Enterprise Field Notes · Issue #4

The Productivity Disconnect

Why workers say AI changed nothing while executives project gains. Both are telling the truth. The gap is the story.

By Ben Pickett · May 14, 2026

New here? Subscribe to Enterprise Field Notes, one new issue every week.

By Ben Pickett

I sell AI for a living. So when I tell you that the productivity numbers most enterprises are reporting are quietly disconnected from what their workers are experiencing, take it as a confession, not a critique.

Last week I wrote about The Single Model Trap. The way one AI model produces wrong answers, and the apology loop that hides them.

One of you wrote back with the question that comes after.

“Even when the answer is right, why isn’t it showing up in the productivity numbers?”

This piece is the answer. Two numbers from the same company. Four multipliers. Twenty-five years of history that already happened once. And the six percent of enterprises that closed the gap.

Two numbers

A working paper out of the National Bureau of Economic Research, released in March 2026 and covered by Fortune the following month, surveyed corporate executives across the United States, the United Kingdom, Germany, and Australia. Nearly nine in ten firms reported that AI has had no measurable effect on their employment or productivity over the past three years. The same executives projected average productivity gains of 1.4 percent over the next three years, a number small enough to fall well inside normal measurement noise.

In March, Goldman Sachs analysis covered in Fortune put 2025 global AI spending at roughly $410 billion. Goldman’s chief economist Jan Hatzius told clients the measurable impact on US GDP growth in 2025 was, in his phrasing, basically zero. Goldman’s published analysis attributes the gap partly to a measurement issue (large parts of AI capex flow to imported chips, which counts toward foreign GDP not US GDP) and partly to the fact that the productivity gains visible at the task level have not yet aggregated up to anything macroeconomic.

And the McKinsey State of AI 2025 report, the most widely-cited enterprise AI study of the last year, found that ninety-four percent of companies have deployed AI and report not seeing significant value from it. Six percent are.

Workers do not lie about productivity. CEOs do not lie about productivity.

Workers report that nothing has changed in their day. The economy reports that nothing has changed at the aggregate level. Executives project gains they cannot point to. Four hundred and ten billion dollars has been spent. Six percent of companies are seeing the kind of returns that AI was supposed to produce. Ninety-four percent are not.

The gap is the story. The story is that productivity gains from AI are not a function of AI. They are a function of what enterprises do around AI. Most enterprises have not done it.

A composite scene from inside the gap

I have watched this play out from the inside of more than one enterprise. The shape of it does not change much. The names change. The story does not.

A team buys an AI tool. Leadership is confident. The budget is approved. The vendor is onboarded. There is a launch deck. There is a Slack channel. There is, somewhere, a slide with a productivity target on it.

Six months later, the data comes back. The team’s output looks the same as it did before. Tickets closed per week, lines of code shipped, customer service handle time, whichever metric the team owns. The needle has not moved.

I ask the team how they are using it. Most of them are not, regularly. The ones who are use it for the tasks they would have done anyway, in roughly the same time. The shape of the work is unchanged.

A handful of people have figured something out. They are running two or three concurrent workflows. They are using AI to draft, to summarize, to scaffold the first version of a thing, and then their human time goes to the parts that need judgment. They are doing twice the work. They look like they are bored.

Nobody has institutionalized what those people figured out. Nobody has rewritten the workflow. Nobody owns the gain. The productivity these people are generating is a personal habit, not a company capability. When they leave, it leaves.

Leadership sees the budget line and a couple of wins they can point to in a board deck. The workers see no measurable change in their day. The CEO reports that productivity is improving. The team says, accurately, that nothing has changed for them.

Both are reporting what they see. Neither is lying. The numbers do not agree because they are describing different layers of the same problem.

Two ways to use AI

Before the math, there is a thing nobody on a productivity dashboard is measuring. The workers who say AI changed nothing and the workers who say AI changed everything are doing two different things with the same tool.

Most enterprise AI usage today is a surface layer. The worker keeps the same workflow they had before. AI gets pasted into one step. The draft email. The summary. The first cut of an analysis. It saves a few minutes per task. It does not change the day.

The workers who are getting real productivity lift do something different. They lean in. They use AI to think out loud, to pressure-test their reasoning, to draft something they then take apart and rebuild. They run two or three models in parallel and look at where the answers disagree. They spend more time thinking about a problem before they touch it, not less.

