The choice in front of you is almost never continue or cancel. With AI enablement it is nearly always a third thing: keep the goal, throw out the method. Collapsing those two decisions into one is how a recoverable program becomes a write-off, because the two wrong answers cost about the same and look nothing alike on the way out.
It helps to know the company you are in. S&P Global Market Intelligence, surveying more than 1,000 respondents across North America and Europe for a March 2025 report, found 42% had abandoned most of their AI initiatives, up from 17% the year before. Walking away is now ordinary corporate behavior rather than a confession. That lowers the political price of stopping, which is usually the real blocker.
So the question is not whether to abandon. It is what, exactly, you are abandoning.
The only test that counts is forward-looking
One question settles most of these cases. If you had spent nothing yet, and knew what you know now, would you approve this plan today?
A no means the money already gone is not an argument for the next invoice. That sounds obvious and is never how the meeting runs. Hal Arkes and Catherine Blumer measured the pull in 1985 with the ski-trip problem: given a $100 trip and a $50 trip booked on the same weekend, 54% of people took the costlier trip they expected to enjoy less. The spend, not the expected result, drove the choice.
The AI enablement sunk cost trap works the same way, with one extra twist. A company that has paid for two sessions and seen nothing often buys a third, because a third purchase is the only thing that makes the first two look like a plan instead of a mistake. The budget line defends itself.
Forward-looking does not mean starting from zero. Some of what you bought survives the decision: a baseline timing of one workflow, a roster of who can do what, a named reason the first attempt missed. Those are inputs to the next call, not losses. What you are writing off is a method, and methods are the cheapest part of the stack to replace.
Three separate things you can abandon
Pull the decision apart first. There are three separable objects, and their price tags are nowhere near each other.
| What you abandon | When that is right | Cost to switch |
|---|---|---|
| The method | A lecture-shaped session, generic sample data, no follow-up slot. The most common correct answer. | Low. Same budget, different format. |
| The provider | Method was right, delivery was weak, or the second-phase quote no longer matches the result. | Medium. New procurement, baseline survives. |
| The goal | The team’s work genuinely does not contain the tasks assistants help with. Rare. | High, and usually wrong. |
Most programs that feel dead are a method problem wearing a goal problem’s clothes. A team that sat through a vendor demo on invented sample data learned nothing transferable, and no amount of repetition fixes that format. Swapping the format is cheap; the market range for a hands-on workshop sits at roughly $2,000 to $15,000, which is normally a fraction of the license spend already committed.
The provider question is its own decision with its own evidence, and it is worth separating from the method question before anyone opens a contract. Our guide to renewing, renegotiating, or walking away from an enablement vendor works through the bargaining room you hold at each point in a term.
Abandoning the goal is defensible in one case: the work the team owns holds no recurring text, analysis, or document task. Check that reading against the work rather than against the complaints. In our experience, teams that call their work “too specialized” for an assistant usually mean the training they got never touched their own documents.
Four triggers to write down before you have results
A trigger agreed after the numbers arrive is a rationalization. Gary Klein’s premortem method in Harvard Business Review exists for this reason: the team commits to what failure will look like while nobody is defending anything yet. Four triggers cover most enablement programs. Write them into the program plan, with dates, before the first session.
Day 14, under 20% of the trained group used an assistant for real work. The session did not land. This licenses replacing the method, not the goal, and the replacement is cheapest inside 60 days while the material is still warm.
Day 90, under 25% weekly active. The content was generic. People can do the thing in a classroom and not in their inbox, which almost always traces to training built on sample data rather than the team’s own files. Change provider or format.
Day 180, usage healthy and no named output has moved. Stop buying training. The capability is in the building and the work was never redesigned to use it. More sessions here add cost and nothing else.
A second-phase quote above about 1.5x the first with the same measurement instruments. The cost broke the business case. If the vendor cannot show what the second phase measures that the first did not, you are funding repetition.
Two notes on using these. Measure usage from self-report alongside the admin console, because the console undercounts anyone working in a personal account. And pick a reporting rhythm before the triggers go live, since a monthly read and a quarterly read hide different things, as we lay out in monthly versus quarterly enablement reporting.
