Most AI enablement renewals get decided backwards. Finance asks what next year costs, the vendor offers a discount for a two-year term, and somebody signs before anyone has settled whether last year’s sessions produced anything. The question that belongs first is cheaper to answer and harder to dodge: did the work change, and if not, was that the vendor’s doing or yours?
Settle that and the renew, renegotiate or walk-away call falls out almost on its own. Skip it and you will be wrong about half the time, because the two common failures look identical on a usage dashboard and call for opposite moves. One argues for buying more enablement. The other argues for buying none.
Here is the order, with the numbers that decide each branch.
Diagnose the failure before you price the renewal
Two failures wear the same clothes, and telling them apart is the whole decision.
Nobody uses the thing. Weekly use among the people you trained sits low, the sessions got good reviews, and three months on the assistant is open on four desks. That is a transfer problem. Either the content worked in the room and died on Monday, which is on the vendor, or your rollout gave nobody a reason to change, which is on you.
Everyone uses it and nothing moved. Use looks healthy, people are fluent, and no deliverable takes less time or costs less than a year ago. More training will not touch this. The capacity exists and nobody redirected it, which is a management job you cannot buy.
That split decides the renewal. The first failure can justify a second term with a changed format. The second cannot justify one at all, and a vendor who tells you the fix is more sessions is selling. Our walkthrough on diagnosing an enablement push that did not land separates the two, and it belongs before the notice window opens.
Run the diagnosis with the vendor in the room, not about them. The good ones will name the failure themselves.
There are four outcomes, not three
Renewal checklists offer three. The fourth saves the most money and gets written about least.
| Outcome | What has to be true | The common mistake |
|---|---|---|
| Renew as written | Use held, one output moved with a date on it, and the next group needs what the last one got | Re-signing because the relationship is pleasant and the invoice too small to fight |
| Renegotiate | A second year is worth it but the format, scope, reporting or price is wrong | Reopening price only, the lever with the least room in it |
| Replace | The need is real and this vendor cannot meet it, usually on format or on refusing to show data | Starting a search you cannot finish, then re-signing the incumbent in week eleven |
| Stop buying enablement | Fluency is fine and the constraint has moved to who owns the work | Treating the budget line as permanent because it exists |
The fourth row is a win, not a retreat. A company whose people are fluent and whose bottleneck is now a product or staffing decision should stop paying for training and go fix that. Enablement spend has a natural end, and no vendor will point at it.
The evidence that decides it
An AI training renewal decision needs one page of evidence. Every figure below is a share of the people you trained, never of headcount: the wrong denominator makes a small program look like a large failure. Return on the spend sits in the bottom three rows.
| What you measure | Renew | Reopen the terms | Walk away |
|---|---|---|---|
| Used an assistant on real work in the last 5 days | Over 50% | 30–50% | Under 30% after a follow-up pass already ran |
| Sustained use: active in 9 of the past 12 weeks | A quarter of the group clears it | 10–25% | Under 10% |
| Named workflows where an assistant is now standard | 2 or more | 1 | None at month 12 |
| Output change you can date | One measured | One named, not yet timed | Use is healthy and nothing moved |
| People who can run a short clinic for peers | 2 or more | 1 | None, and the vendor resists building any |
| Paid seats converting to real use | Over 60% | 40–60% | Under 35% |
Two rows need their reasoning spelled out. The nine-of-twelve-weeks test is not ours. Microsoft published it in its own analytics reference, where a habitual user averages 1 to 14 assistant actions a week and is active in at least 9 of the past 12 weeks. Borrow it whatever your stack runs on. It survives a quarter-end lull, and one enthusiast running bulk tasks cannot fake it.
The seat-conversion row rests on the widest public read of the gap between licensed and used. Recon Analytics surveyed more than 150,000 US paid subscribers between July 2025 and January 2026 and found 83.1% of people with workplace ChatGPT access converting into users against 35.8% for Copilot. Same firms, same policies, a 47-point spread. Below 35% the tool is provisioned and unused, and no training contract is your binding constraint.
Our read on what good looks like at six and twelve months names the studies behind the first row and which are the wrong size to compare against.
