An AI Implementation Framework in Five Stages

An AI implementation framework is the order of work you follow once the job already has a name. NIST calls the risk side of that work Govern, Map, Measure, and Manage.¹ We use five stages in plain words, because that is how the talk goes when we sit down with an owner. First we name the job. Then we score the data and write the process. After that we pick a checker, and last we keep the work or stop it.

Cited in this note

88%

Most small firms we meet are still in the Forming band while those stages are unfinished. A tool is open on someone’s laptop and the rest is a hope. A hope is not a stage, and if you skip one you get the stall we describe in the stall note. The score that should come before any build is in the readiness framework. This note covers only the order of work after that score exists.

We like a plain, slow order, but most owners like a demo. A demo feels like progress because someone in the room finally relaxes. An order feels like a delay because it is only a list. The list is the work, though, and a framework you cannot walk through on one page is just a brochure. A brochure will not run your business on Monday morning.

What the numbers say about the middle

McKinsey’s 2025 survey says that 88% of people report regular AI use in at least one part of their business.² That is the easy number, and a seller will quote it before the coffee arrives. The harder numbers sit right beside it. An AI agent is a tool that takes actions on its own. Some 23% of companies are rolling out an AI agent system somewhere, while 39% are still trying things out. Among firms that earn less than $100 million, 29% are rolling AI out widely. Among firms that earn more than $5 billion it is nearly half. No more than 10% are rolling out agents in any one part of the business.

If you run a shop, a clinic, or a trades desk, please read that last line slowly. Regular use only means a person opened a tool. Rolling it out in one part of the business means that part runs on a written process, with a person who can stop it. The gap between those two is where projects end up as a browser tab that nobody closes. We do not think a small firm is behind just because it has not rolled out an AI agent system. We think it is right on time if it can finish one job and say honestly who checks the work.

Cisco’s 2025 survey of 8,000 leaders found a group called pacesetters, near 13% of that sample. They were 4 times more likely to move a first test run into daily use.³ Those leaders work in companies with more than 500 employees. A small business can still copy the order the pacesetters followed, which was to write the process first and then put it to work. You cannot copy their staff or their in-house experts. A test run in that survey is also not the same as a receptionist trying a new inbox on a Tuesday. Take the order they followed and leave the rest behind.

Stop at the stage you cannot finish. If you cannot name a checker, you are not ready to keep the work. We would rather you pause on a page with a name on it. That beats pushing a half-built flow into a week when the office manager is already behind on the phones.

A worked example

Take a neighborhood dental clinic. We use this walk-through when an owner asks what the stages look like in a real week, and it is not a result we measured. The owner is the dentist, and she still does the exams herself. The office manager runs the schedule and answers the billing questions. She also covers the front desk when someone calls in sick. The front-desk lead answers the phone and sends the reminders. A hygienist knows which patients are in the middle of treatment and which ones only need a cleaning.

The job is narrow. Patients who are due for a cleaning get a short reminder that sounds like the clinic. The dentist does not want a computer booking anything clinical or guessing at a balance. She asks for the reminder message and a person who reads it before it goes out.

We write the job in words the front-desk lead can say out loud. The tool drafts the cleaning reminder from the appointment record and never mentions treatment or a price. A person sends it. If a second job is hiding inside that sentence, we start over. A seller will try to make the job wider in the proposal. This one sentence is how we politely say no.

The data gets scored before anyone connects an account. The appointment book lives in the practice software. A paper list in the manager’s drawer names people who asked not to be texted. Insurance notes sit on another screen, and the hygienist knows which patients are upset about a bill. Those sources do not agree with each other. A draft built only from the appointment book would write to someone on the paper list. That mistake shows up on the first Monday if anyone looks.

The process is a chain of hands. On Monday morning the front-desk lead pulls the list of patients who are due, and the draft uses the appointment record only. Anything that sounds like a treatment word goes to the hygienist. The office manager reads the rest and sends it. If she is out, the messages wait, because nobody sends one just because the draft sounded sure of itself.

The checker is the office manager, by name, with time held on her calendar. If she cannot do it, the hygienist is the named backup. A checker who is simply whoever happens to be free is not really a checker at all.

After a few real Mondays, they look at what the manager sent back. If the problems are about wording, they keep the job and tighten the line. If the problems are the wrong people, they stop sending and fix the paper list and the appointment book first. Stopping is a finished stage too. The demo took an afternoon, but the paper list and the rule about waiting took the whole month.

