Replacing an AI training vendor mid-program

Replacing a provider partway through an AI enablement program fixes exactly one of the three things that make these programs fail, and it is not the most common one. Diagnose the cause first, then pick a switch date off the cancellation ladder in your agreement rather than off the date your patience ran out.

One disambiguation. Most published material on switching an AI training vendor is written for a machine learning team changing its data-annotation supplier or model host. This is the other reading: you bought a program to raise your team’s AI capability, sessions have landed badly, and someone has to decide before the next invoice.

Three causes, one set of symptoms

A failing enablement program looks the same from the outside whatever is wrong with it. Attendance slides from 28 people to 19 to 11, survey scores come back polite and middling, and four weeks later nobody can point to a task that is done differently.

Those symptoms have three plausible causes, and only one of them is the vendor.

The provider is wrong for the work. The curriculum is generic, the exercises run on their sample documents instead of yours, and the facilitator explains capabilities rather than sitting with someone while they rewrite a real deliverable. Replacing them helps here, and this is the case buyers assume they are in.

The scope was wrong when you signed it. The provider is competent and is delivering precisely what the statement of work describes: a 90-minute overview for a mixed room of 40 people. You were hoping for changed workflows. That gap is a specification failure, and a new vendor given the same brief delivers the same afternoon.

Nobody inside owns the follow-through. The sessions were fine and the reversion happened anyway, because no manager asked anyone to use the thing in the next sprint and the day-7 check-in nobody owned never happened. Replacing the vendor changes the logo on a deck.

Symptom	Likeliest cause	Does replacing the vendor fix it?
Generic examples, not your documents	Wrong provider	Yes
Facilitator could not answer workflow questions	Wrong provider	Yes
Mixed-seniority room, content pitched to the middle	Wrong scope	No
Nothing promised after the final delivery date	Wrong scope	No
High energy, no change by week three	No internal owner	No

Two of the three survive a vendor swap intact. That is the reason to spend a week on diagnosis: the real failure mode is not choosing badly twice, it is paying twice for the same unchanged outcome and concluding that AI enablement does not work.

The three questions that separate them

You can usually tell which cause you have in under an hour.

Open the session materials and look at the exercises. Were participants working on their own documents — a real proposal, a real board memo, a real client email — or on a case study about a fictional company? Practice on the team’s own work is the mechanism by which this training transfers at all, so generic exercises mean a provider problem and you can stop here.

Read the scope section of the agreement, not the proposal deck. List every deliverable with a date attached. If all of them are session dates and topics, you bought an event, and the provider is delivering one. The clauses worth rewriting are in our piece on the contract terms that decide whether AI training lands.

Name the person who followed up on day 7. Not whoever was supposed to — whoever did. If you cannot name them, the program had no internal owner and the next provider hits the same wall at the same depth. This is the most common cause and the least comfortable to find, because the remedy is internal.

Your contract decides when, not whether

Most corporate training agreements contain no termination clause in the sense procurement teams expect. They contain a cancellation schedule, a different instrument: it does not ask whether you may stop, it prices stopping by how much notice you give.

Where a real termination-for-convenience right exists, it is worth knowing why it was hard to get. Morgan Lewis’s sourcing practice, writing in July 2026, puts it plainly — “termination-for-convenience provisions are not just boilerplate” — and names what suppliers protect in resisting one: unrecovered upfront costs, term-based discounts that exist only because you committed, and capacity decisions made on the assumption you would stay.

Mid-program you have whatever you signed, so the question is not whether you may leave but what leaving on a given date costs. Published terms give you the shape. NobleProg’s US terms, which are public, set penalties as a percentage of the published course fee by written notice before the course date: nothing at 31 days or more, 50% at 21 to 30 days, and the full fee at 20 days or fewer. A transfer to another date carries a $50 per-delegate administration charge, and substituting a different attendee is free. Those are open-enrollment terms and catalogs change quarterly, so read your own agreement. The asymmetry is typical too: the provider may reschedule at any time and owes nothing for your costs.

