The '19%' problem: what Gartner's productivity surveys actually measured this spring

Gartner published two surveys in May 2026 that should be uncomfortable reading for every chief executive with an AI line item in their 2027 budget.

The first, released 13 May, found that 19% of employees surveyed in Q1 2026 reported no time saved with AI

—a result drawn from

12,004 employees and managers across 40 countries

.

The second, released 19 May, found that while AI saves sellers an average of 4.8 hours per week, sales organisations that reinvest that time into high-impact activities are 2.2× more likely to exceed customer growth goals

. Read separately, each looks like a vendor cheerleading deck. Read together, they describe a measurement crisis that most enterprises have not yet named: we can count time saved, we cannot yet explain why the profit does not arrive.

I have defended enough pilot budgets to know when a metric is hiding a governance failure. The '19% reporting zero time saved' figure is not an adoption problem. It is a confession that the productivity question we are asking—'Did you save time?'—is the wrong level of abstraction for enterprise decision-making. Time saved is a task metric. It becomes a business metric only when the organisation can redeploy that time, measure the output quality of the redeployed capacity, and attribute margin movement to the shift. Most boards I work with have none of those three controls in place. They have seat counts, prompt volumes, and vibes.

The metrics executives cite vs. what the surveys actually captured

BCG's research found 60% of companies globally were not generating material value from AI despite substantial investment, which aligns with McKinsey's finding of only 39% of organisations reaping EBIT impact

.

A landmark NBER survey of nearly 6,000 senior executives across the United States, United Kingdom, Germany, and Australia found 69% of businesses actively use AI, yet 89–90% of those firms report no detectable impact on employment or productivity over the prior three years

. That gap—between 'we use it' and 'it moved the number'—is where $644 billion in 2025 GenAI spending went to find an addressable use case.

But the controlled task studies tell the opposite story.

Workers completed tasks 25.1% faster with 40%+ higher quality ratings in Harvard Business School studies, programmers produced 126% more coding output per week according to Nielsen Norman Group

, and

developers completed tasks 55% faster using GitHub Copilot in research involving 4,800 developers

. The divergence is not subtle: individual contributors report extraordinary gains, finance reports flat productivity.

The honest explanation is that we are measuring different things and calling them by the same name.

Microsoft provides productivity metrics through Copilot usage reports—usage frequency, features accessed, user activity data—but these metrics measure adoption rather than value; the fact that a user interacts with Copilot regularly does not demonstrate that those interactions are producing business outcomes that justify the investment

.

Microsoft updated its M365 Copilot usage reporting in mid-July 2026 to

refresh faster (48 hours), change the default reporting window to 28 days, separate agent metrics into a dedicated report, and roll out worldwide by late July 2026

. The change is administrative—faster dashboards, cleaner attribution—but it highlights the underlying problem:

GitHub updated Copilot usage metrics on 2 July 2026 to improve reporting accuracy by adding CLI suggested-line telemetry and correcting AI credit attribution that previously showed real usage as 0.0; the change is an admission that AI developer tooling has outgrown the neat accounting models built for seat-based software, and Copilot now lives in editors, terminals, cloud-side workflows, pull requests, and agentic features

. Usage is fragmenting faster than the instrumentation can follow it.

The 'time saved but not redeployed' trap

Sales organisations that achieve moderate to large AI time savings and then reinvest that time into high-impact sales activities are 2.2× more likely to exceed customer growth goals and 3.1× more likely to exceed lead-to-opportunity conversion goals

, according to Gartner's survey of 210 CSOs conducted in January and February 2026. That 'and then reinvest' clause is doing the analytical work.

Gartner's Dan Gottlieb stated: 'AI is not the hero of this story; AI is the accelerant. The opportunity is not simply using AI to improve sales productivity. It is using AI to break through the constraints that limit sales output'

.

The constraint is not technological.

Without governance—permission remediation, sensitivity labels, adoption programmes—most deployments see zero measurable ROI, and the licence spend is absorbed as an IT expense with no business outcome; EPC Group observes 60–80% of Copilot deployments fall into this ungoverned category

. That range—60 to 80 per cent—is consistent with every large-scale deployment I have reviewed in the past year. The tool works. The workflow redesign required to convert task-speed into margin does not happen.

The Work AI Index 2026 surveyed 6,000 workers and found 87% using AI, 11 hours saved per week, and only 13% of organisations getting real business value; the gap is where the value is being absorbed

.

'Botsitting'—re-pasting documents into prompts, supervising output, debugging confident-but-wrong answers—consumes an average of 6.4 hours per week, more time than workers spend producing work

.

Stanford and BetterUp researchers coined the term 'workslop': AI-generated content that is unhelpful, low-effort, or low-quality; 40% of workers received workslop in the past month, spending nearly 2 hours per incident deciphering or correcting it, costing $186 per employee per month in lost productivity

.

If I were presenting to a board that had spent eight figures on Copilot seats in 2026, I would show them this: time saved per seat is real, but unmanaged output quality destroys the margin before finance can book it. The model is not the risk. The lack of a quality gate between 'the AI said it' and 'we shipped it' is the risk.

What the measurement gap tells you about maturity, not tools

A December 2025 Gartner survey of 197 CxOs and senior business leaders revealed only 27% of executives have a comprehensive AI strategy, and just 20% believe their workforce is truly AI-ready

.

Swagatam Basu, Senior Director Analyst at Gartner, said: 'The survey revealed that in the shift to an AI-powered workforce, most leaders are mistaking basic access or adoption metrics for transformation. This "enablement illusion" is hiding risks and draining ROI'

.

The illusion is not just enablement. It is the belief that time saved is productivity gained.

