When an AI enablement program that worked stops working, the cause is usually drift rather than failure. The team moved past the syllabus, the tools changed under the material, the facilitator who built the room got reassigned, or the sponsor stopped showing up. Three of those four are the provider’s to fix. The fourth follows you to whoever you hire next, which is why switching AI training vendors resolves this less often than buyers expect.
This is not the same problem as a bad buy. A bad buy is obvious in week one: generic examples, a facilitator reading someone else’s deck, nothing in the room that resembles your work. The harder case is the engagement that landed well, produced real wins in the first quarter, and has quietly flattened by the third. Attendance holds. Feedback stays polite. Nobody can name a task that changed since spring.
Four ways a working program stops working
Level drift. The first sessions worked because the easy ground is enormous. Teach a room of 30 people to stop pasting a one-line prompt and start giving a model the document, the audience, and the house style, and you get visible wins in two weeks. That ground gets used up. By month six the room has split: a third run multi-step work with real judgment, a third are solid on the basics, a third never started. A syllabus written for a cold room talks past all three. The provider either re-levels or repeats itself.
Content half-life. The assistants move on a quarterly cadence. A session built around what ChatGPT, Claude, Gemini, and Microsoft Copilot could do nine months ago teaches workarounds for limits that no longer exist, and misses the features your team is already fumbling with alone. This is the most common reason a competent provider goes stale without anyone behaving badly: maintenance was never priced into the engagement, and nobody noticed until the material aged.
Facilitator substitution. The person in the room is the product. When a provider grows, the practitioner who ran your first cohort moves to selling and designing, and a newer facilitator delivers. The deck is identical. The room is not, because half of what makes a session land is somebody who can answer an unscripted question about your workflow without reaching for a slide.
Sponsor decay. The COO opened session one and the room was full. By session four she is on another fire, the managers took the hint, and the follow-through nobody formally owned stopped happening. The sessions are as good as ever. Nothing survives them.
| What you observe | Likeliest pattern | Whose problem it is |
|---|---|---|
| Strong attendance, flat feedback, no new use cases | Level drift | Provider’s, if they won’t re-level |
| Material covers limits your team already works around | Content half-life | Provider’s |
| Sessions feel thinner, questions get deflected to follow-up email | Facilitator substitution | Provider’s |
| Good sessions, no change by week three, managers absent | Sponsor decay | Yours |
| Enthusiasm in the room, reversion to old habits in the work | Sponsor decay or unchanged process | Yours |
The split matters because all four look alike from a buyer’s seat, and a change of supplier only resolves three of them. Replacing a provider to fix sponsor decay buys the same outcome twice and teaches the business that AI enablement does not work.
Your dashboard cannot see decay
License telemetry is the worst instrument for this question and the one most teams have. Seat counts, weekly active users, and message volume all rise while capability sits still, because people learn to use an assistant for the same three low-stakes tasks all year.
The survey data says the gap is near-universal. Docebo’s 2026 study The AI Readiness Gap, run across 2,000 respondents at enterprises of 1,000 or more employees in six countries, found 85% of employees say their training does not help them use AI well in their actual role, 57% say it is not relevant to their role, and 66% do not feel supported by their organization. On the buying side of the same survey, only 24% of learning leaders say learning is fully aligned with business strategy and 43% feel confident connecting learning to business results.
BCG’s AI at Work 2026 survey of 11,749 people across 14 countries puts a harder number on the same wall: 88% expect to need major upskilling within five years, while 36% feel properly trained. Both figures are unchanged from 2025.
None of that shows up in a usage chart. What does show up, if you look for it, is narrowness: the same handful of prompts, the same two tools, nothing that touches the work the business gets paid for. The premise that buying licenses is not the same as building capability runs through our AI enablement manifesto, and the corollary is that license data cannot tell you whether a program is still earning its fee.
The attribution test
Three questions, answerable in an afternoon, separate a provider problem from a deployment problem.
Has the material changed since the first cohort? Ask for the current session plan and compare it to the one you bought. If the exercises, the example documents, and the tool coverage are identical nine months on, the provider is selling a fixed curriculum and calling it a program. That is the one a replacement genuinely fixes.