The exit math runs on the calendar

Buyers are systematically wrong about how fast an AI vendor can be swapped. A Zapier survey of 542 US executives with active AI vendor contracts, reported by The Register in April 2026, found nearly 90% believed they could switch within four weeks and 41% believed they could do it in two to five business days. Of those that had attempted a platform migration, only 42% said it went smoothly — and the write-up does not say how many had tried, so treat that as directional.

For a training engagement the arithmetic is small enough to produce a date rather than a decision. Take a four-session program in the middle of the market range — an LLM fluency program for one team generally runs between $2,000 and $15,000 depending on scope and headcount — with two sessions left. If the next is 26 days out, a ladder like the one above costs half that session’s fee to cancel now and all of it inside 20 days. Moving the date may cost a two-figure administration charge.

That gap is the most useful thing on this page. A date transfer is cheap, a cancellation is not, and the diagnosis you have not done yet takes about a week. Pushing the next session out four weeks buys that window at almost no cost and keeps your negotiating position intact, because the provider still holds unperformed work and an interest in your goodwill. An emergency cancellation spends money to buy the same week.

There is a second reason not to sit with it. Arkes and Blumer’s 1985 paper in Organizational Behavior and Human Decision Processes named the sunk-cost effect: having invested, people keep investing so as not to have wasted what they already spent. “We have already paid for two more sessions” is that reasoning, and what it protects is the invoice. The money is gone either way; what is still depreciating is the team’s willingness to attend, and that is harder to re-buy.

What to get out before you give notice

A departing provider is the best placed party to brief a successor and the worst motivated to do it. Guidance on supplier exits from Lorman puts it crisply: settle transition assistance in advance, or “the vendor will likely extract a larger fee for the privilege when the customer’s need arises.” Mid-program you are past advance. What you still hold is unperformed work, an unpaid balance, and the ability to end things civilly.

Trade those for a handover. Ask in one email, before you mention replacing them:

• Session recordings, with written confirmation you may keep and reuse them internally.
• The prompt and workflow library built during the exercises, as a file you hold rather than a login to their portal.
• Attendance and exercise records per session, by name.
• The question log — what the team asked in sessions and in the chat. The most valuable item here, and the one nobody requests.
• The assessment instrument and baseline scores, if any were taken.

Do the rights check while you are there. Owning the inputs is not the same as owning something portable: a provider can hold copyright in every slide while you hold the documents its examples came from, leaving you a recording you may watch and a curriculum you may not hand to anyone else.

Overlap beats a hard stop

The expensive failure in a switch is rarely the exit. It is the quiet collapse afterward, when the replacement starts from a blank page and relearns by week three what the incumbent already knew.

So do not hard-cut. Keep the remaining dates on the calendar as placeholders and run one paid pilot session with the replacement before canceling anything. Three rules:

1. Run it with the hardest group, not the friendliest. The skeptical finance team is the test; a room of enthusiasts rates anyone well.
2. Use the same yardstick you used on the incumbent, held in your own hands: one short task-based measurement before and after.
3. Tell them it is a pilot. Providers who are good at this behave differently when the next decision rides on one session, and the ones who do not want to be measured say so now rather than in month three.

The pilot is also where you settle what shape the replacement should be: a single practitioner, a firm with a bench, and an internal owner with outside help are different purchases with different failure modes, compared in hiring a freelance AI trainer against an enablement agency. Vet the proof harder this time, too — the case study that persuaded you first time around deserves the checks in what to verify before you believe a vendor’s case study.

Briefing the replacement

What makes a mid-program replacement work is a one-page brief that stops the new provider running their standard opening session for a room that has already sat through one. Five things:

• What was covered, by session, one line each.
• What the team tried and rejected — which prompts failed, which workflow nobody adopted, which tool the sales team quietly stopped opening.
• The recurring objections, verbatim. “It makes things up.” “Faster to write it myself.” “Legal said no.” A provider who cannot answer those three in twenty minutes will not get the room back.
• Where the baseline sits, and what you will measure at day 30.
• Who the holdouts are, and whether they have authority. One unconvinced VP outweighs nine willing analysts.

Then set the expectation internally that this is a continuation, not a restart. A team’s tolerance for a second introductory session is far lower than for the first, and burning it is how a wrong-supplier problem becomes a company-wide belief that none of this works.

