What separates the six percent of companies capturing real value from the other ninety-four percent isn’t better models. It’s people who want the tools to work.

A quality engineer at a medical device plant sat through the demo and nodded in the right places. The AI assistant the company had spent eight months building would help her draft deviation investigations. It would pull prior cases. It would flag patterns. It would draft most of the regulatory write-up before she ever touched it. Her director was proud of the thing. Her CFO had signed the check. Six weeks after launch, the usage logs told the story. She had opened the tool twice. She was still drafting investigations by hand, the same way she had since 2017. She was not lazy. She was not afraid of change. She was an experienced engineer who knew that one day an FDA investigator might sit across the table and ask her to explain a paragraph in one of her deviations, and she was not going to put words in front of that investigator that she had not written herself. The tool was excellent. The workflow it had been bolted onto was not hers. The trust was not there, and no amount of fine-tuning was going to close the gap. This is the part of the AI conversation that gets the least airtime in the boardroom. It is also the part that decides everything.

The Adoption Illusion

The headline numbers from McKinsey’s 2025 State of AI report read like a victory lap. Eighty-eight percent of organizations now use AI regularly in at least one business function, up from seventy-eight percent a year earlier. Two-thirds use it in multiple functions. Stop there and you would think the job was mostly finished.

Read past the headline and the picture changes. Only thirty-nine percent of those companies attribute any EBIT impact at all to their AI use. Among the ones that do, most report less than five percent of EBIT attributable to AI. Just six percent qualify as high performers, where AI is contributing five percent or more to enterprise EBIT.

Adoption is mainstream. Value is rare. BCG’s own research finds that seventy-four percent of companies struggle to achieve and scale value from AI. The gap between the dashboards that show “AI in use” and the income statement that shows AI actually moving the number has never been wider.

For years we assumed the constraint was the technology. It isn’t, and it hasn’t been for a while. What decides the outcome now is what happens between deployment and the moment a busy person decides whether to open the tool on a Tuesday morning when there are already fourteen other things to do.

Why Engineering is the Easy Part

The models work. The APIs are stable. Generative AI can summarize a ninety-page contract, draft a customer reply,

surface a quality trend buried in three years of complaint files. Agentic AI chains those actions together, takes a goal, and executes against it across systems. In our work with regulated manufacturers, we have watched a well-built agent collapse a four-hour root-cause investigation into a forty-minute review. The engineering is not where the difficulty lives anymore.

The difficulty lives in the workflow the agent is supposed to navigate. Most of the processes inside large companies are not designed. They are accumulated. A step gets added because a regulator asked about it in 2014. An exception path gets carved because one customer threatened to leave in 2019. A spreadsheet gets bolted on because the ERP could not handle a corner case. By the time the AI team shows up, what looks like “the process” is twenty years of well-meaning patches stacked on top of each other, and nobody alive in the building can explain why all the steps are there.

You automate that and you have not transformed anything. You have scaled dysfunction at machine speed.

The data on this is blunt. McKinsey reports that AI high performers are 2.8 times more likely than peers to redesign workflows fundamentally. Fifty-five percent have done it. Twenty percent of everyone else has. The companies winning are not the ones with the best models. They are the ones who cleaned the floor before they laid the tile.

Clean the workflow first and generative AI and agentic AI stop looking like replacements. They start working like force multipliers. People get hours back. The quality floor holds even on a bad day, and attention shifts to the parts of the job that actually need a human in the chair.

Trust is the Real Infrastructure

There is a second reason adoption stalls, and most leaders would rather not say it out loud. Employees do not trust how AI will be used on them, or that the tools they are handed will protect them when something goes wrong.

Forty-seven percent of organizations have already experienced at least one negative consequence from generative AI use. Every employee has heard the stories. The legal brief with the fabricated case citation. The customer service bot that promised a refund the company would not honor. The hiring tool that quietly screened out candidates the company actually wanted to interview. People whose work gets audited, regulated, sued, or read by a customer have absorbed those stories, and they are doing the rational thing. They are waiting.

BCG’s survey of nearly twelve thousand frontline employees, managers, and leaders found that seventy-two percent say AI has changed the expectations for the skills they should have. Read that number carefully. It is not a statistic about excitement. It is a statistic about pressure. Three out of four people in the workforce believe the bar has moved on them, and most are figuring out alone whether the company will help them clear it or quietly replace them for not clearing it fast enough.

You will not close that gap with a town hall, or with a mandatory training module. You close it the way trust has always been built, which is by showing people, in their own work, that the tool helps them be better at the thing they care about being good at.

