Almost nobody. Ask the AI enablement market to tie a fee to whether your team is still using these tools in six months and you get refund windows, free trials, and satisfaction language. You will not get a priced outcome.
That is not pure vendor cowardice. Two of the three tests an outcome clause has to pass are missing in training work, and the third runs into an insurance problem most buyers never hear about. One thing is still worth putting in the contract, narrower than a guarantee and more useful.
What the market sells today
Start with the paper that already exists.
Coursera for Business publishes $399 per user per year for 2 to 499 employees with a 14-day refund policy, and revokes earned certificates if you refund. The same page cites 25% more productivity and 38% higher retention as figures its customers reported. Those are results other people got. Nothing commits them to you.
Udemy Business sells its team plan on a 14-day free trial. The two catalogs are now one company: the merger closed on May 11, 2026, after an agreement signed in December 2025. The biggest merger in corporate learning in years, and neither side brought an outcome clause into it.
General Assembly, which sells the longest and priciest programs in this set, puts it in writing on its own FAQ: it “can’t guarantee a job to any student or graduate.” Refund terms sit in state and country catalogs and vary by both. Smaller AI training shops land in the same place: a 14-day money-back window is the house standard.
A refund window guarantees you can change your mind in two weeks. It guarantees nothing about whether anything changed.
The three tests an outcome clause has to pass
The pricing literature agrees on what an outcome clause needs: the outcome has to be observable, attributable to the vendor, and contractible between the two parties. Run AI enablement through all three.
Observable: partly. Weekly active use can be seen if somebody instruments it before kickoff. Hours returned is not, and the honest measured figure is small. Bick, Blandin and Deming, in The Rapid Adoption of Generative AI (NBER working paper 32966, published in Management Science in 2026), recorded an average saving of 5.4% of work hours among people who use these tools, about 2.2 hours a week on a 40-hour week. You cannot see a shift that size in a 120-person company without having measured the week before.
Attributable: this is where it breaks. Adoption in month six moves because of the session, and because a manager started asking for it in standups, and because licenses finally reached everyone, and because two new hires arrived already fluent. No vendor can separate those four. No buyer can either.
Contractible: only with work you have not done yet. A clause needs a number, a window, and a named data source. Most buyers asking for a guarantee cannot name the data source, so the clause has no referee.
One test out of three is a weak base for risking money.
Where outcome pricing did take hold
It works in two adjacent markets, and both show why it does not work in this one.
Strategy consulting. Roughly a quarter of McKinsey’s global fees now come from performance-based arrangements. Michael Birshan, the firm’s managing partner for the UK, Ireland and Israel, said in November 2025 that the firm is “doing more performance-based arrangements with our clients.” Treat the number as the firm’s own account at a media event, not an audited figure.
The metric is what makes it work there. Clients arrive wanting procurement cost down 15% or factory throughput up 20%, and both sit in a system of record each side can read without arguing. Corporate training has no such system.
AI software. Armilla sells an insurance-backed performance warranty for AI products, underwritten through Lloyd’s of London with Chaucer Group, Greenlight Re and Swiss Re behind it. It pays when a model misses a defined threshold on output accuracy, hallucination rate, or jailbreak resistance.
Note what it does not cover: workforce enablement. A model’s behavior can be tested in a lab against a fixed input. A sales team’s behavior in March cannot.
The experiment that already ran
The purest outcome-based education contract ever run at volume was the income share agreement. Know how it ended before you ask a training firm to sign one.
BloomTech, formerly Lambda School, let students pay a share of their income once they earned $50,000 or more, and originated more than 11,000 of these agreements. On April 17, 2024, the Consumer Financial Protection Bureau entered a consent order against the company and its chief executive. The agency found the agreements were loans carrying an average finance charge of around $4,000, that students had been told otherwise, and that the company had misled them about the odds of getting a job.
Corporate procurement is business-to-business, so those consumer rules do not reach your contract. The structural lesson does. Once payment depends on a future outcome, the deal starts behaving like a financing instrument with disclosure duties attached, and training firms are not lenders.
The legal ceiling nobody mentions
This is the part that gets a guarantee struck by the vendor’s counsel, not their sales team.
