Two consulting firms published research this summer that describes a gap neither of them explains. The explanation is one question long.
Start with six figures. None of them are mine.
McKinsey, June: 88 percent of organizations now use AI in at least one business function. Around 1 percent consider themselves fully mature. Roughly two-thirds have yet to scale beyond isolated pilots.
Deloitte, July: worker access to AI expanded by half in a single year. Fewer than 60 percent of workers with access actually use it in daily work. 84 percent of organizations have not redesigned jobs or workflows around AI. And only half of executives regularly verify the quality of AI outputs when making decisions.
Eighty-eight to one. That is not a slow adoption curve. It is a wall.
And a third set, which neither paper cites and which changes how the first two read. Statistics Canada, July: 35.9 percent of Canadian workers used generative AI in their main job in the year to March 2026, against 12.2 percent of Canadian businesses using AI at all.
Both papers are worth reading, and I want to be clear before I go further that I think both are right about what they set out to address.
What they get right
Deloitte’s argument is that adoption metrics are easy to produce and easy to fake — a worker opens the tool, clicks enough to register as active, and goes back to working exactly as before. Their proposed answer is adaptation: judgment, experimentation, divergent thinking, measured by behavior rather than by usage counts.
That is correct, and the distinction between did they use it and did they use it to change anything is a real one that most organizations are not drawing.
McKinsey’s argument is that the difference between companies getting value from AI and companies running pilots is not a technology gap. It is an operating model — agents treated as teammates, a knowledge layer that agents can actually reach, architecture designed for interchangeability, autonomy earned progressively, platform centralized and tasks decentralized.
Also correct, and the observation that fifteen companies converged on the same practices independently is genuinely interesting.
Neither paper is wrong. I want to point at something both of them contain and neither follows home.
The finding Deloitte published and did not pursue
Buried in the section on judgment, Deloitte cites research showing that sustained AI use can erode professionals’ confidence in challenging AI recommendations — even when they possess the expertise to do so.
Read that again. Not that people lack the judgment. That people who have the judgment progressively stop applying it.
And a few lines later, on safeguards:
“Human-in-the-loop” principles risk turning workers into passive validators rather than active decision-makers.
They also report that only half of executives verify AI output quality, and cite research on models responding to fact-checks by deploying persuasive tactics to defend their original answers — which, in Deloitte’s own words, calls into question whether human-in-the-loop safeguards are truly effective.
So in one paper, published by a Big Four firm, we have three things established:
- the capacity to disagree with the machine degrades under sustained exposure;
- the primary safeguard turns the human into a validator rather than a decision-maker;
- and the verification that safeguard depends on happens about half the time.
That is a complete description of a control mechanism failing.
One thing I cannot leave out
I would rather not interrupt the assessment here, but leaving this unsaid would be a form of dishonesty.
Those three findings do not describe how I have worked with these systems — across twenty-eight years and four generations of the technologies that preceded them. The capacity to disagree has not eroded. The verification has not become occasional. And I have not been turned into a validator.
That is not one anecdote against a survey. It is a record, published as it happened and never edited, running from a 1998 defense engagement through every wave since.
What it establishes is that these three outcomes are not properties of working with AI. They are properties of working with AI under particular conditions — and Deloitte measured organizations that never established different ones. The finding is accurate about what was sampled. It does not generalize to arrangements built to prevent exactly this.
The distinction determines the remedy. If the erosion is inherent, coaching is all you have. If it is conditional, the conditions are what you build.
And the conditions are not aspirational. The tendency toward agreement is real and it is constant — which is why I measure it. I track the depth at which agreement is reached rather than whether agreement occurred, because fast consensus is the warning sign rather than the result. When that depth falls, it is visible before the output degrades, and it gets reset.
That is the difference between a trait to be coached and a condition to be instrumented. Deloitte’s own data says half of executives verify AI outputs. The question is not how to make them more diligent. It is what would tell them when they had stopped.
Where I part company
Deloitte’s remedy for all of this is behavioral. Coaching agents that analyze cognitive patterns. Nudges. Peer connections. Reflection sessions. Simulations. Experience design that builds judgment over time.
Every one of those is a reasonable thing to do. None of them addresses what the paper just described.
You cannot coach someone into having an independent basis for disagreement.