The check that makes stopping the wrong answer
Run one check before you act on any trigger, because it reverses the decision more often than anything else on this page.
Ask whether the capability is absent, or present and unused.
If people can show you work they now do differently and the hours still disappeared, the constraint is not the training. It is whose job it is to decide where saved time lands. Stopping the program at that point buys you nothing and quietly discards capability you already own. Microsoft’s 2026 Work Trend Index, run across 20,000 knowledge workers, attributes 67% of reported AI impact to organizational factors such as manager support and talent practices, against 32% for individual mindset. The thing that broke was often never in the room.
The older learning literature says the same in blunter terms. CEB Global, now part of Gartner, measured in 2014 that 45% of training delivered is never applied on the job, falling to 33% in organizations that measure and manage transfer. A program with no slot in the week for the new skill decays on schedule, and that is a calendar failure, not a curriculum one.
There is also a shape to usage that changes what counts as failure. The national panel run by Alexander Bick with Adam Blandin and David Deming found 33.3% of daily users saved four hours or more in the prior week against 11.5% of once-a-week users. Forty people using an assistant daily is a working program. Two hundred people using one monthly is a headcount statistic. Before you call a program dead on breadth, look at how hard the group that stuck is using it. Our measurement guide sets out the full instrument set, and the seven-mode diagnosis sorts which failures a second session makes worse.
What the failure statistics actually license
Three numbers get quoted at these meetings. Each means less than it is used to mean.
The MIT 95%. MIT’s Project NANDA reported in July 2025 that 95% of organizations were getting zero return on GenAI investment, with 60% of firms evaluating enterprise-grade tools, 20% reaching pilot and 5% reaching production. Read the methodology before you cite it: the authors label it preliminary, built from 52 structured interviews, 153 survey responses collected at four industry conferences, and a review of 300 public initiatives. It is a directional read on a self-selected sample, not a population estimate. The same report found external partnerships succeeded at twice the rate of internal builds, which is a finding a provider like us obviously likes, so treat it as directional too.
The Gartner 40%. Gartner’s June 2025 prediction that over 40% of agentic AI projects will be canceled by the end of 2027 is about building and running autonomous agents, on cost, unclear value and weak risk controls. It is not about teaching people to use assistants. Those are two purchases with two failure modes, and importing one failure rate into the other decision is how a cheap fix gets killed by a borrowed statistic.
Your peer group. The Census Bureau’s Business Trends and Outlook Survey put 32% of US firms with 100 to 249 employees using AI in a business function as of May 2026, with 37% among firms of 250 or more. Your comparison set is no longer companies doing nothing.
Together these license a stop. None of them makes one correct. A failure rate tells you that abandoning is defensible, not that abandoning is right for your cohort, and the difference is the four triggers above.
Where the money goes when you stop
A stop without a destination for the money is just a budget cut, and budget cuts do not survive the next planning cycle. Rank the alternatives by what each dollar buys.
- One measured workflow. Pick a recurring deliverable, time it this week, time it again in 90 days. Costs nothing and produces the number finance will accept.
- A provisioning fix. If self-reported use is healthy and the console is near empty, people learned the skill and are using it on accounts you do not sanction. Sanctioning what they already use is cheaper than any session.
- A manager decision. Name the backed-up work that freed hours go to. Free, and the most common missing piece. At the Bureau of Labor Statistics’ $46.89 per hour worked in June 2026 for private industry, the hours are real money only once they land on work that was genuinely waiting.
- A narrower hands-on pass. One team, their own documents, two recurring deliverables assigned by name. This is where the $2,000 to $15,000 market band earns its keep, and narrow beats company-wide every time.
- Custom automation. If the diagnosis says process rather than skill, you are looking at a build in the $25,000 to $150,000 market range. A different purchase, and one worth separating before anyone quotes you.
Options one through three cost almost nothing, which is the uncomfortable part. Programs usually stall for reasons that were never purchasable, a point we argue at length in why buying licenses is not the same as building capability.
