Most conversations about AI in small business start with what the tools can do. Earn the Change starts somewhere else: with who decides.
That question gives three modes, and they work best as rungs on a ladder.
The three rungs
- Assist. A person asks, the AI drafts, the person decides. Assistants and copilots live here. The AI helps, but a human makes every call.
- Embedded. The AI handles one defined step inside a workflow. People handle the exceptions.
- Agentic. The AI carries several steps on its own, with people at checkpoints chosen in advance.
Each rung up hands the AI more responsibility. So each rung up has to be earned.
Exhibit 1
Highlighted boxes mark where a person makes the call.
Assist
Decides: the person
Person asks
a question, or for a draft
AI drafts or suggests
nothing goes out yet
Person decides
edits, sends, or throws it out
Needs: the normal path
a one-page SOP is enough
Embedded
Decides: the workflow
Work arrives
an invoice, an email, a form
AI runs one defined step
sorts, tags, or summarizes
Inside the rules?
yes flows on · no, a person decides
Needs: exceptions written
what doesn't fit, where it goes
Agentic
Decides: the AI, within limits
Goal, inside set limits
what it may do on its own
AI plans and acts
several steps, using tools
Person signs off
at checkpoints chosen first
Delivers, can be undone
start with reversible actions
Needs: decision rights
do, ask first, never touch
The more the AI does on its own, the more the process underneath has to be proven.
Source: Earn the Change, Troy Green
Climb one rung at a time
The same belief that runs through the whole book applies here: prove it at a lower level of autonomy before you let it act on its own.
Start at Assist. Run it as a small test, on real work, with your three numbers. Only when that holds do you consider moving one step up. Skipping straight to Agentic because the demo was impressive is one of the most common, and most expensive, mistakes.
And plenty of processes should stay at Assist forever. That's a perfectly good answer.
Exhibit 2
Prove each level before climbing. If the protect number slips, step down.
Agentic
Protect: Real consequences
AI acts, people at checkpoints
↑ proven: move up↓ slips? step down
Embedded
Protect: Its error rate
AI runs one defined step
↑ proven: move up↓ slips? step down
Assist
Protect: Time spent checking
Person decides, every time
Source: Earn the Change, Troy Green
You can only safely automate livestock
Embedded and agentic AI need a process that's standard, stable, and written down, including its exceptions. A one-page description of the normal path is enough for a person, and enough for Assist, because a person still decides. It isn't enough when the AI acts alone. For that, you also need to write down what doesn't fit, where it goes, and who decides.
Who decides, and who owns the mistakes
At every rung, someone has to own the AI's mistakes. Not the vendor. A named person on your team, whose role has changed to match. If nobody can say who that is, you're not ready to move up.
You also need a step-down rule: what result tells you to move the AI back down a rung, and who is allowed to switch it off.
Exhibit 3
Before AI
Who decides
The person
Who answers for it
The person
Blame matches authority
AI added, blame stayed
Who decides
The AI
Who answers for it
Still the person
Expect overrides
Blame moved too
Who decides
The AI, by set rules
Who answers for it
Owner of the rules
People let it work
Source: Earn the Change, Troy Green
Watch the hidden cost of checking
At Assist, the most important protect number is the time people spend checking and fixing what the AI produces. A tool that saves ten hours a week and quietly costs eight hours of checking hasn't saved ten hours.
There's a simple way to find out. Sit with one person while they finish one real task with the tool. Make a tally mark every time they correct or redo its work. If fixing takes longer than doing it by hand, stop there. The step with the most marks is usually your next obstacle.
Exhibit 4
An HVAC company's after-hours calls, 4 p.m. to 8 a.m., split by what's at stake.
After-hours call
4 p.m. to 8 a.m.
Seasonal tune-ups · Agentic
AI answers and books open slots; every booking can be undone
Repairs · Embedded
AI sorts urgent or routine; routine calls go to the morning queue
Emergencies · always a person
Urgent calls page the on-call technician
Source: Earn the Change, Troy Green
Ask the vendor the right question
A demo is built to impress, on clean data, in a controlled setting. So ask any vendor to run their tool on your own messy cases. Ask where your data goes, who can see it, and whether it's used to train anything else. If you can't get clear answers, treat that as a no.
Before you buy
Pick one place your business uses AI today, or is thinking about it. Write down which rung it's on, who decides, and who owns its mistakes. If any of those are blank, that's your first step. Not a new tool.