Judgment is not a personal attribute that gets applied in a vacuum. A person can disagree with a recommendation when they are standing somewhere the recommendation is not — holding information the system does not have, seeing a condition the system cannot see. That is not confidence. That is position.
Take the position away and no amount of training restores the capacity. You get a better-trained validator. Which is exactly what Deloitte observed, and exactly what its own remedy would produce more of.
Redesigned for whom?
The erosion is not a behavioral failure. It is a structural one — and the tell sits in their most-quoted number, though not in the direction it is usually read.
84 percent of organizations have not redesigned jobs or workflows around AI.
Read the phrase carefully. Around AI. Around the technology. The implied measure of progress is how far the organization has been reshaped to fit the tool.
Which raises a question the paper does not ask: redesigned for whom?
Deloitte’s own argument is that the value comes from human judgment — knowing when to challenge the machine, when to verify it, when not to use it at all. And they name the difficulty directly: the more work AI absorbs, the more judgment is needed to direct and evaluate it, and the fewer occasions workers have to develop that judgment in the first place.
Redesigning jobs around AI is how you absorb more work. It is the mechanism that produces the paradox, offered as the remedy for it.
To be fair to the authors, that is not the only available reading. They also write about protecting deep work, preserving space for divergent thinking, and designing the interaction rather than subordinating people to the tool. A generous reading of around AI is “designed thoughtfully for both.”
But even under that reading, the direction of fit runs one way — from the operation toward the technology — and the metric counts movement in that direction as progress.
And there is a further reason not to read that 84 percent as a failure to act.
The workforce has already redesigned.
Statistics Canada, drawing on supplementary questions to the Labour Force Survey, reports that 35.9 percent of Canadian workers used generative AI as part of their main job in the twelve months to March 2026. Over the same period, 12.2 percent of Canadian businesses used AI at all — and 66.7 percent have no plans to adopt it.
Roughly three workers using it for every business that has adopted it. Statistics Canada’s own observation, in an earlier release comparing the two, is that the pattern may indicate that adoption is not solely firm-led.
Microsoft’s 2026 Work Trend Index reaches the same place from a different direction: across 20,000 knowledge workers in ten countries, 16 percent of AI users have redesigned their workflows around AI.
Sixteen and eighty-four. Two independent studies, different populations, the same split.
So the 84 percent are not standing still. Their people redesigned without them — individually, informally, in the absence of any policy to defy. Statistics Canada found that only 5 percent of non-users cited company policy as a reason; the dominant answers were that it did not apply to their job, or that they were not interested.
That is not defiance. It is a vacuum being filled, and the formal operating model now describes an operation that has already moved.
The question is not whether to redesign. It is what you redesign around — and whether you have noticed that some of it happened already.
In an excerpt from a paper I wrote in the fall of 2004 titled Acres of Diamonds, I made the following statement:
“It is my position that a true centralization of procurement objectives requires a decentralized architecture that is based on the real-world operating attributes of all transactional stakeholders starting at the local or regional level. In other words, your organization gains control of it’s spend environment by relinquishing centralized functional control in favor of operational efficiencies on the front lines. This is the cornerstone of agent-based modeling.”
That was written about procurement software, two decades before any of this. The architecture is derived from the real-world operating attributes of the people doing the work. Not chosen because it performs better. Derived — you find out how the front line actually operates, and the structure follows from that.
Reverse the direction and you get an operation shaped by whatever the current technology happens to require.
And a note on where the seven truths came from
McKinsey’s operating truths were drawn from fifteen AI-centric companies — a four-person agriculture technology venture, a Series A marketplace, a Series D fintech, a seed-stage AI company, a digital health scale-up.
I want to be careful here, because this is an observation about sampling rather than a criticism of the work.
Those organizations had no legacy operating reality to validate. A four-person venture has no accumulated set of declared processes that quietly stopped matching what people actually do. The map was drawn last week; of course it matches the territory.
Which is why the truths begin at design. There was nothing to trace.
The 88 percent are not in that position. They have decades of process documentation, org charts, workflow diagrams and system configurations describing an operation — and no established way of knowing whether the description still corresponds to what happens.
Advice derived from organizations that cannot have the problem is not wrong. It is simply addressed to a different situation than the one most readers are in.