AI can improve productivity before it transforms workflows, but not everyone observes these distinctions; only about a tenth of workers agree that AI is transformative for their organisations while three times that amount of C-level executives believe they have achieved sustained, enterprise-wide impact from AI; at least some of these leaders are confusing individual efficiency gains with enterprise reinvention

.

Gallup's Q1 2026 data shows 65% of employees say AI has improved their productivity, but only 12% say it has transformed work

—a 5.4× perception gap that is larger than any measurement error.

Employees who are proficient with AI across multiple use cases are twice as likely to be highly productive, 2.3 times more likely to deliver high-quality work, and 3.2 times more likely to drive effective process improvements; leaders should move beyond tracking basic adoption and implement a 'True ROI Index' focused on the depth and diversity of AI use

.

That recommendation—depth and diversity, not seat count—is the first honest metric I have seen in a Gartner deck this cycle. Measuring 'how many people prompted' is as useful as measuring 'how many people opened Excel'. The question is whether the prompt improved a decision, shortened a cycle, or eliminated a manual reconciliation that was costing you margin. Most organisations cannot answer that question because they never instrumented the workflow to measure it.

The developer productivity wedge: faster and slower

The cleanest controlled experiment came from the coding domain, and the result should make every CTO revisit their Copilot ROI model.

A randomised controlled trial by METR found experienced developers required 19% more time to complete coding tasks when using AI tools compared to working without them, yet reported a 20% perceived speed increase, creating a perception-reality gap of roughly 39–44%; developers predicted a 24% speed increase before beginning the tasks

.

Read that twice. Developers believed they were 20 per cent faster. The stopwatch said they were 19 per cent slower. The gap is not rounding error; it is a cognitive mismatch between effort and outcome that every performance dashboard is currently blind to.

GitHub Copilot's code acceptance rate averages between 27% and 30%, and developers retain 88% of accepted code in final submissions, indicating that Copilot suggestions are production-ready rather than starting points requiring extensive modification

—but acceptance rate and cycle time are measuring different things. If you accept 30 per cent of suggestions and each suggestion costs you review overhead, you can be accepting more code and shipping more slowly at the same time.

GitHub Copilot now generates 46% of code written by developers and Microsoft CEO Satya Nadella confirmed that Copilot now represents a larger business than GitHub itself at the time of the 2018 acquisition

. That revenue milestone is important—it tells you the product has commercial traction—but it does not tell you whether the buyers are measuring cycle time, defect rates, or just seat utilisation. My bet: most are measuring seats.

What to measure if you want the truth

If I were writing the measurement brief for a client deploying Copilot or any LLM-based assistant at scale in 2026, I would bin the vendor dashboards and track these six instead:

1. Cycle time by work type, not average time saved. Measure 'days from brief to shipped feature' for engineering, 'hours from request to closed ticket' for support, 'days from RFP to submitted proposal' for sales. If AI is working, cycle time should compress. If it is flat or rising, you are botsitting.

2. Rework rate. What percentage of AI-drafted content required material human correction before it could ship? If that number is above 30 per cent, your quality gate is a human doing the work twice—once to prompt, once to fix.

3. Headcount redeployment. Did you reassign capacity freed by AI to higher-margin work, or did people fill the time with more of the same?

A financial services executive told Fortune: 'A product development cycle tracking to take 24 to 36 months was completed in six months once his team incorporated AI capabilities; rather than reduce staff, he redeployed those developers to additional projects'

. That is the reinvestment Gartner is measuring. Most firms I work with are not doing it.

4. Agent-mediated margin contribution. For any autonomous or semi-autonomous agent, measure: revenue per agent-hour, cost per agent-hour, and net margin per agent-hour. Compare to the human baseline. If you cannot do this calculation, you do not know if the agent is profitable.

5. Shadow AI surface area.

Only 21% of workers use AI daily despite 91% of organisations claiming to use it; 56% of workers received no AI training; and 73.8% of workplace ChatGPT accounts are personal unsecured accounts

. That last figure is the governance time bomb. Measure it, then eliminate it.

6. Executive–worker perception delta.

Only about a tenth of workers agree AI is transformative for their organisations, while three times that amount of C-level executives believe they have achieved sustained, enterprise-wide impact

. If your board believes AI is transformative and your IC population does not, someone is looking at the wrong dashboard.

None of these are in the default Copilot admin centre report, and that is the point. The vendors will give you adoption, engagement, and prompt volume. You have to build the instrumentation that tells you whether any of it mattered.

The bet I would make

Gartner forecasts 40% of enterprise applications will embed task-specific agents by end-2026, up from under 5% in 2025, while over 40% of agentic AI projects are forecast to be cancelled by 2027

. Both predictions can be true. Embedding is cheap—wrap an LLM call in a workflow trigger and call it an agent. Graduating from pilot to production requires measurement discipline that most enterprises do not have, so the cancellation rate will be high.

If I were allocating capital in this cycle, I would bet against vendors selling 'time saved' as a proxy for ROI, and I would bet on the 13 per cent of organisations that have built the workflow redesign and quality controls required to convert task-speed into margin. That 13 per cent—

the organisations getting real business value from AI despite 87% of workers using it and 11 hours saved per week

—are not running better models. They are running better systems around the models. They know what good output looks like, they have a process to reject bad output before it ships, and they have redeployed the saved capacity to work that moves the EBIT line.

The '19% reporting no time saved' problem is not an AI problem. It is a governance problem that shows up as a measurement problem that gets reported as an AI problem. Fix the governance, instrument the workflow, measure the cycle time and the margin—not the prompt count—and the productivity will follow. Or keep counting seats and wondering why the number your CEO cites in the all-hands does not match the number your CFO sees on the P&L.

I know which approach I would defend in front of a board. And I know which one most enterprises are still running.

---

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.