over the plans. The context window is the box of screws: only so many fit at once. Give it this job's files (the plan set and spec book), not everything you own. Long jobs lose early details, so keep handoff notes like a shift turnover report.
- Post the rules. A short rule file works like the company's procedures and QA manual. Write your standards once (company name, pricing, markup, what always needs a permit line) and every session reads them.
- Write the scope. An agent is a subcontractor with a written scope of work. A skill is a standard detail or install procedure, pulled when that task comes up. Vague scope gets a guess. A clear scope gets the job.
- Add an inspector and a hold point. A checker is the third-party QA inspector, because the crew doesn't inspect its own work. A gate is an inspection hold point: nothing sends, posts or pays until someone signs off.
The same job, two ways
Ask a chat app to "write a change order for 3 extra outlets" and you get a clean-looking form with a made-up price, no markup and no permit line. Give the same AI your rule file, price list and change-order template, and it fills them in, lists what it couldn't confirm, and waits for your yes. Same AI. Different setup.
Example scenario for illustration, not a measured result.
Quick translation table
| AI word | On the jobsite |
|---|---|
| Model | A skilled tradesperson who forgets yesterday unless it's written down |
| Chat app on a phone | The walk-up front counter: quick answers, no jobsite |
| Context window | The box of screws: only so many fit at once |
| Memory and files | The job binder and daily logs |
| Connection (MCP) | A standard utility hookup: any compatible tool plugs in |
| API | A supplier account: direct orders, an account number (the key), billed per order |
| Checker | Third-party QA inspector |
| Gate | An inspection hold point |
Where the comparison breaks
Real crews remember last week and push back when a plan looks wrong. AI only remembers what's in the job binder, and it can be confidently wrong. Anything a client, inspector or bank will see gets checked against a real source.