The missing question
In 1998 I worked on a defense maintenance operation delivering parts on time 51 percent of the time against a 90 percent requirement.
Everyone involved believed they understood the problem. Procurement was being held accountable. Supplier performance was being scrutinized. The process documentation was accurate and current, and everybody had read it.
The question that changed the outcome was not a procurement question, a technology question, or a change management question:
What time of day do orders come in?
Late afternoon. Almost all of them, around four o’clock — because service technicians were holding order releases and batching them at the end of the day. Releasing parts orders as they arose interrupted service calls, and technicians were measured on call volume.
A rational response to how they were being measured. One department away from the people being blamed. And not represented anywhere in the declared process.
Delivery performance went from 51 percent to 97.3 percent in three months. No new system. The technology was selected afterward, once the result was substantially in hand — which is why no platform can be credited with it.
What that question actually is
It is not a clever question. It is a validation question. It tests whether the declared operating model corresponds to the operating reality.
Both of these papers assume that correspondence.
Deloitte’s adaptation loop adapts people to a system. McKinsey’s seven truths design an operating model around a business. Both are sound engineering. Both begin one layer above the question of whether the thing being adapted to, or designed around, describes what actually happens.
And if it does not, then everything downstream inherits the error — faster, more consistently, and with better instrumentation than before.
That is what the 88 and the 1 have between them. Not a technology gap. Not a behavior gap. The distance between the operation people believe they are running and the one they are actually running.
What I would do with these papers
Read both. They are good.
Then, before adopting either framework, spend a week answering one question about your own operation: where does what we say we do stop matching what we actually do?
Not through a survey. Not through process mapping, which will return the declared version, because that is what it is designed to capture. By going and looking at when things actually happen, who actually decides, and what people are actually measured on.
If the answer is nowhere, you are in a very small minority and both papers will serve you well.
If you find a four-o’clock — and most operations have one — then you have just learned something worth more than any adaptation program, and you learned it before spending the budget.
Their numbers are the case. The question is the part neither paper contains.
-30-
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Sources: “The seven operating truths of AI-native companies,” McKinsey & Company, 11 June 2026. “AI adoption to adaptation,” Deloitte Insights, 9 July 2026. “Use of generative artificial intelligence tools among Canadian workers, March 2026,” Statistics Canada, 30 July 2026. Work Trend Index 2026, Microsoft.
Related
Their Numbers, Not Mine
Posted on August 7, 2026
0
Two consulting firms published research this summer that describes a gap neither of them explains. The explanation is one question long.
Start with six figures. None of them are mine.
McKinsey, June: 88 percent of organizations now use AI in at least one business function. Around 1 percent consider themselves fully mature. Roughly two-thirds have yet to scale beyond isolated pilots.
Deloitte, July: worker access to AI expanded by half in a single year. Fewer than 60 percent of workers with access actually use it in daily work. 84 percent of organizations have not redesigned jobs or workflows around AI. And only half of executives regularly verify the quality of AI outputs when making decisions.
Eighty-eight to one. That is not a slow adoption curve. It is a wall.
And a third set, which neither paper cites and which changes how the first two read. Statistics Canada, July: 35.9 percent of Canadian workers used generative AI in their main job in the year to March 2026, against 12.2 percent of Canadian businesses using AI at all.
Both papers are worth reading, and I want to be clear before I go further that I think both are right about what they set out to address.
What they get right
Deloitte’s argument is that adoption metrics are easy to produce and easy to fake — a worker opens the tool, clicks enough to register as active, and goes back to working exactly as before. Their proposed answer is adaptation: judgment, experimentation, divergent thinking, measured by behavior rather than by usage counts.
That is correct, and the distinction between did they use it and did they use it to change anything is a real one that most organizations are not drawing.
McKinsey’s argument is that the difference between companies getting value from AI and companies running pilots is not a technology gap. It is an operating model — agents treated as teammates, a knowledge layer that agents can actually reach, architecture designed for interchangeability, autonomy earned progressively, platform centralized and tasks decentralized.
Also correct, and the observation that fifteen companies converged on the same practices independently is genuinely interesting.
Neither paper is wrong. I want to point at something both of them contain and neither follows home.