Jensen Huang put it cleanly in March: “Our software engineers 100% use coding agents now. Many of them haven’t generated a line of code in a while, but they’re super productive and super busy.”

That is the pattern, and it is the part most enterprise AI rollouts miss. NVIDIA’s highly-paid engineers at the most demanding AI company on earth are not coding less and doing less. They are coding less and doing more. The work shifted up the stack. They specify. They architect. They review. The AI executes. The hours did not disappear. They moved.

The data is starting to confirm this. The 2025 Stack Overflow Developer Survey reports 84% of developers use AI tools and 51% of professional developers use them daily. CodeRabbit’s December 2025 State of AI vs Human Code Generation report analyzed 470 GitHub pull requests and found AI-coauthored PRs contain about 1.7 times more issues than human-only PRs (10.83 issues per AI PR versus 6.45 per human PR). Logic and correctness issues are 75% more common in AI-authored code; security vulnerabilities are 2.7 times higher; performance inefficiencies show up nearly 8 times more often. The developers writing AI-assisted code report spending more time reviewing it than they used to spend writing the original. The work didn’t get easier. It got different.

This is the counterintuitive truth most enterprise AI dashboards do not capture. The workers who are using AI well are busier, not less busy. They are shipping more, not less. The hours did not go away. They moved.

Think of it like chess. The surface-level player makes the move that comes to mind. The strong player stops, considers the possibilities, evaluates the position, and then acts. The strong player thinks longer per move and wins more games. AI is the same. The lean-in usage requires more thinking up front and more review on the back end. The output is multiples of the surface usage. But you cannot get there by adding AI to a workflow that does not have the thinking time built in. Most enterprises did not budget for the thinking time. They budgeted for the seat license.

Productivity is a multiplication problem

Here is the math nobody is doing out loud.

Productivity gain from AI = access × usage × efficiency × workflow redesign factor.

Four numbers. All of them less than one in most enterprises. Multiplied together, the final number is small in a way that is genuinely counterintuitive if you have not done the arithmetic.

1. Access

The percentage of workers in the company who have sanctioned AI in front of them. Wharton’s 2025 AI Adoption Report finds 82% of senior leaders now use Gen AI weekly, up from 72% a year ago. Rollout to the rest of the workforce lags that, often substantially. Most enterprises still have meaningful populations of workers who either do not have a sanctioned tool, or have one but cannot get an account without going through procurement or IT.

2. Usage

The percentage of workers who have access and actually use the tool in their daily work. Slack’s 2025 Workforce Index found that 60% of desk workers now use AI at work, and 42% use it weekly or more. The compound of access and usage is the number that ends up multiplied through everything else. The workers who are locked out, who ignored the rollout, or who quietly went back to their previous workflow do not contribute to the company-wide productivity gain.

3. Efficiency

The per-user time savings on the tasks that AI actually touches. Where AI is used well, the efficiency gains are real. McKinsey’s State of AI 2025 reports 14% productivity gains in customer service and 26% in software development. Slack’s data shows daily AI users report 64% higher productivity than colleagues who have not adopted AI. The per-user lift, on the tasks AI is involved in, is genuine.

But efficiency only applies to the tasks AI is actually involved in, performed by the workers who are actually using it. The 26% software-engineering number is not the company-wide number. It is the per-task number on the subset of work AI is touching, for the subset of engineers using it well.

4. Workflow redesign factor

The multiplier on all of the above. Did the company change how the work gets done so that the time savings turn into output? Or did it add AI to the existing process and hope the gains would compound on their own?

When workflow is redesigned around AI, this factor is greater than 1. When AI is bolted onto an existing workflow, this factor is roughly 1.0, sometimes less, because the new tool adds friction to a process that was not waiting for it.

McKinsey’s data sharpens this. AI high performers are about three times more likely than other firms to have fundamentally reworked their processes when deploying AI. Fifty-five percent of high performers report having done that redesign work. Among everyone else, the share is under 20%. The high performers did the redesign. The 94% did not.

The compound math

Here is what most enterprises are running.

Access 60%. Usage 60%. Efficiency 20%. Workflow redesign factor 1.0.

