Back to Guides
Enterprise Software 15 min read

AI enablement, reconsidered: what the term gets right and what it hides

AI enablement, reconsidered: what the term gets right and what it hides

AI enablement is a useful term wrapped around a weak evidence base. It gets one thing right: buying access to a tool is not the same as being able to use it, and the distance between those states is work somebody has to pay for. What it hides is that almost every number used to sell it comes from self-report, that its central claim cannot be disproved, and that the word has become cover for two things that are not capability building at all. We run these workshops, so this is criticism of our own category. A buyer who cannot tell the defensible version from the rest will overpay for the wrong one.

What the term gets right

Take the strongest version of the case first, because the criticism only matters if the term is worth keeping.

Organizational AI adoption reached 88 percent in the Stanford HAI AI Index, 2026 edition. When nearly every company has bought in, having bought in stops being a position, and a second question arrives: of the people holding those licenses, how many can brief a model on a real deliverable, spot where it is confidently wrong, and decide which parts of their job should never go near it? That number is a different variable from seat count, and companies that confuse the two end up nine months into a rollout with nothing to show a board. We defined the term and its scope elsewhere; the short form is converting purchased access into demonstrated capability.

So the word names a gap that exists, which is more than most category labels manage. The problem is what got built on top of it.

The readiness argument cannot lose

The standard pitch runs like this. Enterprise AI results have disappointed. The explanation is never that the technology underdelivered; it is that the organization was not ready, with messy data, untrained staff, and governance bolted on after the first incident. The fix, in every version, is more enablement.

That argument has no losing condition. If a program produces no measurable change, the available conclusions are that it was too small, started too late, lacked sponsorship, or ran before the data was clean. Disappointment becomes evidence for the thesis rather than against it, and the firms supplying the explanation are the firms selling readiness assessments.

This is not a reason to dismiss the field. Unfalsifiable framing is a reason to insist on a falsifiable purchase. A program that states in advance what number would mean it failed is a different object from one that cannot fail, and little of what carries this label meets that bar.

The statistics AI enablement runs on do not survive a read

Four numbers do most of the work in AI enablement marketing. The most-quoted comes from MIT’s NANDA report, The GenAI Divide: State of AI in Business 2025: roughly 95 percent of enterprise generative AI pilots produced no measurable financial return.

Read the figure closely and it shrinks. The report is preliminary, and coverage cannot agree on its sample: some accounts describe 150 leader interviews and 350 employees surveyed, others 52 organizations and 153 senior leaders. The bigger problem is the definition. A pilot counts as failed if it did not reach full production, or showed no quantifiable benefit, inside six months, which makes a sensible nine-month rollout a failure. The 95 percent describes the companies in that dataset, not all enterprises. The same report carries a finding quoted far less often: external partnerships succeeded at roughly twice the rate of internal builds.

The pattern repeats. Vendor pages cite an 80 percent project failure rate, an 86 percent leadership unpreparedness figure, and a 40 percent productivity gain, almost always without the study, the sample, or the outcome measured. Corroboration for the direction exists, including S&P Global’s finding that a large share of companies abandoned most of their AI initiatives. The direction is not the issue. The issue is a category quoting precise percentages about something as loose as whether a pilot “worked.” A finance reader who checks one number and finds a six-month deadline inside it will discount every other number in the deck, which is why we have written separately about which statistics to cite and which to leave out.

Self-report is the weakest evidence in business

This is the criticism that should change how you buy.

In 2025, METR ran a randomized controlled trial with experienced open-source developers, assigning AI assistance issue by issue across 246 real tasks with 16 participants. They forecast that the tools would speed them up by about 24 percent, and afterward believed they had worked about 20 percent faster. Measured, they were 19 percent slower. Expert forecasters were wrong in the same direction and by more.

The honest reading cuts both ways, and citing this study without the caveats repeats the error the enablement pages make. Sixteen developers is a small sample in a specific setting, in codebases they knew well, with tools from early 2025. METR says plainly that the result does not show AI fails to speed up most developers, and now marks the finding as out of date. Its attempt to repeat the experiment in February 2026 could not be completed, because developers would no longer join a control group without the tools they had come to rely on. Nobody should present 19 percent as a standing fact.