What to reopen, in the order that moves money
Buyers go at price first because price is the number on the page. It is the fourth-best lever.
1. Format. The biggest swing in value per dollar, and most vendors concede it because it costs them less too. A catalog nobody finishes becomes two facilitated cohorts. Two full training days become a half-day on patterns plus four 45-minute clinics across the quarter, which is how a habit survives past week three. Ask for the format that fits the failure you diagnosed, then price that.
2. Scope. Year one trains the willing. Year two has to reach the function furthest behind, usually finance or legal and never engineering. Name that group in the scope and cut the volunteers who no longer need a session. Smaller scope, sharper target, lower invoice.
3. Reporting and data rights. Ask for raw numbers rather than a satisfaction summary: attendance, follow-up activity, and what the vendor saw people struggle with. Ask also that everything built in your sessions, prompt patterns included, is yours to keep. A vendor who will not put that in writing is telling you next year’s dependency is the product.
4. Price, against a floor you can name. Self-serve platforms publish list prices, and those set the floor a custom engagement has to beat on something other than course count. Coursera’s Teams plan is posted at $399 per user per year for 2 to 499 licenses, and the other large catalogs land nearby for small teams. Live, hands-on work on your own documents sits in a different market, roughly $2,000 to $15,000 per engagement by scope and group size. One number is what breadth costs, the other what transfer costs. Make the vendor defend the gap.
5. Terms tied to something. Partial outcome linkage is reachable even where full outcome pricing is not, and the honest version is a holdback against a measurement milestone rather than a share of claimed savings. See who in this market will price against outcomes.
6. Term length and the renewal clause. A two-year term at a discount is a fair trade once a program has proven out and a bad one before. Either way, strike the evergreen clause or shorten the notice window. A second year with a 90-day auto-renewal buried in it means you make this decision once and then stop making it.
Walking away costs less here than in software
Procurement habits trained on software overstate what leaving a training vendor costs, which skews the negotiation. A software exit means migration, re-integration, retraining and a stretch where something is broken. A training exit means none of that: no runtime dependency, no data model to port, nothing that stops working on the termination date.
The switching cost is internal: hours to run a search and brief a new facilitator on how your team works. Call it a few weeks of somebody’s attention, which is not a lock-in.
The market gives you more room than most buyers use. ATD’s 2026 State of the Industry, reporting 2025 data, put direct learning expenditure at $846 per employee against $1,254 the year before while average formal learning hours rose from 13.7 to 16.7. Read it with care: the sample shrank from 539 organizations to 340, so some of the drop is composition rather than behavior. The direction holds. Buyers are getting more hours for less money by moving to cheaper delivery, the backdrop your vendor negotiates against whether or not they raise it.
One more reason to keep commitments short. Recon Analytics watched one assistant’s share of US paid subscribers fall from 18.8% to 11.5% in seven months. Where workers had both ChatGPT and Copilot available, 76% chose the first and 18% the second. People migrate toward whatever works, regardless of what you licensed. A curriculum welded to one product ages faster than a three-year term, so keep deals short and teach transferable patterns rather than one interface. That gap between buying tools and building capability is the case we make in why licenses are not capability.
Exit mechanics nobody plans until week two
Start 90 to 120 days out, not 30. Contracts here commonly carry a 30 to 90 day notice window, and a file opened inside it has narrowed to renew or scramble.
- Day 120 to 90: pull the numbers, run the diagnosis, pick your outcome.
- Day 90 to 60: tell the vendor, in writing, with the real reason. Good vendors save accounts here. The rest confirm your read.
- Day 60 to 30: if you are replacing, keep two alternates in play and ask each for a paid pilot session rather than a pitch. One session on your own documents beats any reference call.
- Day 30 to 0: notice delivered in the form the contract names, then the handover.
Three things to extract before the relationship ends, because afterwards nobody returns the call. Session materials and recordings, with a written license to reuse them internally. The prompt patterns built on your documents, exported somewhere you control. Attendance and follow-up data in raw form, the baseline for whatever comes next.
