Why AI Pilots Stall Without Controlled Work

What recent research shows about GenAI return, workflow redesign, and worker bypass, and how finishing work under named control answers the gap.

The demo works. The pilot gets a slide. Then Monday comes, and the work still moves the old way. Someone pastes a draft into a chat window, checks it by eye, and files the result by hand.

The tool sits beside the work. The process that already runs the business never takes the result back.

If you've inherited a pilot like that, you already know the feeling in the room. The technology did what the vendor showed. The operation stayed where it was. Three studies from the last two years describe that stall with more precision than the headlines, and they describe it with limits attached. Those limits belong in the telling.

I have written before about running a tech project and calling it a transformation. This piece stays with the next cut of that problem: the pilot that already demos, and the process that never receives a finished result. If you're still learning the tool itself, start with the executive's guide to getting started with AI. This article assumes the pilot exists.

What the research actually says

In July 2025, Project NANDA published The GenAI Divide: State of AI in Business 2025. Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari are the authors. The cover note calls the paper "Preliminary Findings from AI Implementation Research." The method is a review of more than 300 public initiatives, interviews with people from 52 organizations, and survey responses from 153 senior leaders, gathered from January through June 2025. The paper sets its result against $30 to $40 billion in enterprise GenAI investment.

The executive summary states the return finding in one sentence: "95% of organizations are getting zero return." The next sentence stays with it: "Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact."

That is a finding about organizations and about measurable profit and loss. It is preliminary. It comes from interviews, a survey, and public initiatives, and the authors say so.

The same summary gives a second finding. It counts a funnel for enterprise-grade systems, custom or vendor-sold: "Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production." The authors attribute most of that drop to "brittle workflows, lack of contextual learning, and misalignment with day-to-day operations."

Later the paper uses the phrase "the 95% failure rate for enterprise AI solutions" for the return split. Headlines shortened that phrase again, into a claim that 95 percent of pilots failed. The executive summary says that 95 percent of organizations are getting zero return. It separately counts how many custom or vendor tools reach a pilot and how many reach production. The funnel's limitation note is on the same spread: "These figures are directionally accurate based on individual interviews rather than official company reporting. Sample sizes vary by category, and success definitions may differ across organizations." The appendix adds that a six-month window can understate success for complex systems.

Keep both findings, and keep the caution. In this sample, most organizations saw no measurable P&L return. Custom and vendor tools fell off hard before production. The paper is preliminary, and it tells you that.

McKinsey's global survey, "The state of AI in 2025: Agents, innovation, and transformation," was published November 5, 2025. Fieldwork ran from June 25 to July 29, 2025, with 1,993 respondents in 105 nations.

McKinsey defines AI high performers as respondents who attribute an EBIT impact of 5 percent or more to AI and who report "significant" value. That group is about 6 percent of respondents. Two practices travel with them. "High performers are also nearly three times as likely as others are to say their organizations have fundamentally redesigned individual workflows." Redesigning workflows "has one of the strongest contributions to achieving meaningful business impact of all the factors tested." High performers are also more likely to have "defined processes to determine how and when model outputs need human validation to ensure accuracy."

The rest of the survey keeps the result in scale. 88 percent report regular AI use in at least one function. About one-third say their companies have begun to scale. 39 percent attribute any enterprise EBIT impact to AI. Among organizations using AI, 51 percent report at least one negative consequence, and nearly one-third report consequences from inaccuracy.

Use is widespread. Enterprise return shows up for a minority of respondents. The respondents who do report a meaningful return are much more likely to say they redesigned the workflow and decided when a person has to check the output.

WalkMe's State of Digital Adoption 2026 is the worker's side of the same picture. The company announced the report on April 9, 2026. The method section describes two surveys in the first quarter of 2026: 3,750 participants at organizations with 1,000 or more employees, across 14 countries. That total is 1,700 senior leaders and 2,050 office and hybrid workers.

The report says: "In any given month, more than half of workers abandon their enterprise tools at least once and do the work manually instead. For a third of them, it's a recurring pattern." It also says 37 percent of workers "skip AI entirely because using it means breaking their workflow and manually moving data between systems."

On the path around the official tools: "45% of workers used unapproved AI tools in the past 30 days. 36% used them with confidential company, customer, or employee data." On trust: "Only 9% trust AI enough to use it for high-impact work," while "61% of executives trust AI for complex work."

The work is getting finished. A large share of it is getting finished by hand, or in a tool the employer did not approve, sometimes with confidential data in the prompt. That is a control problem sitting inside an adoption problem.

The strongest objection

Here is the objection in the form a sharp operator would actually make.

Approval gates are why enterprise AI moves slowly. The firms that get value move fast with copilots, let skilled people improvise, and accept some mess. The widely shared line that 95 percent of AI pilots fail is a media simplification of a preliminary working paper, so a product story built on that line is standing on a soft fact. Workers bypass official tools because official tools are clumsy, and another layer of logging and approval makes them clumsier. A survey that shows high performers redesign workflows still leaves the cause open. Those firms may simply be better operators. They might have done well on any stack.

Most of that objection holds.

The NANDA paper is preliminary. It says so on the second page. The funnel figures rest on interviews, success definitions differ, and the authors warn that a short observation window can understate success. The executive summary says 95 percent of organizations are getting zero return, and it counts the path to production separately: evaluated, piloted, and in production.

