An AI Explained My Own Methodology to Me, and Made Up the Numbers
By Ann Marvin, President and CEO, IMA Worldwide
An AI wrote an article about my methodology last week. It was confident, well organized, and scored a 92 out of 100 on the content grader. It also invented two metrics that do not exist.
I want to be precise about this, because it is funnier and more useful than a general complaint about AI slop.
The article described something called the EMR Ratio, which it defined as the ratio of installation effort to implementation success, and used as a risk measure roughly nine times. EMR is real. It stands for Express, Model, Reinforce, and it describes what a leader says, what a leader does, and what a leader rewards. The ratio attached to it is 1x, 2x, 3x: expressing counts once, modeling counts twice, reinforcing counts three times. It is not a ratio of installation to implementation. It is not a risk measure. You can read what EMR actually is in about four minutes.
The article also referred, repeatedly and with great authority, to the 40 to 60 percent Adoption Decay curve. There is no such curve. There is no such term. I have been in this field a long time and I would like to meet whoever plotted it. I might start to incorporate it since it is a cool name.
The tell is which numbers get repeated
Here is the part worth sitting with. The two invented metrics were the ones the article leaned on hardest. Every section came back to them. They were the spine.
That is backwards from how a person writes. When I write about this work I lean on the things I have watched move, and I am cagey about the rest. The numbers I repeat are the ones I have defended in a room. An AI has no such asymmetry. It has no memory of being wrong in front of a client, so a metric it invented three paragraphs ago is exactly as load bearing as one with forty years behind it. Confidence is flat across the whole surface.
Which is the thing itself, is it not. Not a technology failure. A judgment failure, outsourced.
Why it reached for a framework it had to invent
I keep coming back to the question of why the AI did not simply use what was there. It had the real material. It chose to build new scaffolding anyway.
I think it is because it was asked to write about AI adoption vs traditional change management, and that framing carries a hidden instruction: AI is new, therefore it needs new apparatus. So the model went looking for AI-native change metrics, did not find any, and helpfully manufactured some. It gave me an Adoption Decay curve because the question implied there ought to be one.
Humans are doing the exact same thing right now, at scale, in real companies, with real budgets. Every week there is a new AI adoption framework with a new acronym and a new maturity model. Almost all of it is a rediscovery of things that were settled before any of us had a chatbot. The pitch decks are new. The problem is not.
What the old models actually get right
I am not here to defend the old models uncritically. Let me be fair to them and then say where they stop.
Kurt Lewin gave us unfreeze, change, refreeze, and the genuinely important insight that you have to loosen the current state before anything new will hold. John Kotter, in Leading Change in 1996, was right that nothing moves without urgency and a coalition with actual weight behind it. Though everything cannot be on fire. Jeff Hiatt's ADKAR at Prosci put the individual at the center, which was the correct place to put them. I use all three vocabularies without embarrassment.
Where they stop is the same place in each one. They tell you what has to be true. They do not tell you whether it is true at your company, on this initiative, before you have spent the money. And they tend to leave the building at go live, which is roughly when the interesting part starts.
That is the actual difference, and it is not a philosophical one. In the methodology I steward there are ten practice areas, each paired with one of ten scored diagnostics, running on a cycle of plan, implement, monitor. You get a number before the check clears and you stay through the window after launch. That is it. That is the whole differentiator, and it is unglamorous enough that no AI would invent it.
The part that cannot be generated
You knew where this was going, because I always end up here.
Everything I have written this year has run into the same wall. AI made me faster at making things and did nothing for the part where people change what they do. Installation is not implementation, and AI has paved the road right up to the edge of that gap without narrowing it by an inch.
Look at the five conditions people need before they change and ask which of them a model can produce for you. Information, maybe. It can tell people what is expected. Willingness, ability, confidence, control: no. Not one. You cannot generate someone's belief that this will work, and you certainly cannot generate their sense of having had a say, because the having of a say is the thing itself. I will start to argue ability and confidence soon since I am learning a lot from AI.
Same with reinforcement. Reinforcement carries about three times the weight of communication, and AI is spectacular at communication and useless at reinforcement, because reinforcement is mostly a leader deciding what gets rewarded. There are six things a leader cannot hand to anyone else, and no assistant has ever done any of them.
So when the article told me AI adoption needs a new methodology, it had it exactly inverted. AI adoption is the case where the old methodology matters more, because AI has stripped away every other place the work could have been hiding. Sustainable AI adoption does not come from a new framework. It comes from the same disciplined attention to readiness, reinforcement, and honest measurement that has always separated adoption from mere installation.
What I would ask before your next AI rollout
Not a framework. Three questions, and you already know your answers.
What, specifically, gets rewarded differently now that this tool exists? If nothing does, the tool is running against every incentive already in the building and the incentives are undefeated.
Who was asked, and who was told? Those are two different groups and they behave differently for years.
What are you measuring after the launch communications stop? Usage during launch week measures your launch. What survives the silence measures your adoption.
None of that requires an Adoption Decay curve. It requires someone willing to look at the honest number, which has always been the hard part, and is now very nearly the only part.
I did keep the article, by the way. It is a genuinely useful artifact: a perfect specimen of the thing it was describing. It installed beautifully. It implemented nothing.