The readiness questions we asked at the start of an AI enablement engagement were mostly the wrong ones. We asked about data, platforms, and strategy documents. Two years of running these programs taught us the variance lives somewhere far less impressive: whether one named person has calendar time, whether a manager intends to review work differently afterward, and whether anyone has decided what the reclaimed hours are for. Each is checkable in an afternoon. Here is which signals predicted a program that landed, which predicted nothing, and which gaps are worth a two-week delay.
What we got wrong in the first year
We inherited the industry’s readiness checklist without testing it. Every framework on the market rates the same eight dimensions — strategy, data, governance, technology, use cases, people, culture, measurement — and treats them as equally load-bearing. We asked those questions too.
The items are not wrong. They are unweighted, and almost none of them are about the thing being bought. A company can score well on data maturity and still have forty people who open ChatGPT once, get a mediocre paragraph back, and never return.
The shape of the failures made that obvious. Programs that changed nothing were rarely blocked by infrastructure. They were blocked by a manager who never asked for the new method, an owner whose calendar had no room for follow-through, or sessions run on example documents because nobody cleared real ones.
That matches one of the franker statements in the vendor literature. Microsoft’s adoption guidance puts it plainly: “The most common mistake is treating this change as a technology deployment. It’s a behavior change program. Licenses don’t become usage without enablement” (Microsoft Learn, updated September 2026). In the same document, organization and culture is the only capability driver rated a full maturity level above the rest: “Technology is the easy part. Adoption is the challenge.”
So we rebuilt our intake around four questions and dropped most of the others.
The four signals that predicted whether AI enablement landed
These four preconditions separated engagements that changed working habits from engagements that produced a well-reviewed session and nothing else. Each is observable before a contract is signed.
| Signal | What to check | Why it predicts |
|---|---|---|
| A named owner with protected time | Hours blocked for the eight weeks after the session | A title without time produces a launch and silence |
| A manager who will review work differently | What they will ask for in the next review cycle | Expectation from above makes the new method the default, not an option |
| Real documents cleared in advance | Three to five approved work artifacts per team | A session on invented examples teaches a tool; real work teaches a job |
| A decision about the reclaimed time | A written answer to “spent on what instead” | Without it, saved minutes dissipate and nothing is left to point at |
The third surprised us most. Document clearance sounds administrative. In practice it became our most reliable advance indicator, because getting it requires a manager to chase a sign-off, a compliance view of what can go into a model, and participants who know which parts of their job consume the hours. A company that cannot produce five cleared documents in two weeks is telling you something about all three.
The fourth is the one buyers resist and the evidence supports. Microsoft’s guidance is unusually blunt for vendor documentation: “Time saved is a leading signal, not the goal. A trial can show real-time savings on written tasks while overall productivity stays flat, because saved minutes leak away instead of moving to higher-value work.” With no named destination for the hours, the success condition is a feeling.
Collecting all four late is how a kickoff slips a month, which is why order matters as much as presence — see our rules for sequencing a program.
The readiness questions we stopped asking
Dropping questions was harder than adding them, because each one sounds responsible. In two years, none of these four changed our recommendation.
Data maturity. It matters enormously for building systems and barely at all for teaching people to work with a general-purpose model on their own documents. Teams with messy data learn fluency at the same rate as teams with clean warehouses. We still ask when the conversation turns to custom automation, a different purchase at a different price.
Whether an AI strategy document exists. A strategy deck correlates with executive attention and with nothing else. Several of the best programs began with no document and one frustrated COO.
Self-reported comfort scores. Asking people to rate their AI confidence from one to five produces a number that moves when the office mood moves. It does not predict whether they can get a usable draft out of a model on a Tuesday. Collect a work sample instead. The instruments built on those scores have their own problems, covered in our guide to reading a readiness assessment.
A count of volunteer champions. Volunteers were already going to use the tools. The useful champion is whoever the team already asks for help, hand raised or not.
Those dimensions matter to a business. They are poor forecasters of whether training changes behavior.