0.60 × 0.60 × 0.20 × 1.0 = 7.2%.

Seven percent company-wide productivity gain on a deployment where the per-user efficiency lift is 20%. Compound that down through reporting noise and the lagged effect of any new tool requiring onboarding time, and the number a CFO can actually point to is closer to zero. Which is the number Goldman Sachs saw in the GDP data.

Now look at the same math from the worker’s seat. The workers who are using the tool are seeing the real 20% lift on the tasks they touch. That is the number they describe when they say AI changed their workflow. They are telling the truth. Their CEO is also telling the truth. The two numbers describe the same company.

If the redesign factor moves from 1.0 to 1.5, the same deployment yields 10.8%.

If access moves to 80% and usage to 75%, holding the others, the same deployment yields 9%.

If all four levers move modestly, the compound starts to clear 15%. That is the territory the high performers are in.

Solow’s paradox names the phenomenon. It does not name the timeline.

In 1987, the economist Robert Solow wrote a one-line review of a book on the new information economy. “You can see the computer age everywhere but in the productivity statistics.”

He was describing the same thing we are describing now. Workers reported that computers were changing their work. The macroeconomic numbers showed nothing. Executives projected gains the data did not confirm. The phenomenon repeats. The diagnosis repeats. That part of Solow translates directly into 2026.

The timeline does not.

The 1980s and 1990s history often gets summarized as “the productivity gains from computers took 25 years to show up.” That is technically what happened, but it is the wrong lesson to carry into the AI era. The reason the computer-productivity surge waited until the mid-1990s is that the technology itself rolled out slowly. Personal computers in 1985 were rare and expensive. Office software was primitive. Networking inside companies barely existed. The commercial internet did not exist. The rate-limiting step was getting the machines into people’s hands.

None of that is true of AI in 2026. ChatGPT reached 100 million monthly users in two months, the fastest consumer technology adoption in recorded history. Slack reports 60% of desk workers are already using AI at work. Wharton’s senior-leader weekly AI use rose from 72% to 82% in a single year. The tool is already in the workforce’s hands. The bottleneck is not deployment speed. The bottleneck is what happens after deployment.

Which is the part of the computer history that does translate. Erik Brynjolfsson and his collaborators at MIT spent the late 1990s and early 2000s publishing study after study showing one thing: the productivity gains from computers concentrated in firms that combined the technology with business-process redesign. The firms that bought the machines and kept the workflow the same did not get the gains. The firms that redesigned around the technology did.

The 6% are doing that work today, with 2026 technology, on a deployment timeline measured in months instead of decades. They are not waiting for a 1995-style surge to arrive in 2050. They are capturing real productivity gains right now. McKinsey’s data is unambiguous on this. The 6% high performers are seeing 5% or more EBIT impact from AI this year.

This is the trap of the Solow analogy if you read it the easy way. The phenomenon is the same. The timeline is not. Anyone telling themselves “AI productivity gains will arrive eventually, like they did for computers” is giving themselves permission to wait through a window that has already opened. The 6% are not waiting. They are not catching some future curve. They are doing the redesign work now and getting paid for it now.

What the 6% are doing differently

The McKinsey 2025 report defines AI high performers as companies seeing 5% or more EBIT impact from AI. Six percent of surveyed companies qualify. Five behaviors show up consistently when you look at what those companies have in common.

They redesign workflows

They do not bolt AI onto existing processes. They map the work, identify the redundant or low-value steps that AI can eliminate or compress, and rebuild the workflow around what is left. The team’s job changes. Their daily flow changes. Their meetings change. AI is not an additional task. It is a removed task.

They scale fast

Once they validate that something works on one team, they roll it across teams in weeks, not quarters. The validation-to-rollout gap is the place where most enterprise AI pilots die. Pilots that work and then sit in a steering committee for three quarters end up rebuilt by someone else, or replaced by the next vendor that walked in. The 6% close that gap by making rollout the default and steering review the exception.

They invest in literacy

Wharton’s 2025 report finds that 43% of organizations now see a real risk of worker-skill decline as AI use rises. The 6% close that gap deliberately. They run internal training. They surface their early super-users and turn them into instructors. They treat AI literacy as a competence on par with spreadsheet literacy in 1995. Slack’s data on the productivity differential between daily AI users and non-adopters (64 percentage points) is the size of the literacy gap, not the size of the model gap. The bottleneck is people who know how to use the tool well, not the tool.

