We don’t start with AI.We start with the work.
Most AI projects fail before the technology gets a vote: wrong process, dirty data, nobody owning the output. The method exists to decide what gets automated, what gets killed, and what earns autonomy. Five stages, in order, every client, no exceptions.
- 1
Map
Lay the business out in three lanes and find where it hurts.
- 2
Triage
Four gates. Most processes don't survive them, and that's the point.
- 3
Score
Six weighted drivers decide the first project on numbers, not vibes.
- 4
Install
Agents get hired like people: a name, a job, a schedule, a boss.
- 5
Run
Autonomy is earned one dial-stop at a time, in the open.
Map
Every business fits on one grid. Three lanes across: Growth, Operations, Back Office. Three jobs an agent can do, down: Observe, Identify, Build. An agent is a trigger, a judgment call, and an output. It’s a workflow, not a person, and it lives in exactly one box.
| Growth | Operations | Back Office | |
|---|---|---|---|
| Observe Watches your data and turns the lights on. | Yesterday's leads by source, and what each one cost | Flag the job going sideways before the client calls | Cash, pipeline, and open tickets in your inbox by coffee |
| Identify Triages and preps so a human starts with context. | Score and research every inbound lead before the call | Sort the inbox and route each thread to its owner | Classify every card transaction with a written reason |
| Build Does the work: drafts with a human in the loop. | Draft the follow-up after every sales call | File crew photos to the right job folder | Draft the weekly finance summary for review |
Observe first. Watching is the cheapest, safest first agent. You earn your way down the rows.
The team picks the box. The workflow they hate most is your first agent. Built with them, they become believers.
One box at a time. “I need a marketing agent” is too big to start and too vague to finish.
Triage
Never automate a broken process. It just breaks faster. Every candidate passes four gates, in order, and we stop at the first yes.
Eliminate
Does anyone actually consume the output? Would anything break if it stopped tomorrow?
Kill the process. No build, and you just got hours back for free.
Simplify
Are there approval steps, duplicate entries, or handoffs that add no value?
Strip them first. A smaller process is a cheaper automation.
Standardize
Is it done differently by different people or locations?
Standardize before building. You cannot automate five versions of one process.
Automate
Only what survives the three gates above.
Now, and only now, the technology conversation starts.
Half the value of this stage is what never gets built. Killing a report nobody reads costs nothing and buys back hours. We would rather lose a build than sell you a faster version of a mistake.
Score
What survives triage gets scored on six weighted drivers, out of 55. The number picks the first project, not the loudest opinion in the room.
| DRIVER | WEIGHT | WHAT IT MEASURES |
|---|---|---|
| Value | ×3 | Hours saved times loaded rate, plus error cost, annualized |
| Data readiness | ×2 | Digital, accessible, decent quality |
| Feasibility | ×2 | APIs or exports exist, process is stable, exceptions manageable |
| Workflow fit | ×1 | Output lands where people already work |
| Sponsorship | ×2 | A named owner who wants it and can approve it |
| Risk, inverted | ×1 | 5 means failure is cheap and recoverable, 1 means a lawsuit |
The verdict bands
- 44 and up
- Strong first project
- 33 to 43
- Phase two
- Under 33
- Park it, fix the weak driver
And the tiebreak rule: the first project must be visible to the owner weekly, shippable in under 30 days, and cheap to unwind. Its job is to build belief, not to maximize ROI.
Install
Nobody adopts a browser tab. Every agent we install gets the same four things every hire gets, written down:
- A name. The team talks about it like a teammate, because it is one.
- A job description. One page: the task it owns, what it escalates, what it never touches.
- A schedule. When it runs and where its output lands.
- A boss. A named human on your team who owns the output.
Three numbers on every job description: tasks completed · intervention rate · pass rate on sampled runs
Your first two agents are live by day fourteen. Typically the Morning Briefing and the Meeting Notes Brain, because they observe, they’re safe, and the whole team feels them in week one.
The right tool, not the fashionable one
The intake answers route each process to its technology. AI is one row of four, and we tell you when it isn’t the answer.
| Same every time, starts from a system, under 10% exceptions | Deterministic workflow. No AI needed, and we'll say so. |
| Same every time, starts from email, PDF, or paper | AI reads, rules act. Extraction plus workflow. |
| Judgment or mix, language-heavy, a human can review the output | An agent with a human in the loop. |
| Deep judgment, high stakes, hard to verify | AI assists, a human decides. Or we don't touch it. |
Run
Autonomy is earned, not configured. Every agent starts at the left end of two dials and moves one stop at a time, only when its numbers say it earned the next one.
When it acts
How much it acts
Everything runs in the open. Every agent action is logged, every run is auditable, and every week you get a plain-English digest: what your agents did, what they flagged, where a human stepped in, and the one thing we’d change next.
And it stays run. The AI platforms change monthly. AI Care keeps every installed agent monitored, fixed, and current, at the flat price you saw before you enrolled.
We’ll run your business through the method, live.
That’s what the free AI Demo Hour is: one hour on site, your real workflows on the matrix, and a written summary of what we found and what we would tackle first.