Why the room they return to matters more than the session
The training-transfer research explains the pattern better than any vendor framework. A meta-analysis of 89 empirical studies by Blume, Ford, Baldwin and Huang found that transfer of training is predicted by trainee motivation and a supportive work environment alongside individual ability (Journal of Management, 2010). The finding that reframed our scoping is more specific: those predictors were stronger for open skills, the judgment-based kind, than for closed skills, the procedural kind — with computer software given as the example of a closed skill.
Working with a language model looks like a closed skill and behaves like an open one. There is no correct sequence of clicks. What a person learns is when to trust an output, how much context to supply, and which tasks are worth delegating. That is judgment, and judgment is where the environment a trainee returns to carries the most weight.
The field data agrees. Slack’s Fall 2024 Workforce Index, covering more than 17,000 desk workers, found that workers whose employer had provided guidance showed a 13-percentage-point rise in AI adoption since January, against 2 points for workers with none (Slack, November 2024). The same survey found 61% of workers had spent under five hours in total learning to use AI, and 30% reported no training at all.
Guidance from an employer is not a curriculum. It is permission plus expectation, delivered by someone whose opinion affects a performance review — the cheapest intervention available and the one most programs skip. It is also why we treat habit formation as the deliverable rather than instruction, argued at length in the case for building habits rather than buying tools.
The statistic everyone quotes, and what it actually says
If you have sat through a vendor pitch this year, you have heard that 95% of AI pilots fail. It appears in nearly every readiness deck in the category, usually unattributed.
The underlying document is MIT’s NANDA report, The GenAI Divide: State of AI in Business 2025, published July 2025 as preliminary findings from more than 300 publicly disclosed AI initiatives, 52 structured interviews, and 153 survey responses collected at four industry conferences. The report’s own claim is that roughly 5% of integrated pilots are extracting millions in value. The 95% is press framing of that sentence, and the two are not the same statement. A preliminary study with 153 responses cannot establish an industry failure rate. What it can do, and what its authors argue, is locate the failure in integration with real work rather than in model quality.
The genuinely sobering figure is quieter. McKinsey’s November 2025 State of AI survey found 88% of respondents reporting regular AI use in at least one business function, while only about a third had begun scaling across the organization. Widespread access, concentrated capability.
One vendor puts an order of magnitude on that gap: Correlation One reports that “enterprises typically hold 2–3x more AI licenses than trained, active users,” sourced to 2026 client work across more than ten industries (Correlation One, July 2026). No sample size or instrument is published, so treat it as a practitioner observation, not a measured statistic. It matches what we see, and the price of that gap is the subject of what an unsequenced rollout actually costs.
What we now baseline before day one
A program with no before-measurement cannot be defended at renewal, and retrofitted baselines never survive a finance review. Four things, collected before the first session:
- Three to five named recurring tasks per team, with a timed before. Not “writing takes too long.” The specific weekly task, who does it, and how many minutes it took last time.
- The active-seat percentage from your license console. Export it rather than estimating. Leaders are reliably wrong in one direction: McKinsey’s Superagency research found C-suite respondents estimated 4% of employees used generative AI for at least 30% of their daily work, while employees themselves reported 13% (McKinsey, January 2025, CxO survey n=118, employee survey n=3,002).
- One work sample per participant. The actual output they produce, saved before training. It is the only before-and-after a skeptical reader finds persuasive.
- The written answer on reclaimed time. One sentence per team, signed off by whoever owns that team’s targets.
None of this needs a platform or a consultant. It needs four hours of somebody’s attention, and it is the difference between a program you can renew and one you re-justify from scratch.
Gaps you can close in two weeks, and gaps that should postpone the work
Readiness is not binary. Most gaps are logistics. A few are structural, and running a program into one of those wastes a budget.
| Gap | Verdict |
|---|---|
| No acceptable-use note for staff | Fixable. A one-page statement of what can and cannot go into a model is a two-day job |
| No cleared documents | Fixable, and the fixing is itself valuable |
| No baseline | Fixable in four hours |
| Owner has the title but no calendar time | Fixable, but only by whoever controls their priorities. Settle it in writing first |
| Managers not briefed | Fixable. Half a day with the people who approve the work, before the sessions |
| Licenses not provisioned, or blocked by IT | Postpone. Do not teach a tool people cannot open on Monday |
| Leadership will not use the tools | Postpone, or shrink to one willing team |
| A reorganization landing inside the quarter | Postpone. Nobody learns a new method while waiting to hear whether they still have a job |
| No budget for anything after the session | Postpone or rescope. Microsoft’s guidance calls this “one-and-done training,” where “skills and confidence fade” |
One distinction deserves emphasis. A company whose leadership declines to use the tools can still run a good program inside one function with a committed VP, and that is often the right first move. A company mid-reorganization cannot, at any price.