In March 2025, Holland & Knight published a regulatory assessment of AI performance guarantees. The analysis turns on the principal object and purpose of the agreement. A guarantee can stay a non-insurance warranty when it is incidental to a real business activity and when it excludes, in the authors’ words, “fortuitous events outside of the offeror’s reasonable control.” They describe the area as “somewhat of a legal gray area” with little AI-specific precedent.
Now apply that to the clause you most want: a guarantee that your people are still using these tools at week 40. That depends on your managers’ calendars, your reorg, your license renewals, and whether anyone made room in the workweek for a new way of working. All outside the vendor’s reasonable control.
Ask a vendor to guarantee it anyway and you have asked them to underwrite your own management, which is the shape of an insurance product sold by a company with no license to sell one. A good lawyer reaches that conclusion in a single read, and the clause comes back redlined.
Why most clauses would be unenforceable anyway
Suppose the legal problem vanished. A clause still needs a number to pay against, and the published evidence says most buyers do not have one.
ATD surveyed 199 talent development professionals for Evaluating Learning: Getting to Measurements That Matter. Thirty-five percent said their organization evaluates the business results of learning programs to any extent. That report is several years old and still the most-cited figure in the field, which tells you how little ground has been broken since.
The TalentLMS 2026 L&D report, fielded in September 2025 across 101 US HR managers and 1,000 US employees trained in the prior year, found completion rates are still the primary measure of L&D success. Business impact was a secondary measure for 37%, roughly level with career growth at 31% and satisfaction at 28%.
Picking the right number is itself the work. The Wharton and GBK Collective study The Adaptive Advantage, published October 13, 2026, surveyed 1,416 US enterprise leaders and employees. Among teams tracking at least one growth measure, 35% reported strongly positive returns from AI, against 13% of teams watching efficiency alone. Pick the wrong metric for your clause and you buy an argument about a figure that was never going to predict anything. The measurement playbook works through which ones hold up.
Four at-risk structures, ranked by who will sign
Four shapes exist for putting part of a fee at risk. Only the first two get signed without a fight.
| Structure | What is measured | Who controls it | Will a vendor sign | What it costs you |
|---|---|---|---|---|
| Reach at risk | Who came, who finished, and whether exercises used your documents | The vendor | Yes, in writing | Nothing |
| Redelivery | The session runs again at no charge if a completion or quality bar is missed | The vendor | Often | A scheduling commitment |
| Milestone holdback | 10% to 20% of the fee released on a 90-day usage read | Shared | Sometimes, if you own the baseline | Instrumenting usage before kickoff |
| Keep-rate or hours-saved guarantee | Weekly active use or hours returned at 6 to 12 months | Mostly you | Almost never, and be wary of yes | A higher price and a future argument |
A fifth shape, the fee as a share of audited savings, comes up in conversation and never survives it. In a market where a workshop runs $2,000 to $15,000, the hours needed to agree on the savings number can exceed the fee at risk. That shape gets more plausible on the automation side, where market pricing runs $25,000 to $150,000 and the deliverable is a system you can test against a spec.
What we put at risk, and what we will not
We quote and run these programs, so this is the live answer rather than a framework.
Reach is the one we will put in writing before any baseline exists: who came, who finished, whether the exercises ran on the team’s own documents or a demo dataset, and a remedy if we miss. All four are ours to control, so all four belong in the contract. Redelivery too, since running a weak cohort again is cheaper than arguing about it.
The keep rate, no. Not without a baseline we helped you capture and a named owner on your side whose job includes making room for the practice. A vendor who guarantees a keep rate before a baseline exists is pricing a guess with your money, and a paid seat is not a capability no matter who signs for it. That argument is the whole case in the AI enablement manifesto.
One more thing to watch for. A vendor who agrees to every clause you send has priced the risk into the quote. You pay for a guarantee either way, and the only question is whether you can see the price.
How to write a clause that survives review
Six elements. A clause missing any one of them is a wish.
- A number. Not “strong engagement.” Thirty of 40 invited staff finish the second session.
- A window. Measured at 90 days from the last session, not “over time.”