What survives is the gap. Participants were wrong about their own speed by roughly 40 percentage points, in the flattering direction. Set that beside how a typical program reports its return: a post-session survey, a confidence score, a show of hands on whether people feel faster. If skilled practitioners misjudge their own throughput that badly while doing the task, a satisfaction survey measures mood with a capability label on it.

A claim about output needs a measure of output: same tasks, same scoring, before and after, with the second reading late enough to catch retained behavior rather than recall.

Programs that reward activity reward the wrong output

The measurement problem has a second half, and it is worse than imprecision.

Writing in Harvard Business Review in September 2025, Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano and Jeffrey Hancock named a pattern they called workslop: AI-generated work that looks finished and is not, pushing the thinking and the rework onto whoever receives it. A deck that covers the topic without making an argument. A summary that has to be checked line by line. The sender records a win; the cost lands downstream, uncounted.

Set that beside how most programs report progress: seats activated, prompts sent, weekly active users, sessions attended. All of those rise when someone generates more output, whether or not it was worth sending. A program graded on activity pays people to produce workslop and reports the volume as adoption. License usage dashboards overstate real capability for the same reason: an admin console counts whether the product was opened, which says nothing about judgment.

The useful version of fluency training spends its time on the opposite skill: not how to produce more, but how to tell whether what came back is usable, and which tasks should not be handed over at all. That is hard to demo and impossible to count in a console, which is why little of the market sells it.

Two things wearing the word

Two uses of the term are not capability building, and should be named.

Enablement as compliance paperwork. The duty is real. Article 4 of the EU AI Act, in force since 2 February 2025, requires providers and deployers to take measures supporting AI literacy among staff operating AI systems on their behalf. It sets no curriculum and no standard, so a recorded webinar and an attendance log satisfy the auditor. Plenty of what is sold as governance enablement exists to produce that record. It is a legitimate purchase under its own name, but it is not a capability program, and buying it as one is how companies come to believe they have trained a workforce that cannot pass a scored task.

Enablement as transition management. When large training mandates land in the same quarter as headcount reduction, staff read them together, and they are often right to. Citigroup’s January 2026 mandate putting 175,000 employees through AI training was reported as helping people reinvent themselves, while reporting on the wider pattern notes AI-justified layoffs that rarely come with evidence a system replaced the eliminated role. The measured effect of compulsion is resistance: staff routing around sanctioned tools in favor of their own, mandates met with filed paperwork and no change in behavior.

Both patterns spend the word. By the time a real program arrives, half the room has decided what “enablement” means.

The honest ceiling

One more piece of evidence bounds the whole conversation.

MIT’s Daron Acemoglu modeled the macroeconomic effect of current AI capability from task-level exposure estimates. His figure for total factor productivity is about 0.7 percent over ten years, with a GDP effect he puts near 1.1 percent and a ceiling around 1.8 percent. He gets there by noting that about 20 percent of US labor tasks are exposed to these tools, and only about a quarter of those can be done profitably today. MIT Sloan summarizes the prediction as nontrivial but modest.

Economists dispute it in both directions, and it is a forecast rather than a measurement. Its use is as a sanity check. If the ten-year economy-wide effect is single-digit percentage points, a vendor promising your 200-person company a transformation in a quarter is promising something the aggregate evidence does not support.

What survives the criticism of AI enablement

Strip out what the evidence does not carry and something smaller is left standing. It is unglamorous and it is real.

The claim What the evidence supports
AI transforms how your company works Specific tasks get meaningfully faster; the aggregate effect is modest and slow
Our training lifted productivity 40 percent Nothing, unless output was measured before and after on the same tasks
95 percent of AI projects fail, so you need us A preliminary finding with a six-month definition of failure
A workshop will fix it A workshop plus a baseline, a retest, and an owner might

The defensible shape of the work is narrow. One or two workflows that matter, worked on with the team’s own documents rather than a vendor’s examples. A scored baseline before anything is taught. A retest eight to twelve weeks later with no refresher, so the number means retention. A named internal owner with protected hours. An agreed figure that would count as failure, written down before the invoice. Facilitated fluency workshops run roughly $2,000 to $15,000 in the market depending on headcount and session count, and custom automation is a separate purchase in the $25,000 to $150,000 range that lands better once capability exists. The staged version of that program sits in our enablement playbook, and the argument for why access and capability are different purchases is in the longer piece behind this one.

