The work that never leaves the tab
Three weeks ago you asked for a follow-up sequence for the customers who got a quote and went quiet. What came back was good — better than what you'd have written on a Tuesday afternoon.
It's still sitting in that chat window. Nothing was sent. Nobody was called. The document that would have made you money is a document.
This is the most common place AI stalls in a real business, and it has nothing to do with how you asked. The tool did its job. It just has no hands.
A harness is the wiring, not the AI
The term for this is harness engineering, and in plain English it means what your AI is allowed to touch.
Not the model. Not the prompt. The wiring around it: which of your systems it can look at, which it can change, and where it has to stop and wait for a person. Two shops can run the identical tool with identical instructions and get wildly different value out of it, entirely because one wired it to something and the other didn't.
Read, write, send
Every harness — for any task, in any business — comes down to three questions. Ask them in this order.
- What can it read? Your price list, the on-call calendar, job history, the inbox. Reading is where the value starts and it's the safest thing you'll ever let it do. Nothing changes when it reads.
- What can it write? A draft in the CRM. A row in a sheet. Note the word draft — writing means it can create something, not that anything happens because it did.
- What can it send or trigger? An email to a customer. A text to a tech. A charge on a card. Every guardrail you will ever need lives here, and the right answer is usually nothing, yet.
Most people jump straight to the third question, get nervous, and abandon the idea. The order exists to stop that.
One night on the dispatch board
It's July, it's Midland, and overnight you accumulated the usual pile — voicemails, texts, two emails from a property manager.
A harnessed system reads the voicemail transcripts, the on-call calendar, and the job history for every number that called. It writes a ranked callback list into the dispatch sheet: account name, what broke, whether they're under contract. And it sends absolutely nothing. At 6am your dispatcher opens one tab instead of four and starts calling.
Written out, that's the entire spec:
[READ] On-call calendar, voicemail transcripts, job history
[WRITE] Ranked callback list — dispatch sheet (draft row)
[SEND] NONE — requires dispatcher sign-off
Three lines. That's what a harness spec looks like, and that's the whole document someone needs to go build it.
The limit is the feature
The instinct is to read [SEND] NONE as the unfinished part — the compromise you make until you trust it.
It isn't. The limit is what makes the thing deployable at all. A system that reads everything and sends nothing can be turned on this month, by a business with real customers, without anyone losing sleep. A system that can text your customer list needs a lot more thought than "let's try it."
You aren't giving something up by drawing the line. You're buying the ability to start.
Start read-only
The lowest-risk first harness in any business: pick one task, give it read access to exactly one system, no write, no send. Let it produce its output somewhere you have to go look at it on purpose.
Then watch it for two weeks. Not to see whether it's impressive — to see whether it's right, on your actual data, on a bad week. That tells you more than any demo will.
If it holds up, you've earned the right to let it write. Drafts only. Then you watch that.
Where this stops being a Saturday project
Read-only is genuinely something you can set up yourself. Most modern AI tools connect to a calendar, a folder or a sheet with no code involved, and if that's where you are, go do it this weekend.
Write and send are different work. Connecting to a CRM or a dispatch system means authentication, error handling, and deciding what happens when the system is down mid-task — and someone owns it when it breaks at 2am. That's the honest line between a weekend project and something that belongs in your operations.
Still sorting out whether a task even wants to be a built system? Chat, delegate, or build is the test, and from answers to assets covers wanting a tool rather than paragraphs. Both sit under the four disciplines.
This week
Pick the one task you most want off your plate. Write its three lines:
[READ]
[WRITE]
[SEND]
Be specific — name the actual systems, not categories.
That list is the spec. It's the document you hand to whoever wires it up, and it's the fastest way to find out whether the task is a weekend or a project.
If your [SEND] line has something real on it and you want it built with the guardrails designed in rather than bolted on, that's what we do. Book a 30-minute call and bring your three lines.