- A named data source and reader. Workspace admin analytics, read by your IT lead, reported to both sides.
- A specific remedy. A credit of a stated amount, a free redelivery, or release of the holdback. Not “the parties will discuss.”
- A carve-out. Name what the vendor is not answerable for: a reorg, a license lapse, a hiring freeze.
- A baseline captured before kickoff. One week of measurement, taken before anyone teaches anything.
Then send the identical clause to every bidder. The one who redlines it with reasons has run enough rooms to know where it breaks. The one who signs it unread has not.
When to stop asking for a guarantee
Drop the word and ask three questions instead. What in the design produces weekly use at week 40. What you will measure to show it. And the numbers from their last three jobs of similar size, read at 90 days, with the name of whoever measured them.
A vendor with delivery history has those figures or can say honestly why not. One without will talk about engagement. That exchange costs nothing and tells you more than any clause.
If the quotes still look alike afterward, the next move is arithmetic rather than contract language. Converting each bid to a cost per hour saved reorders the field more often than it confirms it, and a model your finance team cannot dismiss is a better use of a week than negotiating an outcome you cannot measure.
Frequently asked questions
Does any AI enablement vendor offer a true outcome-based guarantee?
Not in any form we have been able to verify. What exists is 14-day refund windows, free trials, satisfaction language, and the occasional promise to redeliver a session. None of those pays against an adoption or productivity result. If a vendor says otherwise, ask to read the clause and look for the number, the window, the data source, and the remedy. One of the four is usually missing.
What portion of the fee is reasonable to hold back?
Ten to twenty percent, released on a 90-day read against a measure you set up before kickoff. Below 10% the vendor stops caring. Above 20% you are asking a small firm to finance your program, and the ones who accept are short of cash, not long on delivery.
Why won’t vendors guarantee adoption?
Because adoption at month six depends mostly on you: whether managers ask for the new way of working, whether licenses reach everyone, whether a reorg eats the quarter. Legal analysis of AI performance guarantees makes the same point from the other side. A warranty is meant to exclude events outside the offeror’s reasonable control, and your management is exactly that.
Is a money-back guarantee the same thing?
No, and conflating the two is how this market gets away with its current terms. A money-back guarantee refunds you for a product you did not want. An outcome guarantee pays you for a result that did not arrive. Every catalog vendor sells the first. Almost nobody sells the second.
What should the data source be?
A record that already exists and that neither side keeps for the clause’s sake. Workspace admin analytics from the AI tool itself, read by a named person on an agreed date, is usually cleanest. Never use a self-reported survey as the payment trigger: the party being paid should not supply the number.
Does outcome pricing make more sense for automation projects?
Much more. An automation has a spec, a test, and a measured output before and after, and the market range of $25,000 to $150,000 justifies the legal work. Workshop-scale engagements rarely clear that bar. If outcome-based structure is a hard requirement, it fits the build side of the relationship better than the teaching side.
What if a vendor does offer a keep-rate guarantee?
Ask what baseline they measure against and what happens if they miss. If the answer is “we’ll establish one during the engagement,” the clause pays against a number they control, which is worse than no clause. Payback-period math pressure-tests the same promise more honestly.
How long should the measurement window be?
Ninety days for a holdback. Long enough for the novelty to wear off, short enough that both sides still remember the terms. A 6-month window sounds more rigorous and mostly guarantees a dispute about what else changed in the business between kickoff and the read.
Do we need a baseline even if we only hold back 10%?
Yes, and it is the cheapest line in the exercise: one week of measurement before anyone teaches anything. Without it the 90-day read produces a number with nothing to compare against, and the holdback gets released on a chat rather than evidence.
Where to start
Draft the six-element clause for the measure you most want moved, before you shortlist. Send the same text to every bidder and keep their redlines. What each one strikes tells you more than anything else you get during the bake-off, and it arrives free.
Then instrument your baseline week. It takes a few hours of someone’s time and converts every later argument from opinion into arithmetic.
If you want a second read on a clause you are holding, or on a quote that claims to guarantee something, a free AI-readiness call is a direct way to get one. We run these programs, so we will also tell you when the guarantee you are chasing is not one you should want.
Arthur Wandzel