The McKinsey data shows what that looks like in practice. Sixty-five percent of AI high performers have defined human-in-the-loop validation processes for model outputs. Twenty-three percent of everyone else have. Call that a trust mechanism rather than a technical control, because that is how the person doing the work experiences it. When she knows her judgment is the final say on what reaches a customer or a regulator, she will open the tool. McKinsey’s broader Rewired research, built on more than two hundred at-scale AI transformations, identifies six dimensions that separate high performers. Five of the six are about people, not technology. The authors of Section 6 of Rewired make the point with a directness most consulting books avoid.

“The user is not a problem to be managed at the end of the project. The user is the project.”

Map the Stakeholders or Stop the Project

Every AI initiative inside a company touches at least three audiences. There is the process owner whose work is changing. There is the internal customer downstream of that process, who has been receiving an output of a certain shape for years and is about to receive a slightly different shape. There is the external customer, or the regulator, or the auditor, who eventually sees the consequence of whatever the AI produced.

Most pilots focus on the first audience and quietly hope the other two will not notice. They notice. Sales operations rolls out an AI lead scoring model and forgets that the field reps were its real audience. Finance automates an accrual workflow and forgets that the controller signs the 10-Q based on numbers the model now produces. Customer service deploys a chatbot and forgets that a customer in a bad moment will project the entire company onto whatever it says.

Before any initiative gets resourced, we make somebody answer for all three. What does winning look like for the person who owns the process once it changes? What lands differently for whoever consumes the output? And what about the customer or regulator at the end of the chain, better, same, or worse? If those answers don’t come back in plain language, the project isn’t ready to build. It’s ready to plan, which is a different and earlier thing.

Make it Compelling Enough to Want

The goal is not just adoption. The goal is ultimately pull.

Adoption is what you chase when people are not asking for the tool. Pull is what you get when they tell you which workflow to clean up next, because they have seen what happened in the last one and they want that, too.

Pull comes from a sequence almost no one follows. Pick a workflow that matters to the people doing it, not to the executive sponsoring it. Spend more time with those people than with the technology. Strip the workflow back to what it is actually trying to accomplish. Build something small that gives them back time on the thing they hate doing. Let them tell their colleagues.

When that loop runs three or four times, something shifts that no transformation office can manufacture by edict. People start scanning their own work for the next candidate without anyone telling them to. That is the culture change. Searching for the next opportunity stops being a quarterly initiative and turns into a habit, at every level, including the levels nobody was managing.

That is the prize. Not a roadmap full of pilots. An organization where the people closest to the work are the ones spotting the opportunities, because they have seen what happens when the tools, the workflow, and the trust line up at the same time.

What to do on Monday Morning

If you want a first move, here is one that has worked. Pick a workflow whose owner is already asking for help, not the one your steering committee circled. Send someone to sit beside that owner for a week before anyone scopes a build. Then change what the pilot is graded on. Counting tools deployed tells you nothing. The number that matters is whether the user came back on her own asking which workflow to clean up next. When that starts happening, the culture is turning.

A Note to the People Writing the Checks

Here is the uncomfortable part for the C-suite reader. Most top-down AI mandates fail because the people writing them are too far from the actual work to know what they are really asking for. The slide says “automate customer onboarding.” The reality is forty-three handoffs across six systems, governed by a policy document that contradicts itself in two places, with a compliance review that nobody has updated since the last reorg.

You cannot fix that from a steering committee. You fix it by sending people who can sit in the chair of the person doing the work and tell you the truth about what they saw. That is the part that gets called “ethnography” when consultants describe it and “common sense” when an operator describes it. Either word is fine. What matters is that somebody who is not trying to defend the existing process tells you what the existing process really is and why they do what they do, before you spend eight months automating it.

The six percent of companies pulling away on AI are not smarter than the other ninety-four percent. They are doing the unglamorous work first. They clean the workflows, earn the trust one team at a time, and let the results get loud enough that the next group of employees shows up asking for their turn instead of waiting to be told.

The companies creating durable AI advantage are no longer distinguished by access to better models. They are distinguished by their ability to redesign work, earn trust, and build organizations that continually discover the next opportunity for AI. Technology may start the transformation, but human systems determine who wins.

The engineering really is the easy part. The rest of it is the work.

Sources

McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation (November 2025)

McKinsey & Company, Rewired: The McKinsey Guide to Outcompeting in the Age of Digital and AI, Section 6 (“The Keys to Unlock Adoption and Scaling”)

BCG, Building the Workforce of the Future (2025), survey of ~12,000 frontline employees, managers, and leaders

BCG, AI at Work 2024: Where’s the Value? (74% scaling-value finding)