Speed and judgment still matter. The same paper says generic tools such as ChatGPT and Copilot are widely explored, and the people interviewed described them as flexible and immediately useful. A copilot in a skilled person's hands is real help on individual work. The paper's point is that this help has been hard to see in the P&L.

The clumsy-tool point holds too. WalkMe's report says workers skip AI when using it means breaking the workflow and retyping data between systems. A logging layer pasted onto that kind of tool will get the same walk-around. People finish the job. They finish it on the path that lets them finish.

The operator point holds as well. McKinsey's high performers are a small group defined by reported results. The survey can tell you which practices they report alongside those results. It leaves open whether the practices caused the results, and whether a weaker operator would get the same outcome by copying the list. NANDA makes a similar caution about one of its own findings: firms that choose an outside partner may already differ in capability, so the correlation is not yet a cause.

Those concessions shrink the claim. They should. What remains is still enough to act on this week.

What still stands

McKinsey's high performers report two design choices together. They redesign the workflow. They define how and when a person has to check the model's output for accuracy. WalkMe shows the official tool losing the work. People complete it by hand, or they complete it in a tool the firm did not approve, and a meaningful share of that second path includes confidential data.

The control gap is already open. It is open in the workaround.

Between the Curves describes that gap as two curves pulling apart: what the tools can do and how widely people use them, and the risk that grows with that use. The risk sits between the work people actually do with AI and the work the firm can govern. The person who owns the work is the person who can judge whether the output is complete and accurate. IT can help with access and with the log. Accountability for the output stays with the function that does the work.

Put the approval on the moments that can spend money, touch a client, or send something outside the firm. Let the ordinary steps run, and write the finished result back to the record. That design is how the two high-performer practices land in a process you already run: a workflow that has been redesigned, and a defined point where a person checks the output. It is a practical answer you can run on Monday. The studies show which practices travel with reported return. They also leave room for the explanation above, that better operators would have found some way through on another stack.

What controlled work looks like on Monday

Start with a process the business already runs. A file lands in a folder. A form is submitted. A Slack message arrives. That trigger starts the firm's own steps. Each step is logged under a name. When a step would move money, touch a client, or send something outward, it waits. The owner approves it or sends it back. The finished result is written into the system of record the team already uses.

That is controlled work. The run produces a finished artifact. The log shows what happened and whose name is on it. The approval is narrow, and it is real.

I use the name Systems Intelligence for this idea. It is one concept with two angles. One angle is the business around the work: an owner who can make it move, decision rights that are clear, people who know the work, information you trust, and systems that fit the workflow. The other angle is the harness, the part that runs the automation. The harness takes the trigger, walks the firm's steps, keeps the log, holds the sensitive items for the owner, and returns the finished work.

The front door is live. The agents that read and write, and the learning loop that files a missing step and builds the next automation, are still being built. You can run the check below without waiting on those pieces.

Consultancies and MSPs can deliver an AI Execution Plan and the install. Unleash Epic provides the platform. If you want a conversation about one process, the company path is on the services page.

Why a rollout gets walked around

A rollout hands people a tool and asks them to change a habit, while the old process is still the one that finishes the job. The file still has to reach the record. The client still has to receive an answer. When the new tool makes that harder, people do the task the old way, or they do it in a tool they picked themselves.

WalkMe's report is the footprint of that choice. People abandon the enterprise tool and finish by hand. Many skip AI when it breaks the workflow. Others finish in an unapproved tool, and a share of that path includes confidential data. Executives trust these tools for complex work far more often than workers trust them for high-impact work.

That distance is the trust gap. The unapproved path is the control gap, already in motion. A process that finishes the work, logs the steps, and waits only where the risk is real gives people a path they can stay on.

Run the check on one live process

Five questions. Yes or no. Use them on one process that already has a trigger you can point at: a dropped file, a form, or a Slack message. The score tells you how many of the five pieces are in place, and which missing piece to fix first.

Do this in the next working session

Write the trigger in one line. Under it, write three things: the owner's name, where the log of steps will live, and the one action that must wait. Money, a client, or an outbound send. If you can't name the owner, that's the whole task. Name the owner before you add a tool.

The check above is the artifact. It is one page, and it is about a single live process. Run it this week.

FAQ

What did the MIT NANDA research actually say about generative AI return?

Project NANDA's July 2025 paper, The GenAI Divide, says 95 percent of organizations in its research are getting zero return, and that 5 percent of integrated AI pilots are extracting millions in value while the vast majority remain stuck with no measurable P&L impact. On custom or vendor tools, 60 percent of organizations evaluated them, 20 percent reached a pilot, and 5 percent reached production. The paper calls itself preliminary. It says the funnel figures come from interviews, and that success definitions differ by organization.

What does controlled work mean in practice?

Controlled work means a trigger starts a process the business already runs, each step is logged under a name, and money, client contact, and outbound sends wait for the owner. The run returns a finished result to the system of record.

Where should a person still approve model output?

McKinsey's 2025 survey found that AI high performers are more likely to have defined processes for how and when model outputs need human validation. In a controlled process, that check sits on the steps that spend money, touch a client, or send something outside the firm. The other steps can finish and write the result back to the record.