The finding Deloitte published and did not pursue
Buried in the section on judgment, Deloitte cites research showing that sustained AI use can erode professionals’ confidence in challenging AI recommendations — even when they possess the expertise to do so.
Read that again. Not that people lack the judgment. That people who have the judgment progressively stop applying it.
And a few lines later, on safeguards:
They also report that only half of executives verify AI output quality, and cite research on models responding to fact-checks by deploying persuasive tactics to defend their original answers — which, in Deloitte’s own words, calls into question whether human-in-the-loop safeguards are truly effective.
So in one paper, published by a Big Four firm, we have three things established:
That is a complete description of a control mechanism failing.
One thing I cannot leave out
I would rather not interrupt the assessment here, but leaving this unsaid would be a form of dishonesty.
Those three findings do not describe how I have worked with these systems — across twenty-eight years and four generations of the technologies that preceded them. The capacity to disagree has not eroded. The verification has not become occasional. And I have not been turned into a validator.
That is not one anecdote against a survey. It is a record, published as it happened and never edited, running from a 1998 defense engagement through every wave since.
What it establishes is that these three outcomes are not properties of working with AI. They are properties of working with AI under particular conditions — and Deloitte measured organizations that never established different ones. The finding is accurate about what was sampled. It does not generalize to arrangements built to prevent exactly this.
The distinction determines the remedy. If the erosion is inherent, coaching is all you have. If it is conditional, the conditions are what you build.
And the conditions are not aspirational. The tendency toward agreement is real and it is constant — which is why I measure it. I track the depth at which agreement is reached rather than whether agreement occurred, because fast consensus is the warning sign rather than the result. When that depth falls, it is visible before the output degrades, and it gets reset.
That is the difference between a trait to be coached and a condition to be instrumented. Deloitte’s own data says half of executives verify AI outputs. The question is not how to make them more diligent. It is what would tell them when they had stopped.
Where I part company
Deloitte’s remedy for all of this is behavioral. Coaching agents that analyze cognitive patterns. Nudges. Peer connections. Reflection sessions. Simulations. Experience design that builds judgment over time.
Every one of those is a reasonable thing to do. None of them addresses what the paper just described.
You cannot coach someone into having an independent basis for disagreement.
Judgment is not a personal attribute that gets applied in a vacuum. A person can disagree with a recommendation when they are standing somewhere the recommendation is not — holding information the system does not have, seeing a condition the system cannot see. That is not confidence. That is position.
Take the position away and no amount of training restores the capacity. You get a better-trained validator. Which is exactly what Deloitte observed, and exactly what its own remedy would produce more of.
Redesigned for whom?
The erosion is not a behavioral failure. It is a structural one — and the tell sits in their most-quoted number, though not in the direction it is usually read.
84 percent of organizations have not redesigned jobs or workflows around AI.
Read the phrase carefully. Around AI. Around the technology. The implied measure of progress is how far the organization has been reshaped to fit the tool.
Which raises a question the paper does not ask: redesigned for whom?
Deloitte’s own argument is that the value comes from human judgment — knowing when to challenge the machine, when to verify it, when not to use it at all. And they name the difficulty directly: the more work AI absorbs, the more judgment is needed to direct and evaluate it, and the fewer occasions workers have to develop that judgment in the first place.
Redesigning jobs around AI is how you absorb more work. It is the mechanism that produces the paradox, offered as the remedy for it.
To be fair to the authors, that is not the only available reading. They also write about protecting deep work, preserving space for divergent thinking, and designing the interaction rather than subordinating people to the tool. A generous reading of around AI is “designed thoughtfully for both.”
But even under that reading, the direction of fit runs one way — from the operation toward the technology — and the metric counts movement in that direction as progress.
And there is a further reason not to read that 84 percent as a failure to act.
The workforce has already redesigned.
Statistics Canada, drawing on supplementary questions to the Labour Force Survey, reports that 35.9 percent of Canadian workers used generative AI as part of their main job in the twelve months to March 2026. Over the same period, 12.2 percent of Canadian businesses used AI at all — and 66.7 percent have no plans to adopt it.
Roughly three workers using it for every business that has adopted it. Statistics Canada’s own observation, in an earlier release comparing the two, is that the pattern may indicate that adoption is not solely firm-led.