Who is in the room, and are they the person you bought? Name the facilitator on your first session and the facilitator on your last. If the second name is newer and the sessions feel thinner, you are buying a different product at the same price.
Can you name the manager who asked someone to use this in the following week’s work? If nobody comes to mind, stop the vendor conversation. BCG’s data found only 33% of frontline employees say leadership communicates clearly about AI, and only 28% see strong alignment between what leaders say and what the organization does. Docebo found 91% of learning leaders say their organizations have not fully redefined workflows around AI. A session cannot redefine a workflow it is not allowed to touch.
Two clean yeses on the first two and a blank on the third means the provider is doing the job and the business is not. That is the most common finding and the least welcome, because it cannot be solved with a purchase order.
The call to make before the renewal call
Most buyers go from quiet dissatisfaction straight to a shortlist, skipping the one conversation that is free. Put the diagnosis to the incumbent directly, with evidence, and ask them to re-level.
A provider who can do this answers with specifics inside a week: a split cohort structure for the three tiers your room has become, exercises built on documents you supply rather than their samples, a named facilitator, and a written refresh cadence for the material. A provider who cannot answers with a reskinned deck, an offer of more seats, or a certificate track. That answer is your decision, made for you, at no cost.
Three things to ask for in writing, because they reveal the business model:
- Who maintains the curriculum, and on what cadence. Quarterly is the honest answer for LLM fluency work. “Continuously” without a mechanism means never.
- Which named person delivers, and what happens if they are unavailable. A substitution clause with your approval attached is reasonable; silence is a tell.
- What the next tier looks like for the people who already passed the first one. If there is no answer, you have bought an introduction, not a program.
Those are the same questions worth asking a replacement. Our vendor evaluation framework scores them properly, and if you take the decision to a formal process, the guide to what good RFP answers look like covers the evasions to watch for in the responses.
Four moves, and when each is right
| Move | Right when | Cost | Risk |
|---|---|---|---|
| Re-level with the incumbent | Material is stale but the practitioner knows your business | A scoping call, usually a modest rise per session | Buys time without fixing a fixed-curriculum provider |
| Narrow the scope | The program serves everyone and reaches nobody | Lower spend, fewer people, deeper work | Reads as retreat to anyone tracking headcount reached |
| Replace the provider | Curriculum has not moved, or the person you bought is gone | Re-baseline, a new brief, four to eight weeks | Repeating the mistake with a better-pitched version |
| Build it internally | You have a credible internal practitioner and a standing need | A salary line and year-two maintenance | Year two, when the owner’s day job reclaims them |
Narrowing is the underused move. A program that covers 200 people lightly and changes nothing is worse value than one that takes 25 people in a single function and rebuilds how they produce one deliverable. The spend goes down and the evidence goes up.
The internal option deserves arithmetic rather than reflex. Our buy-versus-build playbook splits an enablement program into its four assets and prices them separately, and the case for an internal AI academy sets out the four conditions that disqualify a build before you cost one.
If the decision is replacement and you cannot wait for the renewal date, the mechanics of replacing a provider mid-program cover the cancellation ladder, the notice math, and what to extract before you give notice.
What switching AI training vendors actually costs
The invoice is the smallest part. Three costs do not appear in any quote.
Re-baselining. A new provider inherits no record of what your team can already do. Either you supply that record or the first session is an introduction your senior users sit through twice. The wasted hour is minor; the credibility cost of making your best adopters repeat a beginner session is not.
Relationship restart. The holdouts in any room are won by a specific person over several sessions. A new facilitator starts that work from zero, and the same four skeptics tend to decide whether the rest of the floor engages.
Comparison opacity. You cannot build a switching comparison from public prices. Coursera, at the time of writing, publishes $399 per license per year for its Teams tier at 2 to 499 licenses and sends anything larger to sales. Every instructor-led provider in the category quotes privately, including the ones that publish ranked buyer’s guides naming themselves first. Market ranges are the only honest reference point: a scoped LLM fluency engagement generally lands between $2,000 and $15,000, and the custom automation work that often follows runs $25,000 to $150,000.
Which also means the proposals you receive are shaped by who sent them. The map of the trainer market separates practitioners from platforms and resellers, and is worth reading before you take a quote as a market rate.