They tie outcomes to performance reviews

If no one’s quarterly review reflects an AI-driven productivity outcome, no one will capture that outcome. The 6% put a name next to every measurable AI gain. The gain is reviewed. The lack of gain is also reviewed. Ownership turns AI deployment from a procurement decision into an operating discipline.

They run more than one model

Last week’s piece was about The Single Model Trap. The 6% have read that lesson and built it into the workflow. They do not deploy a single AI on a critical task. They orchestrate. One model drafts. Another verifies. Two providers cross-check on the high-stakes outputs. The disagreement is the signal.

Hallucinations do not vanish because you deployed an AI assistant. They vanish when the output is checked by another model that does not share the same blind spots. The 94% that bought one AI and called the deployment done are also the 94% that keep getting surprised by wrong answers in production. Single-model deployments are how single points of failure get rebranded as productivity tools.

They deploy where work happens, and integrate into the work itself

Not in a separate portal. Not behind another single sign-on prompt. Not in a new tab that workers have to remember to open. AI ships inside the tools the team is already opening at 9 AM. Slack. Teams. WhatsApp. Zoom. SMS. The email client. The CRM. The ticketing system. Wherever the work lives.

But channel placement is only the surface. The deeper move is the integration behind those channels. The AI has to actually be wired into the worker’s ecosystem through their own auth. It needs to know the company. It needs to know the worker, their role, their permissions, their team’s context. It needs to know the company’s HR policies, the security guardrails, the engineering standards the org follows, and what lives in the internal knowledge base. The AI should be able to read the worker’s inbox, summarize the meeting they just left, pull the latest numbers from the spreadsheet their team owns, draft the customer follow-up using last quarter’s call notes, and answer the policy question without forcing the worker to dig through a wiki. All under the worker’s existing permissions. Without copying anything. Without switching tabs. Without exposing data to a model that has no right to see it.

This is the integration layer most enterprise AI deployments are missing. The AI is technically deployed. It is not actually wired in. It cannot act inside the workflow because the auth and integration plumbing is not there. So it stays an island. So the workers stay on the surface. Forty-seven percent of digital workers say they struggle to find the information they need to do their jobs (Gartner, 2023). That is not a problem AI created. It is the problem AI was supposed to solve. Most enterprise AI deployments missed the step where solving it lives.

The other half of the integration is the admin side. The company needs a place where IT controls what is deployed and to whom. What models are approved. What data sources are connected. What guardrails apply by role. What is logged. What is allowed to leave the perimeter. Without that control plane, governance does not exist. Without governance, IT blocks the deployment by default and adoption stalls before it starts. With it, the same deployment moves materially faster from pilot to production because the rollout has a known, repeatable path instead of a per-team negotiation. The admin layer is not a tax on adoption. It is the thing that lets adoption scale.

And then there is the rollout itself. The 6% are deliberate about which data sources connect first, which roles get access in what order, what minimum context each integration needs to deliver a useful answer on day one, and how the rollout is measured week over week. They keep the data requirements minimal at the front edge so a new user gets a real answer on day one, not on day thirty. They surface the wins early so the rest of the org wants in. Frictionless is the design constraint. Not because friction kills a post. Because friction kills a deployment.

I am building toward AI that ships where people already work and is wired into the systems they work with. Slack, Teams, WhatsApp, Zoom, SMS, and the inbox, CRM, calendar, knowledge base, and policy library behind them. It knows the company. It knows the worker. It honors their permissions. It is governed at the admin layer so the company controls what is deployed, to whom, and what data ever leaves. That is what makes adoption frictionless. Anything less is a new tab nobody opens twice.

Why the gap will not close on its own

The most common response to the productivity disconnect is to wait. Surely the gains are coming. Models are getting smarter. The tools are getting better. The integration vendors are catching up. Give it another year.

The Solow paradox argues against waiting. The computer-era productivity surge concentrated in firms that did the organizational work alongside the technology investment. The firms that only deployed watched their competitors pull away. The AI version of that story is unfolding on a much faster clock. The 6% are capturing gains today, not in 2050. The 94% are not waiting for a curve to turn. They are forfeiting gains that are already on the table.

