Name One Thing You Deleted Because of AI

Now answer for your own company. Name one workflow, one meeting, or one recurring report you deleted because of AI.Not made faster. Not made easier. Deleted, so that nobody does it anymore.

Most teams cannot name one, and that single blank explains a set of numbers the industry has been arguing about for a year.

The two numbers that will not reconcile

MIT’s Project NANDA published research in 2025 finding that roughly 95 percent of enterprise generative AI pilots produced no measurable effect on profit and loss. The work drew on 150 leader interviews, 350 employee surveys, and analysis of 300 public deployments. About 5 percent were extracting real value. The authors attributed the gap not to model quality but to what they called a learning gap in how organizations integrate the tools.

One caveat worth stating before someone else states it in my comments. That report is preliminary, it has not been peer reviewed, and parts of it rest on self reported outcomes. Treat it as a strong signal rather than a settled fact.

Now the other side. The 2026 AI and Data Leadership Executive Benchmark Survey asked about 110 large companies the same broad question and got a different answer. In that survey, 97.3 percent report delivering measurable business value from data and AI, and 39.1 percent say they have AI in production at scale, up from 4.7 percent two years ago.

Both cannot be right. And the more useful question is not which survey to believe, but what the gap is made of.

The Bolt On

Here is the trap, and it deserves a name. Call it the Bolt On.

A Bolt On is AI attached to an existing process without removing any part of that process. The work still happens. The meeting still runs. The report still gets produced. One step inside it now takes eleven minutes instead of forty.

That feels like progress, and on a stopwatch it is. But the organization did not get any capacity back, because the time saved did not leave the building. It got spent producing more of the same output, and more output arrives downstream as more work for somebody else. The team that saved four hours writing first drafts now spends four hours reviewing the extra drafts they generated. Net change to the business, close to zero, which is what the MIT research found when it went looking for profit and loss movement.

You bought capacity. You wanted leverage. Those are not the same purchase, and only one of them shows up on the income statement. It is worth reading next to the argument that product market fit is survivorship bias, because both run on the same error: measuring the thing you can see rather than the thing that moved.

The 2026 benchmark survey contains the confirmation, sitting in plain sight. Asked what stands between them and value, 93.2 percent of those executives named culture and change management. Only 6.8 percent blamed technology. The tools work. What did not change is any process, any meeting, any report, or anybody’s job.

Why nothing gets deleted

Deletion carries a political cost that addition never does.

Adding a tool creates a champion. Somebody sponsored it, ran the pilot, and owns a visible initiative with a budget line and a slide in the quarterly review. Removing a report creates an enemy. Every recurring artifact in your company has someone who requested it, someone who produces it, and someone who reads it, or at least claims to.

So the safe move is universal. Keep the report and use AI to write it faster. Keep the meeting and use AI to summarize it. Keep the approval chain and use AI to draft the request that moves through it. Nobody loses anything, which is the same as saying nobody gains anything.

The vendors are not going to correct this either. Software companies price on seats and usage, so a customer who deletes half their workflows is a customer who needs fewer seats. Nobody in your supply chain is paid to help you do less. That leaves the decision inside your building, made by people who each own a piece of the thing that would have to go.

Then there is the reporting layer, which rewards the wrong thing without anyone noticing. Adoption gets measured as usage. Licenses assigned, prompts run, weekly active users. Every one of those metrics goes up when you bolt on and stays flat when you delete. So the dashboard tells leadership the program is working at the exact moment it is producing nothing, which is the pattern I described in my piece on how good strategy gets killed, running on a new budget line.

Run the test. Two minutes.

Open the tool your team uses to track work. Pick the last two quarters.

List every recurring meeting, report, review, and approval step that existed before your AI rollout. Now mark the ones that no longer exist.

Most teams mark zero. A few mark one, and it tends to be a report that died for unrelated reasons and got attributed to AI after the fact.

Then run the second half, which is the harder one. For each item still on the list, write down what would have to be true for it to disappear. Not who would object. What would have to be true. Half of them will come back with an answer like the VP likes seeing it, which is not a reason, it is a person. The other half will come back with something real, and that is your actual roadmap.

The output is a two column page. What you deleted, and what you could delete if somebody decided to. Most companies have never produced that page, and it costs two minutes.

What the 5 percent did instead

The MIT research points at integration depth rather than tool quality, and the distinction is worth making concrete.

The 95 percent bought general tools and pointed them at general tasks. Draft this, summarize that, help me write the thing I was already writing. Useful, popular, and invisible on the income statement because the underlying process survived intact.

The 5 percent picked one expensive workflow and rebuilt it so that a step nobody enjoys stopped being a step. Not a faster version of the old sequence. A shorter sequence. That is why their results show up in profit and loss while everyone else reports high satisfaction and flat margins.

The difference is not budget and it is not model access. Everyone has the same models. The difference is whether anyone had the authority to remove work, and whether they used it.

The way out

Pick one process and set a deletion target before you buy anything else.

Name the workflow, name the step that goes away, name the date. Then measure the program on that, rather than on seats or prompts or weekly active users. If the step still exists in ninety days, the initiative failed, regardless of how good the satisfaction scores look.

Give someone the authority to remove work. This is the part most companies skip, and it is the whole thing. Adoption programs tend to be run by people who can add tools and cannot cancel meetings, which guarantees the outcome before anyone starts. If the person leading your AI effort cannot kill a recurring report without a committee, you have staffed a purchasing function and called it transformation.

Then audit the output side, not just the input side. Ask whether the extra volume your team now produces is reaching anyone who needed it. Faster drafts that create more reviews, more messages, and more meetings have moved cost rather than removed it, and the person absorbing that cost is seldom the person who bought the tool.

None of this is an argument against the technology. The technology works, and the survey data says so from every direction. The argument is that adoption got measured as addition, and addition was never how any of this paid off before.

Every technology that ever paid for itself removed something. Name what yours removed. If the answer is nothing, that is not a slow start. That is the result.