Take the baseline before you switch
The step almost nobody takes, and the one that decides whether this happens again: capture what your team can do before the new provider arrives.
It does not need to be elaborate. Pick four tasks the business cares about, have 10 people do each one with whatever AI help they choose, and keep the outputs with timings. Repeat it 60 days after the new program starts and you have an answer no completion report can give you. Skip it and you will be in this same conversation next year with no way to tell whether the second provider was better than the first.
One note on timing: start the exit work well before the end date. Reporting on AI vendor exits, including InformationWeek’s June 2026 guidance, puts the sensible lead time at 90 days rather than 30, and recommends running the replacement in parallel for one cycle rather than cutting over cold. Services engagements reward the same patience for a simpler reason: a provider with unperformed sessions on the books is far more cooperative about handing over materials than one whose final invoice has cleared.
Frequently asked questions
How do I know whether the vendor is the problem or our rollout is?
Three checks. Compare the current session plan against the one you bought, name the facilitator on your first and most recent session, and name the manager who asked someone to apply the training in the following week’s work. Unchanged material or a swapped facilitator points at the provider. A blank on the third question points at you, and that is the most common result.
What are the signs an AI enablement program has stalled?
Narrowing rather than falling usage. The same few prompts, the same two tools, no new use cases in a quarter, and polite mid-range feedback instead of complaints. Attendance usually holds, which is why stalling gets missed.
Should we fix the program or replace the provider?
Fix first if the practitioner understands your business and the material is the only stale part, because re-leveling costs a conversation. Replace when the curriculum has not changed in three or more quarters, or when the person you bought no longer delivers the sessions.
What does it cost to switch AI training vendors?
Beyond the new fee, budget four to eight weeks of calendar, one re-baselining exercise, and a session’s worth of credibility with senior users if you make them repeat an introduction. Exit charges depend on your notice window and any remaining session commitments, which is why renewal is the cheapest place to make this call.
Who owns the training materials built on our own documents?
Whatever your agreement says, and most say nothing. Ask for internal reuse rights to recordings, the prompt and workflow library as files you hold rather than a portal login, and the attendance and exercise records. Ask while unperformed work remains on the contract.
Can we switch providers in the middle of a program?
Yes, at a price set by your notice terms rather than your patience. Cancellation is usually priced on a ladder tied to days before the session date, so waiting three weeks often costs less than acting today. Pause rather than cancel if the next session is over a month out.
Is bringing AI enablement in-house cheaper than switching?
Only with a credible internal practitioner and a standing need. The recurring cost is not the sessions, it is keeping the material current as the tools change each quarter. Companies between 50 and 500 people usually end up splitting it: internal ownership of follow-through and measurement, external delivery of the sessions themselves.
How do we avoid repeating the same mistake with the replacement?
Write the diagnosis down before you shortlist anyone, then test every candidate against it. Ask who maintains the curriculum and on what cadence, which named person delivers, and what the next tier looks like for people who already have the basics. Hold back part of the budget until the month after the final session.
How long should a program keep producing results before it flattens?
Expect the easy gains in the first one to two quarters and a plateau after that unless the program re-levels. A plateau is normal and is not by itself evidence of a bad provider. What matters is whether they have a credible next tier and a maintenance cadence, or are selling the same introduction on repeat.
Key takeaways
- A program that worked and then stalled has usually drifted: the team outgrew the syllabus, the tools moved, the facilitator changed, or the sponsor stopped showing up. Only three of those are fixed by a new provider.
- Usage dashboards cannot detect decay. Capability narrows while the numbers hold, which is consistent with the 85% of employees who say their training does not help them use AI in their actual role.
- Call the incumbent and ask them to re-level before you build a shortlist. How they answer is the decision.
- Four moves are available, not one: re-level, narrow the scope, replace, or build internally. Narrowing is the most underused.
- Capture a capability baseline before the new provider starts, or you will have no way to prove the switch worked.
If you are weighing whether to renew, re-scope, or replace, a free AI-readiness call is a direct way to pressure-test the diagnosis before you spend another budget cycle on it. Bring what your team can already do, and what you expected by now.
Arthur Wandzel