Write the stop down, including the part that stings
One page, written the week you decide, with six lines: what you bought, what you measured, which trigger fired, what you are keeping, where the remaining money goes, and the date you revisit.
That memo is worth more than the program was. It converts a failed purchase into the only thing that makes the next one cheaper, and it protects the person who signed the first invoice. Teams that surface a bad result early tend to get a quick budget correction. Teams that report only progress get believed until renewal, and then get believed about nothing.
Expect pushback on the write-up rather than on the decision. People accept stopping; they resist recording why. Record it anyway, unflattering reading included, because the version that survives next year’s planning meeting is the one with a date and a number on it.
Frequently asked questions
How do I stop an AI enablement program without writing off what we spent?
Separate what you bought from what you learned. The sessions are gone, but the baseline timing, the roster of who can do what, and the named failure mode carry forward into the next attempt at a fraction of the original cost. State those three things explicitly in the decision memo so the spend reads as a first iteration rather than a loss.
What is the AI enablement sunk cost fallacy in practice?
Buying a third session because two have already been bought. The pull is measurable: Arkes and Blumer’s 1985 experiments found 54% of people chose a costlier option they expected to enjoy less, purely because of what had been spent. In enablement it shows up as a second phase that measures nothing the first phase did not.
When is it too early to abandon an enablement approach?
Before day 14, and before anyone has checked whether usage simply moved out of view. Novelty drives week-one numbers up and week-three numbers down, and a flat admin console is often a reporting artifact rather than a verdict. Give the program one honest usage read, including personal accounts, before you decide anything.
Should I change the provider or change the method first?
Change the method first, because it is cheaper and it is the more common cause. A generic session delivered well still fails. If the format already matched your team’s own work and delivery was the weak link, then the provider is the problem, and that is a separate decision with its own evidence.
How do I tell my CFO the program is not working?
Lead with the trigger that fired, not the feeling. Name the metric, the date, the threshold it missed, what you are stopping, and where the money goes instead. Finance will fund a corrected plan with a decision date more readily than a hopeful extension of the original.
Does a high industry failure rate mean we should stop?
No. It means stopping is defensible, which is a different claim. Both the 42% abandonment figure and MIT’s 95% describe populations, not your cohort. Use them to remove the stigma from the decision, then make the decision on your own day-14, day-90 and day-180 readings.
What if usage is healthy but nothing measurable has changed?
That is the one pattern where stopping is usually the wrong move. The capability is present and the work was never redesigned around it, which is a management decision rather than a training gap. Stopping the program discards the capability and leaves the actual constraint untouched.
How much should the next attempt cost after a failed program?
Less than the first, because it should be narrower. One team, their own documents, two named deliverables sits at the low end of the $2,000 to $15,000 range the market charges for hands-on sessions. If a vendor quotes more than the first phase without naming a new measurement, that is the cost trigger firing.
Can we pause instead of stopping?
Only with a date and a condition attached, otherwise a pause is a stop nobody wrote down. A useful pause names what must be true to resume, usually a manager decision about where saved hours land, and sets a review date inside 90 days.
Key takeaways
- The live decision is usually keep the goal, replace the method. Treating it as continue or cancel is what turns a fixable program into a write-off.
- Ask the forward question: knowing what you know now, would you approve this plan today? What you already spent is not an argument.
- Write four triggers with dates before the first session: day 14 under 20% usage, day 90 under 25% weekly active, day 180 with no output moved, and a second-phase quote above 1.5x with the same instruments.
- Run one check first. Capability present and work unchanged means stopping is the wrong call, because the constraint sits with the manager who owns the deliverable.
- Industry failure rates, including MIT’s 95% on a preliminary conference-recruited sample, make a stop defensible rather than correct.
- A stop needs a destination for the money, and the top three destinations cost almost nothing.
To pressure-test your own program against these triggers before you fund another phase, book a free AI readiness call. If the honest read is that your training worked and the problem is how the work is organized, we will say so rather than sell you a session.
Arthur Wandzel