The Wharton data argues against waiting. Forty-eight percent of executives now describe AI adoption to date as a massive disappointment. Disappointment is not waiting. It is the prelude to a budget cut. The companies that cut their AI budgets in 2026 because the productivity did not show up will be the companies that started over in 2028 when the high performers had a two-year head start.

The talent data argues against waiting. Workers who become AI super-users do not stay. They move to companies that let them work the way they have learned to work. The longer an enterprise runs its 60-by-60-by-20-by-1.0 deployment, the more it loses the people who could have raised those numbers from the inside.

What this comes down to

AI works. Worker-level productivity gains are real where they happen.

But enterprise productivity is a multiplication problem. Access. Usage. Efficiency. Workflow redesign.

Most enterprises are running 60% by 60% by 20% by 1.0 and wondering why the numbers don’t move. The answer is in the math. The compound of four fractional multipliers is a small number.

The gap between executive belief and worker reality is not a measurement error. It is the implementation work most companies have not done.

Workers do not lie. CEOs do not lie. The gap is the work that hasn’t been done yet.

Five questions to audit your AI deployment

Run these on the AI deployment you have the most visibility into.

1. What percentage of your sanctioned AI users actually use the tool weekly?

If the answer is under 60%, you have an adoption gap, not an AI problem.

2. What was the workflow before AI? What is the workflow now?

If they are the same, you did not redesign. You added a tool to a process that was not waiting for one. Your workflow redesign factor is 1.0 or less. Your compound math is small.

3. Whose performance review reflects an AI productivity outcome this quarter?

If the answer is nobody, the gain has no owner. It will not be captured. Procurement is not a strategy.

4. If the CFO asked you to point to a measurable productivity gain from AI this quarter, could you?

If you stalled, you are inside the executive side of the disconnect. The fact that you cannot point to the number is the number.

5. Where do your workers actually open AI? Inside their workflow, or in a separate portal?

Adoption survives where work already happens. Everywhere else is friction, and friction is the silent killer of the usage number in the compound.

If three or more of those answers make you flinch, you are running the 60-by-60-by-20-by-1.0 deployment. The gap between what your CEO is reporting and what your workers are experiencing will keep growing until the implementation work gets done. The 6% did the work. The other 94% are still waiting for the productivity to show up. It will not show up. Productivity does not arrive. Productivity is the residue of work that has already been done.

Ben Pickett is Co-Founder and Chief Operating Officer of Swa-AI and former Global Director of Site Reliability Engineering at Nike. He writes about enterprise AI, reliability, and what it actually takes to get technology from pilot to production.

Ben Pickett is Co-Founder and Chief Operating Officer of Swa-AI and former Global Director of Site Reliability Engineering at Nike. Swa delivers enterprise AI inside Slack, Teams, WhatsApp, Zoom, SMS and more channels coming, integrated into the systems behind them through user-scoped auth. Multi-model orchestration built in. Private Data by Default — Retention on your Terms. Flat-rate pricing. Deployment in under three minutes. Swa Solo, our first version for individuals and small teams to try Swa before bringing it to the enterprise, ships soon. swa-ai.com