How we’d run it

We run the stages in order, and we do not open the next one while the current one is still a shrug.

First, we sit with the owner and with one person who does the job by hand. We ask to see last week’s work on the screen they really use. Then we write one sentence that names the job, the facts it uses, and the thing the tool must not touch. If they want several jobs, we pick the one that happens every week and hurts when it is late. The rest go on a closed list. Opening that list is how a simple reminder turns into the whole patient experience before anyone has sent a real one.

Second, we score the data in the room. We ask where the record lives, who types it, and what happens when it is wrong. We also ask which other list says something different. In the clinic that is the appointment book against the do-not-text drawer. On a trades desk it is the job ticket against the text thread where the customer changed the light fixture. In a shop it is the register against the note scribbled on the bag. If nobody can point at the one place the facts live, we stop. We will not agree to clean it up later, since later is how bad names get a green check.

Third, we write a short script for a loud morning. It says who pulls the list and what the draft may see. It also says who reads it and what to do with a strange one. Last, it says what happens when that person is out. We read it back to the person who will run it. We change it until they say they would follow it on a day when the phones are bad. A script they will not follow is only decoration.

Fourth, we name the checker and hold the time on the calendar before anything is switched on. We also write down where a rejected draft goes, because a side chat does not count. If the rejection stays hidden, the same mistake comes back. Then everyone treats it as brand new.

Fifth, we keep the work or stop it, using a rule we wrote before the first live send. Keep means the checker can clear the queue in the time we held and the rejections are about wording. Stop means we unplug the sending step and go back to the files. We do not add another job to soften the stop. We also do not let a seller rename our sentence as the first step of a big platform.

We stay with the owner through all of this, and not only with the person who bought the license. The owner is the one a friend will text when something goes wrong. So our part is to read the original sentence out loud again whenever people forget it.

Common mistakes

Starting with the tool is the mistake we see most. A cousin sends a link, and the first hour goes to logins while the job is still a fog. Connecting an account is easy. You cannot quietly undo messages that were sent to people who asked to be left alone.

The demo is the next trap. The records on the screen look clean because someone prepared them, and the do-not-text list is missing. So is the Monday when two people are out. If you sign while the sample is still on the screen, you have bought a performance instead of a tool.

A checker on a slide, and not on a calendar, will let the queue rot. A line like “the office will review it” is not a name. We have watched reminders go out because each person assumed another person had looked. Then the patient heard about a cleaning while already sitting in the chair. The apology lands on the front desk, who never agreed to catch it.

Making a quiet job wider is how a sensible project turns sour. The reminder feels calm, so someone points the same login at billing. A clumsy reminder is an awkward phone call. A wrong balance is a question of trust, and the two do not belong together just because they share a password.

Copying a big-company rollout flatters the room and wears out the desk. Cisco’s pacesetters were 4 times more likely to move a first test run into daily use. They also work in companies with more than 500 employees. They have people whose whole job is the process, while you have a manager who also covers lunch. Copy the order they followed, and do not copy a pile of test runs.

Feelings are a mistake of their own, because many owners refuse to stop. A stop feels like a public failure, so a weak flow stays alive. If you stop the sending and put the reason on the same page as the job, the framework worked. A tab that is always being tweaked means it did not.

What we’d do

We would make the owner pick one job, and we would walk away if she would not. That sounds stern if you wanted a partner who nods along. It is how we have seen this work finish. Use is common, and rolling AI out in one part of the business is rare, since no more than 10% roll out agents in any one part. The firms in that thin slice finished a process. They did not win a contest for the most logins.

We would not aim a clinic at the Cisco pacesetter share. Near 13% of a sample of 8,000 leaders, in companies with more than 500 employees, describes a very different building. The useful part is what they did, which was to write the process first. That let them move a test run into daily use 4 times as often. A clinic can copy that with a manager and a calendar hold. It cannot copy it by wearing out a part-time desk and calling the tiredness progress.

We would keep the old reminder running beside the draft until the manager trusts what she reads. Sellers dislike that, because it makes their product look optional. It still costs less than a week of apologies. When she trusts the draft, we turn the old way off in a sentence the staff can hear. That way nobody keeps doing both in secret.

If a tool is already on your card statement, please do not ask us to make it sound bigger than it is. Ask which stage you cannot finish. A fuzzy job goes back to the readiness framework. A charming demo followed by a bad Monday belongs with the stall note. A habit that is still forming can sit in the Forming band until the checker has a name. That is a perfectly respectable place to be.