What to fix in the next agreement

Two changes do most of the work, and both are structural rather than adversarial.

Phase the commitment. Buy a paid pilot session, then the rest of the program on the pilot’s result. Morgan Lewis notes that sourcing deals increasingly split a relationship into phases, with firmer commitments only after an initial evaluation period — easy to do here, because one session is a legitimate standalone purchase.

Buy the four weeks after, not the day itself. A date-shaped scope cannot deliver a behavior-shaped outcome. Numbered deliverables dated after the final session — a reusable recording, a workflow library built from your own documents, one working session at roughly day 30, a named contact for 60 days — are what turn a delivery into a capability. That is the argument the case for building capability rather than buying licenses makes at the portfolio level, and the buy-versus-build decision for enablement capability works through whether an outside provider is the right instrument at all.

And name the internal owner in the agreement, as the person accountable for the one cause here that no vendor can fix.

Frequently asked questions

Can we cancel an AI training contract partway through?

Usually yes, at a price set by notice. Most agreements price cancellation on a ladder rather than granting a clean termination right: one published set of terms charges nothing at 31 or more days before the session, 50% at 21 to 30 days, and the full fee inside 20 days. The structure is common, the numbers vary.

How much does it cost to exit mid-program?

The unperformed sessions at whatever your notice window triggers, plus any minimum you committed to. For a program in the $2,000 to $15,000 market range, a mid-ladder exit typically costs half of one remaining session, not half the program.

How do I know whether the vendor is the problem?

Three checks: did exercises run on your documents or theirs, does the scope promise anything after the final session date, and can you name whoever followed up on day 7? Generic exercises indicate the provider, a date-shaped scope the specification, no day-7 owner the sponsorship. Only the first is fixed by switching.

Should we replace the vendor or pause the program?

Pause first if your next session is more than 30 days out. Diagnosis takes about a week and a date transfer costs far less than a cancellation, so you can pause, diagnose, then replace with evidence rather than buying the same outcome twice.

What should we get from the outgoing provider?

Recordings with internal reuse rights, the prompt and workflow library as a file you hold, attendance and exercise records, the log of questions the team asked, and any assessment baseline. Ask before you announce the change: a provider with unperformed work is far more cooperative than one whose last invoice has cleared.

How do we brief the replacement so the first session is not a repeat?

One page: what was covered, what the team tried and rejected, the recurring objections in their own words, where the baseline sits, and who the influential holdouts are. State explicitly that the opening overview has happened and will not be repeated.

Will switching reset the program timeline?

Expect four to eight weeks of added calendar for notice, a pilot session, and the brief. In the 2026 survey cited above, nearly 90% of executives expected to switch an AI vendor within four weeks and only 42% of attempted migrations went smoothly. Plan for the longer number.

Is it better to finish the contract and not renew?

Sometimes. Finish if the remaining sessions are cheap relative to the exit penalty and attendance is holding. Stop if attendance is falling: every weak session raises the cost of the next provider’s first.

Do we tell the team we are changing providers?

Yes, briefly, framed as a continuation: what you learned, what changes, and that nobody is repeating an introductory session. Silence invites the reading that the program is being quietly abandoned, which costs you the attendance the next session needs.

Key takeaways

• Three causes produce identical symptoms, and only the wrong-provider case is fixed by a replacement. Wrong scope and absent sponsorship survive the swap.
• Diagnose in an hour: whose documents were in the exercises, whether the scope extends past the last session date, and who actually followed up on day 7.
• Your agreement most likely prices exit by notice rather than granting a termination right, so the real decision is a date. Published ladders commonly run 0% at 31 or more days, 50% at 21 to 30, and 100% inside 20 — and a date transfer is cheap where a cancellation is not.
• Collect recordings, the workflow library, attendance records and the question log before you announce the change.
• Overlap rather than cut: one paid pilot with the hardest group, on the yardstick you applied to the incumbent, briefed on what the team already rejected.

If you are mid-program and unsure which of the three causes you have, that is a 30-minute conversation rather than a project. We run a free AI readiness call: bring the statement of work and one session’s materials, and we will tell you whether the provider is the problem or whether a new one hits the same wall.