What changed in how we scope AI enablement
Four concrete changes, each traceable to something that went wrong.
Document clearance moved into the proposal. Once a kickoff task, it is now a precondition with a date, because a session on real work and a session on sample work are different products.
A manager session comes before the cohort sessions. Half a day with the people who review the output, covering what to ask for and what good looks like. We used to offer it afterward as an add-on. Almost nobody bought it, and the programs without it faded fastest.
One curriculum per cohort became several. Moderna reported the same thing publicly after its own upskilling program: a single AI fundamentals course was too basic for some employees and too advanced for others, so the company split it into six tracks by level (Fortune, March 2024). Mixed-level rooms punish both ends.
Follow-through is quoted as part of the program, not as optional. Clinics inside real work, spaced over the weeks after the sessions. Market pricing for hands-on workshop programs runs roughly $2,000 to $15,000 depending on scope and cohort count, and follow-through is the line item a buyer trims first. It is also where the behavior change comes from, so we no longer present it as severable.
For how the first-year arc fits together, the playbook that runs from license rollout to measurable capability sets out the sequence, and the broader argument for why access is not capability is made in the case against treating licenses as progress.
Frequently asked questions
How do I know if my team is ready for AI enablement?
Check four things: a named owner with hours blocked for the two months after training, one manager who can say what they will ask for differently, three to five real documents cleared for use, and a written answer to what the saved time is for. All four present means ready. Two or more missing means two weeks of work before you spend the budget.
What is the most common reason a program fails?
No expectation from above. People return from a session to a manager who accepts work produced the old way, so the old way continues. Transfer research identifies a supportive work environment as a predictor of whether learning gets used, and that support shows up in what a manager asks for.
Do we need an AI strategy before we train anyone?
No. A strategy document helps executives agree with each other and does not affect whether a team learns to work with a model. You need a named owner, cleared work, and one function willing to change its method. Strategy becomes necessary when you start building custom systems.
Is it true that 95% of AI pilots fail?
Not as usually stated. The figure is press framing of MIT’s NANDA report from July 2025, whose own claim was that about 5% of integrated pilots were extracting millions in value, from a preliminary study of 300-plus public initiatives, 52 interviews, and 153 survey responses. The useful part is the diagnosis: failures come from poor integration with real work, not weak models.
How long does it take to get ready?
Two weeks for the logistics — document clearance, an acceptable-use note, a baseline, a manager briefing. Structural gaps take longer, because they involve someone’s priorities rather than their calendar.
What should we measure before the first session?
Three to five named recurring tasks with a timed before, the active-seat percentage straight from the license console, one saved work sample per participant, and the written decision on reclaimed time.
Who should own this at a company of 200 people?
Whoever controls the calendars of the people being trained, which usually means a functional VP or the COO rather than IT. The owner needs no AI expertise. They need authority to put a clinic on a calendar in week six and have people attend.
What does a program like this cost?
Hands-on workshop programs sit in a market range of roughly $2,000 to $15,000 depending on cohort count and scope. Custom automation work that follows is a separate purchase, typically $25,000 to $150,000. Treat any quote with no follow-through as cheaper than it looks.
What if leadership will not use the tools themselves?
Shrink the scope. Run the program inside one function with a committed leader and use the result as the argument. Microsoft’s adoption guidance calls role-modeling the strongest adoption lever, so going company-wide without it spends a company-wide budget on a function-sized outcome.
Where to start
Readiness here is not a maturity score. It is four observable conditions and the honesty to postpone when a structural one is missing.
Run the check this week. Name the owner and block their time. Ask one manager what they will require differently next review cycle. Request five cleared documents. Write one sentence per team about where the hours go.
If two of the four are missing, you have a two-week project before you have a training project. If you want a second opinion on where your team stands, book a free AI-readiness call and we will walk the four signals with you.
Arthur Wandzel