If the renewal is a scale-up decision in disguise, a pilot that worked and a question about going company-wide, treat it as that instead. The rules we use for judging a pilot before scaling it fit that case better than any renewal checklist.
What goes in the next contract
Whether you re-sign or hire somebody new, five clauses make the next decision easier.
- A named measurement owner on your side, with the figures from the table above and the date they get pulled.
- Raw data delivery rather than a satisfaction report, on a cadence written into the agreement.
- Your material stays yours, including everything built on your documents.
- A capability transfer clause: two of your people trained to run a refresher without the vendor.
- No evergreen clause, and a notice window short enough that you decide rather than default.
How to run those measures so the answer holds up with a CFO sits in our guide to proving capability moved.
Frequently asked questions
How do I decide whether to renew an AI training contract?
Diagnose the failure mode first, then read the numbers. Renew if weekly use among the people you trained holds above 50% and one deliverable has measurably changed. Reopen format and scope if use sits between 30% and 50%. Buy no further term if use is under 30% after a follow-up pass already ran, or if use is healthy and nothing moved in twelve months.
When should I start the renewal conversation?
Ninety to 120 days before the term ends. Notice windows commonly run 30 to 90 days, so a file opened at day 45 has already lost the option of a credible search. Day 120 leaves time to pull evidence, let the vendor answer a real critique, and talk to alternates.
What should I ask the vendor to prove before signing a second year?
Three things that should already exist: what share of the people they trained were still using an assistant on real work 90 days later, which workflows changed, and how many of your people can run a refresher without them. A vendor who answers with satisfaction scores has measured the room, not the work.
Is it cheaper to bring AI enablement in-house in year two?
Often for the refresher layer, rarely for the first pass. Once two or three of your own people can run a clinic off your patterns, internal delivery beats any vendor on cost and relevance. What it does poorly is first exposure for a skeptical group, where an outsider with no stake in your politics gets further. Usual answer: a smaller external engagement for new groups, clinics in-house for the rest.
Our vendor will not share usage or completion data. Is that a reason to leave?
On its own it is close to decisive. Enablement that cannot be measured cannot be renewed on evidence, which leaves every future decision resting on relationship. Ask once, in writing, with a named list. If the reply is a screenshot rather than the numbers behind it, write the data clause into the next agreement or write the vendor out of it.
We bought a course catalog and almost nobody finished anything. Cancel or switch?
Change format before changing vendor. Self-paced catalogs fail structurally rather than on quality: nothing in them creates a time, a room, or an obligation. The same spend moved into facilitated cohorts on your own work produces more change, and most catalog vendors will blend formats rather than lose the account.
How much should a second year of AI enablement cost?
Less than the first for the same reach. Year one pays for discovery, curriculum fit and the first facilitated sessions. Year two drops the discovery, reuses the patterns built on your documents, and leans on your own clinic runners for reinforcement. Live workshop engagements run roughly $2,000 to $15,000 by scope and group size, and published seat prices near $400 per user per year give you a floor to negotiate against.
Who should own the renewal decision?
Whoever owns the work the training was meant to change, with finance in the room and the training budget holder out of the chair. These renewals default to whoever signed last time, which quietly turns a performance decision into an administrative one. The person who can say whether a deliverable moved holds the pen.
Key takeaways
- Diagnose before you price. Low use is a transfer problem a changed format can fix. Healthy use with no moved output is a management problem more training cannot touch.
- Four outcomes, and stopping altogether is a legitimate one. Enablement spend has a natural end and no vendor will point it out.
- Measure the people you trained, not headcount, on a sustained-use test rather than a single week. Nine active weeks out of twelve is a public standard worth borrowing.
- Reopen format, scope and data rights before price. Format is where value per dollar moves, and vendors concede it sooner than margin.
- Leaving a training vendor costs far less than leaving a software vendor. No migration, so negotiate with the room that gives you, starting at day 120 and striking the evergreen clause.
If you want a second read on your own numbers before a renewal lands, book a free AI readiness call. We will tell you plainly which of the four outcomes your evidence points to, including the one where you stop paying anybody for training.
Arthur Wandzel