Microsoft’s 2026 Work Trend Index reaches the same place from a different direction: across 20,000 knowledge workers in ten countries, 16 percent of AI users have redesigned their workflows around AI.
Sixteen and eighty-four. Two independent studies, different populations, the same split.
So the 84 percent are not standing still. Their people redesigned without them — individually, informally, in the absence of any policy to defy. Statistics Canada found that only 5 percent of non-users cited company policy as a reason; the dominant answers were that it did not apply to their job, or that they were not interested.
That is not defiance. It is a vacuum being filled, and the formal operating model now describes an operation that has already moved.
The question is not whether to redesign. It is what you redesign around — and whether you have noticed that some of it happened already.
In an excerpt from a paper I wrote in the fall of 2004 titled Acres of Diamonds, I made the following statement:
That was written about procurement software, two decades before any of this. The architecture is derived from the real-world operating attributes of the people doing the work. Not chosen because it performs better. Derived — you find out how the front line actually operates, and the structure follows from that.
Reverse the direction and you get an operation shaped by whatever the current technology happens to require.
And a note on where the seven truths came from
McKinsey’s operating truths were drawn from fifteen AI-centric companies — a four-person agriculture technology venture, a Series A marketplace, a Series D fintech, a seed-stage AI company, a digital health scale-up.
I want to be careful here, because this is an observation about sampling rather than a criticism of the work.
Those organizations had no legacy operating reality to validate. A four-person venture has no accumulated set of declared processes that quietly stopped matching what people actually do. The map was drawn last week; of course it matches the territory.
Which is why the truths begin at design. There was nothing to trace.
The 88 percent are not in that position. They have decades of process documentation, org charts, workflow diagrams and system configurations describing an operation — and no established way of knowing whether the description still corresponds to what happens.
Advice derived from organizations that cannot have the problem is not wrong. It is simply addressed to a different situation than the one most readers are in.
The missing question
In 1998 I worked on a defense maintenance operation delivering parts on time 51 percent of the time against a 90 percent requirement.
Everyone involved believed they understood the problem. Procurement was being held accountable. Supplier performance was being scrutinized. The process documentation was accurate and current, and everybody had read it.
The question that changed the outcome was not a procurement question, a technology question, or a change management question:
Late afternoon. Almost all of them, around four o’clock — because service technicians were holding order releases and batching them at the end of the day. Releasing parts orders as they arose interrupted service calls, and technicians were measured on call volume.
A rational response to how they were being measured. One department away from the people being blamed. And not represented anywhere in the declared process.
Delivery performance went from 51 percent to 97.3 percent in three months. No new system. The technology was selected afterward, once the result was substantially in hand — which is why no platform can be credited with it.
What that question actually is
It is not a clever question. It is a validation question. It tests whether the declared operating model corresponds to the operating reality.
Both of these papers assume that correspondence.
Deloitte’s adaptation loop adapts people to a system. McKinsey’s seven truths design an operating model around a business. Both are sound engineering. Both begin one layer above the question of whether the thing being adapted to, or designed around, describes what actually happens.
And if it does not, then everything downstream inherits the error — faster, more consistently, and with better instrumentation than before.
That is what the 88 and the 1 have between them. Not a technology gap. Not a behavior gap. The distance between the operation people believe they are running and the one they are actually running.
What I would do with these papers
Read both. They are good.
Then, before adopting either framework, spend a week answering one question about your own operation: where does what we say we do stop matching what we actually do?
Not through a survey. Not through process mapping, which will return the declared version, because that is what it is designed to capture. By going and looking at when things actually happen, who actually decides, and what people are actually measured on.
If the answer is nowhere, you are in a very small minority and both papers will serve you well.
If you find a four-o’clock — and most operations have one — then you have just learned something worth more than any adaptation program, and you learned it before spending the budget.
Their numbers are the case. The question is the part neither paper contains.
-30-
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Sources: “The seven operating truths of AI-native companies,” McKinsey & Company, 11 June 2026. “AI adoption to adaptation,” Deloitte Insights, 9 July 2026. “Use of generative artificial intelligence tools among Canadian workers, March 2026,” Statistics Canada, 30 July 2026. Work Trend Index 2026, Microsoft.
Share this:
Like this:
Related