References

1. Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives. NBER Working Paper 34984, March 2026. Authors: Baslandze, Edwards, Graham, McClure, Meyer, Sparks, Waddell, and Weitz. Survey of corporate executives covering the US, UK, Germany, and Australia. Nearly nine in ten firms report no measurable AI impact on employment or productivity over the past three years; executives project 1.4% productivity gains over the next three years. Median executive reports about 1.5 hours of AI use per week.
2. Thousands of CEOs admit AI had no impact on employment or productivity. Fortune, February 17, 2026 (follow-up coverage April 19, 2026). Direct coverage of NBER 34984. Cites Apollo chief economist Torsten Slok: “AI is everywhere except in the incoming macroeconomic data.”
3. Goldman finds no relationship between AI and productivity but a 30% boost for 2 specific use cases. Fortune, March 3, 2026. Coverage of Goldman Sachs analysis. AI investment grew from $40B in 2019 to $410B in 2025. Chief economist Jan Hatzius: AI investment’s impact on US GDP growth in 2025 was basically zero. Task-level productivity gains where AI is used well: median ~30%.
4. The State of AI in 2025: Agents, innovation, and transformation. McKinsey & Company, 2025. 94% of companies report not seeing significant value from AI deployments. 6% qualify as AI high performers (5%+ EBIT impact from AI). 55% of high performers fundamentally reworked processes when deploying AI, approximately three times the rate of other firms. Customer service productivity gains of 14%, software development gains of 26% where AI is well-used.
5. 2025 AI Adoption Report: Gen AI Fast-Tracks Into the Enterprise (Year Three). Wharton School with GBK Collective, October 2025. 82% of senior leaders use Gen AI weekly (up from 72% in 2024). 46% use it daily (up from 29%). 74% report positive ROI. 43% see real risk of worker-skill decline as AI use rises.
6. Enterprise AI adoption in 2026: Why 79% face challenges despite high investment. Writer, 2026 Enterprise AI Adoption Report. 79% of organizations face AI adoption challenges. 48% of executives call AI adoption to date a massive disappointment, up from 34% the prior year. 75% admit company AI strategy is more for show than internal guidance. Survey of 2,400 knowledge workers and 1,200 C-suite executives.
7. The New AI Advantage: Daily AI Users Feel More Productive, Effective, and Satisfied at Work. Slack Workforce Index, 2025. 60% of desk workers now use AI at work. Daily AI users report 64% higher productivity, 58% better focus, and 81% greater job satisfaction than colleagues who have not adopted AI. 42% of desk workers use AI weekly or more.
8. We’d Better Watch Out. Robert Solow, New York Times Book Review, July 12, 1987, page 36. The original line that became the Solow Paradox: “You can see the computer age everywhere but in the productivity statistics.”
9. The Productivity Paradox of Information Technology. Erik Brynjolfsson, Communications of the ACM, December 1993, Vol. 36 №12, pp. 67–77. Academic foundation for the IT productivity paradox. The 1995–2005 productivity surge later concentrated in firms that combined IT investment with business-process redesign.
10. Bureau of Labor Statistics Productivity and Costs. Multifactor productivity data covering the 1948–1973 (~2.9% annual growth), 1973–1995 (~1.1% annual growth), and 1995–2005 (surge of ~1.5pp above the post-war norm) productivity eras referenced in the Solow paradox comparison.
11. An Interview with Nvidia CEO Jensen Huang About Accelerated Computing. Stratechery, Ben Thompson, March 17, 2026. Source of the “100% use coding agents now… super productive and super busy” quote describing NVIDIA’s engineering workforce.
12. 2025 Stack Overflow Developer Survey: AI. Released December 2025. 84% of developers use or plan to use AI tools. 51% of professional developers use AI daily. 66% cite “AI solutions that are almost right, but not quite” as their biggest frustration; 45% cite debugging AI-generated code as more time-consuming than debugging their own. Trust in AI output declined year over year despite rising adoption.
13. State of AI vs Human Code Generation Report. CodeRabbit, December 17 2025. Analysis of 470 open-source GitHub pull requests (320 AI-coauthored, 150 human-only). AI-coauthored PRs average 10.83 issues each versus 6.45 for human-only PRs, approximately 1.7× more. Logic and correctness issues are 75% more common in AI-authored code; security vulnerabilities up to 2.74× higher; performance inefficiencies nearly 8× more frequent.
14. Gartner Survey Reveals 47% of Digital Workers Struggle to Find the Information Needed to Effectively Perform Their Jobs. Gartner, May 10 2023. Survey of 4,861 full-time digital workers across the US, UK, India, and China. Cited as the gap that permissions-aware enterprise AI search and integrated workplace AI are positioned to close.
15. Enterprise AI Governance: Complete Implementation Guide, 2025. Liminal, 2025. Organizations with structured governance frameworks see materially faster time-to-value and significantly lower security incident rates than ungoverned deployments. Frames AI governance as an enabler of adoption, not a brake on it.

About the author

I'm Ben. I write Enterprise Field Notes, and by day I'm COO at Swa, after years running reliability, data protection, and database operations at Nike. The lesson that keeps proving itself: anything you cannot run without, and cannot walk away from, is a risk you have not priced yet. What is yours?

Read more of Ben's Enterprise Field Notes at benpickett.com.

© 2026 Ben Pickett · Enterprise Field Notes