None of that requires believing the hype. It requires believing that a team which cannot brief a model gets less out of the tools than a team that can, which is a claim small enough to test.

Six questions that tell the two versions apart

Put these to any proposal that uses the word.

  1. What would count as this program having failed? A vendor who cannot answer has sold you something unfalsifiable.
  2. What do you measure before? No baseline, no after-claim. This eliminates most of the market in one question.
  3. Is the measurement a test or a survey? Confidence scores measure mood. Scored tasks measure capability.
  4. Whose documents do we work on? Vendor examples teach the product. Your deliverables teach the skill.
  5. What happens in week ten? A program that ends with the session is an event.
  6. Where did that statistic come from? Ask for the study, the sample, and the outcome definition. The answer tells you how the rest of the proposal was built.

A proposal that answers all six in writing is describing capability work. One that answers none is a course list with a new label.

If you want a second opinion on which version is sitting in your inbox, we run a free AI-readiness call that works through your license data, your current usage, and what a program would have to measure to prove anything.

Frequently asked questions

Is AI enablement just a buzzword?

Partly. It names something real: the gap between buying AI access and being able to use it well. What makes it feel like a buzzword is that every vendor defines it as the scope of what they sell, so one proposal means a data platform and the next a two-hour workshop. Ask what the word covers before comparing prices.

What is the strongest criticism of AI enablement?

That its results are almost always self-reported. METR’s randomized trial found experienced developers were 19 percent slower with AI tools while believing they were 20 percent faster. If practitioners misjudge their own output by that margin, post-training surveys are not evidence of capability.

Does AI enablement actually work?

Narrow versions do, and they can be checked. A program with a scored baseline, work on the team’s own documents, and a retest eight to twelve weeks later produces a number you can defend. Company-wide mandates graded on attendance and license activity do not.

Is the statistic that 95 percent of AI pilots fail reliable?

Treat it with care. It comes from MIT’s NANDA report, a preliminary study whose sample is described inconsistently and which counts a pilot as failed if it did not reach production or show quantifiable benefit within six months. The direction is corroborated elsewhere; the precise figure is not a fact about all enterprises.

Why do AI training mandates face so much resistance?

Because compulsion signals something other than investment. When mandatory training arrives alongside headcount reduction, staff read the two together. Reporting shows significant bypassing of sanctioned tools and paperwork compliance without behavior change. Resistance answers the framing, not the tools.

What is workslop and why does it matter to training?

It is the term used in a 2025 Harvard Business Review piece for AI output that looks finished but shifts the real work onto whoever receives it. Programs graded on prompts sent and seats active reward producing more of it. Training built around checking and judgment works against it.

Should we buy AI enablement if we think the category is overhyped?

Possibly, at a smaller size than you are being quoted. Buy one program with a baseline and a retest on one workflow, at the low end of the market range, and let the measurement decide whether a second is worth it. Skepticism argues for a falsifiable first purchase, not for doing nothing.

Is AI literacy training legally required?

In the EU, something close to it is. Article 4 of the EU AI Act, in force since 2 February 2025, requires providers and deployers to take measures supporting AI literacy among staff operating AI systems on their behalf. It sets no curriculum, so a recorded session satisfies the duty, which makes it a compliance purchase rather than a capability one.

Key takeaways

  • With organizational adoption at 88 percent in the 2026 Stanford HAI AI Index, holding licenses is no longer a position, so the work of converting access into capability needed a name. That much the term gets right.
  • The readiness pitch has no losing condition, since every poor result is explained as not enough readiness. Insist on a purchase that states in advance what failure would look like.
  • The quoted 95 percent figure comes from a preliminary report that counts any pilot as failed if it showed no quantifiable benefit in six months.
  • METR’s randomized trial is the sharpest criticism of AI enablement available: participants were 19 percent slower while believing they were 20 percent faster. The perception gap is what invalidates survey-based reporting, not the number itself, which METR calls dated.
  • Activity metrics reward volume, and volume produces work that looks finished and is not. Measure scored tasks and rework instead of seats and prompts.
  • Acemoglu’s estimate of roughly 0.7 percent total factor productivity over a decade bounds the realistic promise.

Last Updated: Nov 18, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles