Ask an owner whether the business uses AI and the answer is almost always yes. Someone drafts emails with a chatbot, the help desk suggests replies, maybe an agent prepares the invoices. The yes tells you very little. It does not say what the AI is trusted to do, who checks it, or what happens when it gets something wrong. Carolyn Healey, a fractional CMO who works on AI enablement, published a seven-stage maturity ladder on LinkedIn in June 2026 that asks a better question. In her words: "how much autonomy have we actually earned the right to delegate?"[1]
What AI maturity measures.
AI maturity, in Healey's framing, is how much work you have safely handed to AI and what keeps that work accountable, not how much AI you have deployed. Her seven stages, from the bottom: Manual, Assisted, Embedded, Delegated, Orchestrated, Governed autonomy, Adaptive. Through stage 3 a person decides and does every step; from stage 4 an agent does the work and a person approves it, then supervises the system, then sets the limits it runs inside. Two of her rules do most of the work for a small business. Your real stage is set by your weakest function, not your strongest (in our reading, because the weakest one caps every workflow that reads from it). And you move up one stage at a time: an agent gets more autonomy only after the controls for it (logs, limits, an approval step) are in place. The first question for any function is not which tool to buy. It is whether the work is written down anywhere.
Seven stages, measured by what you hand over.
Most maturity models, Healey argues, stop at "AI deployed at scale," which made sense in 2023 and no longer does. Agents added two questions: how much you can safely delegate, and what you have built to govern it. She calls that second part the harness: the tools, memory, guardrails, logs, evals, circuit breakers and audit trails around an agent. Her line for it: "The model is the brain; the harness is the nervous system, and the harness is where deployments hold or fail."[1]
Read the ladder from the bottom. The first three stages differ in where the AI sits, not in what it is trusted with: in all three, a person decides and does every step. Stage 4 is the first real handover. An agent runs a scoped, multi-step task from start to finish, and a person approves the result before anything leaves the building (Healey's "human-in-the-loop"). That is also where, by Healey's account, the first real harness appears. Stages 5 and 6 move the person from approving each result to supervising the system, then to setting the limits the system runs inside. Stage 7 is a business whose work has been redesigned around agents.
A lens, not a measurement.
Healey's sharpest claim is about honesty: "most enterprises report Stage 5+ and operate at Stage 2–3."[1] Her post gives no source for that sentence, so read it as one practitioner's observation. The pattern is easy to picture. A company with one impressive agent in one department describes itself by that agent. Meanwhile the quotes are priced from memory and nobody can say which AI tools the staff have signed up for.
So use the ladder to ask questions, not to score yourself against other companies. The stages are her drawing, the boundaries between them are soft, and no one has tested whether a business at stage 4 outperforms one at stage 3. What the ladder does well is make two things hard to dodge: what the AI is actually allowed to do, and what would catch it if it went wrong.
Is the work written down?
This question is ours, and we would ask it before any of Healey's. You cannot hand a process to an agent if you cannot describe it. Stage 4 means an agent runs a scoped task from start to finish. Scoped means someone can write down the inputs, the steps, the rules, what finished looks like and who signs off. If the pricing rules for a quote live in the owner's head, no tool will get quoting to stage 4. The bottleneck is the missing description.
That makes the first step up for most small businesses unglamorous. Write a one-page procedure for one function. Name the tool people are allowed to use, what may and may not go into it, and who owns the decision. In Healey's terms, that is most of the move from Manual to Assisted: a shared tool, a policy and an owner. It costs an afternoon, and it is the work every later stage is built on.
Your weakest function sets the floor.
Healey's first rule for moving up: "Name the floor honestly. Your weakest function sets it, not your strongest."[1] Her post does not spell out why. Our explanation is that work crosses functions. An agent that is excellent at one step still reads its inputs from the steps before it, and if those steps run from memory, the agent's output is only as good as the memory.
Picture a commercial cleaning contractor in the Quebec City region, about 25 people. Here is how its functions score.
Ask the owner what stage they are at and the answer will be the invoicing agent: stage 4. It is real. It drafts the invoices, a person reviews the batch, and the logs show every run. But the agent reads its prices from work orders, and the work orders take their prices from quotes that the owner prices from memory and one spreadsheet nobody else opens. When a quote is wrong, the agent invoices the wrong amount faithfully, and the reviewer checking the batch has no rule to check it against. The company's real stage for anything that touches pricing is 1.
That is what "the floor" means in practice. Hiring sits at stage 1 too, but nothing downstream depends on it, so it is a policy gap rather than a bottleneck. Quoting is the floor that matters, because the best agent in the company sits on top of it.
Score by evidence, not intent.
Self-assessment drifts upward when the question is "what are we doing with AI?" It holds when the question is "what can we show?" Healey's stage descriptions translate into evidence you can put on a table. The tests below are ours, written from her descriptions. If you cannot produce the evidence for a stage, you are on the stage below it, whatever the tool is capable of.
| stage | the evidence you can show |
|---|---|
| 1 Manual | Nothing shared. People use AI tools on their own accounts, and nobody knows which ones. |
| 2 Assisted | One shared tool, a written use policy (what may go into it, what may not), and a named owner. |
| 3 Embedded | AI built into the software the work already runs in: the help desk, the accounting system, the CRM. A person still makes and carries out every decision. |
| 4 Delegated | A written scope for one task, the list of tools the agent may use, a log of every run, and a record of who approved each result before it went out. |
| 5 Orchestrated | Several agents on one workflow, a set of test cases someone owns and reruns after every change, and monitoring someone actually reads. |
| 6 Governed autonomy | Limits written as code, not as a slide: an action or spending cap the agent cannot exceed, an escalation rule, a stop switch someone has tested, and an audit trail. |
| 7 Adaptive | A workflow redesigned around the agent rather than the old process with an agent added, a feedback loop that changes it, and one leader answerable for the value. |
The tests are deliberately boring: a log, a written scope, a named owner, a stop switch somebody has pressed once to see that it works. None of them requires a better model. Past stage 3, what separates one stage from the next is the harness around the AI.
Move up one stage, not five.
Healey's prescription has four parts.[1] Here they are with what each means at the size of a small business.
- Name the floor honestly. Score every function with the evidence in figure 3 and write down the lowest. Then check which workflows read from it.
- Pick one core workflow, not ten experiments. One workflow that matters, carried up one stage and held there, teaches more than a dozen trials nobody maintains. If it reads from the floor, raise the floor first.
- Build the harness before the autonomy. Her line: "Guardrails earn you the right to delegate." The logs, the tool list and the approval step come before the agent is allowed to act. Our build order for production agents covers that sequence for one agent, and our note on what an agent may do without asking covers how to grant authority one action at a time.
- Put ownership on value, not pilots. Healey puts that ownership at the C-level. In a business of 25, that is the owner or one manager, answerable for what the workflow saves or earns, not for how many pilots are running.
One stage at a time is also a cost argument. The jump from stage 3 to stage 4 needs a written scope, a log and an approval step. The jump from 3 to 6 needs all of that plus tested stop switches and limits written as code, built at once, on a process that may never have been written down. The first is a contained project on one workflow. The second is how a business ends up with an agent nobody can explain and a bill nobody expected.
Where we put our own pipeline.
We apply the ladder to ourselves. This note was drafted by AI research agents. It then went through separate AI review agents (French parity, French register, content quality, a pass for the habits that make text read as machine-written) and a build check that confirms every diagram is present in both languages. Then it waited for a person to read it and approve the merge. Nothing on this site publishes without that approval.
On Healey's ladder that is stage 4 with some of stage 5: several specialised agents coordinate on one workflow, and a person approves the result before it ships. We have not moved it to stage 6, and we do not plan to. A published page gets read, cached and quoted before anyone can take it down, and a person reading it first costs little. The reviewers are also the same kind of model as the writer, which, as our note on AI reviewers shows, makes their agreement weaker evidence than it looks. That is a harness decision, and it is the kind the ladder is meant to make visible.
What to do this week.
- List your functions, not your tools.
Quoting, scheduling, invoicing, customer email, hiring, purchasing: whatever your business actually does. One line each.
target: every function on one page - Ask whether each one is written down.
If the rules for a function live in one person's head, it sits at stage 1 or 2 no matter what software it uses. Writing it down is the first step up.
- Score each function by evidence.
Use figure 3. A function gets the highest stage whose evidence you can show today, not the stage its tool could reach.
- Mark the floor and trace what reads from it.
The lowest score is your floor. Check which of your better workflows take their inputs from it. Those workflows are capped at the floor, however good they look.
target: the floor named, with the workflows that depend on it - Pick one workflow and move it one stage.
Build the harness for the next stage first: the written scope, the log, the approval step. Then hand over the work.
- Name one owner for the value.
One person answers for what the workflow saves or earns. Count that, not the number of pilots.
What this does not claim.
A practitioner's framing, not a study. The seven stages are one consultant's drawing, published as a LinkedIn post. They are not validated, the boundaries between stages are a judgment call, and other frameworks cut the same ground differently. The format is older than AI: the five-level Capability Maturity Model, built at Carnegie Mellon's Software Engineering Institute to rate software development processes, is the classic example. We use them for the questions they make you ask.
Her figures are not used. Healey attaches survey percentages to several stages, attributed to McKinsey, Gartner, Deloitte and BCG. We have not read the underlying reports, so none of those figures appear in this note. Her claim that most enterprises report stage 5 or higher while operating at 2 to 3 comes with no source, and we repeat it only as hers.
Higher is not always better. The ladder describes what you have earned the right to delegate, not what you should. Plenty of small-business functions belong at stage 3 for good: AI suggests, a person decides. A function earns the next stage only when the work is written down, the harness exists, and the handover is worth what it costs.
Illustrative example. The cleaning contractor, its five functions and their scores are ours, built to show how a floor caps the workflows above it. They are not a measurement and not a client record. The description of our own publishing pipeline is the process this note went through.
If you would like a second pair of eyes on where your functions actually sit, send us your list through the contact form. Scoring a business by evidence, choosing the one workflow worth moving first, and deciding what to leave alone is what our automation audit is for.