What more closed tickets cannot tell us
A team can close more tickets, ship more features, and draft more reports without improving the result its customers receive. The productivity slide shows more activity. The business outcome may be much flatter.
Faros AI looked at telemetry from over 10,000 developers across 1,255 enterprise engineering teams in 2025. Developers on high-AI-adoption teams completed 21% more tasks per person and merged 98% more pull requests. However, the study also examined DORA’s four DevOps metrics — deployment frequency, lead time, change fail rate, time to recovery — the ones that correlate with customer-visible performance. Any correlation between AI adoption and company-level performance vanished.
This is not a software-only pattern. MIT’s Project NANDA found roughly 95% of generative AI pilots show no measurable profit-and-loss impact. An NBER study of 6,000 executives across the US, UK, Germany, and Australia found the vast majority see little impact from AI on operations; 56% of companies in PwC’s 29th CEO Survey said they have gotten “nothing out of” their AI investments.
Robert Solow’s old line — “you can see the computer age everywhere but in the productivity statistics” — is back in print, almost word for word, four decades later.
The tempting explanation is that organizations have not adopted AI well enough yet. Wait, and the metrics will catch up.
These findings suggest that the dashboards measure increased activity without establishing whether business outcomes improved.
You are counting what AI made free
Activity metrics were invented for a world where execution was the bottleneck. Counting outputs was the cheapest way to make the invisible visible — tickets, lines of code, calls, reports, all proxies for effort when effort was scarce.
AI made effort cheap.
Drafts, summaries, code, financial models, slide decks, customer responses — generated in minutes at near-zero marginal cost. The system produces more work than it can meaningfully absorb. We are still grading it on production volume, like grading a restaurant on how much food it cooks instead of how much gets eaten.
73% of companies track speed as their primary AI metric. Only 12% track quality.
The LSE Business Review put it more politely: current measurement approaches are “mostly concerned with time savings and cost reductions, while saying very little about the quality or novelty of what is produced.” A study of AI coding assistants in 50-plus British public-sector organizations found participants saved an average of 56 minutes per working day. Whether the work that filled those 56 minutes was any good is, quoting the same report, “substantially harder to observe and aggregate.”
Minutes saved per day are easy to report, so they become the figure used in policy and investment decisions. Whether the work produced in those minutes helped anyone is left outside the comparison.
A dashboard can faithfully report the easy metric while leaving the outcome out. Rising activity then looks like progress even when the customer’s result has not improved.
Productivity theater is a self-reinforcing loop
Joe Procopio at Inc. named it well: AI productivity theater. Companies adopt AI out of FOMO, employees fire up a chatbot and look more productive, dashboards report gains, dashboards justify expanded deployment, deployment produces more dashboard gains.
AI usage became a proxy for success without a separate check on business outcomes.
The pattern shows up across enterprise rollouts documented in operations literature. AI deployed across marketing, sales, legal. Internal dashboards report a 6% productivity gain. Leadership takes the number to the board. Deployment expands.
Then someone runs the outcomes. Win rate: flat. Cycle time to qualified pipeline: unchanged, in some segments worse. Error rates and rework: up. Hidden costs: overproduction of low-value materials, downstream bottlenecks where processes cannot absorb the new volume, longer legal and sales-ops cycles driven by review burden the dashboard does not see. The 6% gain is real as a measurement. It is theater as a business signal.
The loop is self-reinforcing because the dashboard never gets the bad news. Metrics measure activity, activity rises, dashboard reports gains, deployment expands, activity rises again. Actual outcomes go untouched or quietly degrade. The activity figures can be accurate while giving little information about business value.
The part that closes the loop: 70% of CEOs, CFOs, and senior executives use AI less than one hour per week. 28% never use it at all. They are making organization-wide deployment decisions on the basis of dashboards they do not use the tools to interrogate.
The same NBER cut found that CFOs perceived AI-attributed productivity gains of 1.8% in 2025. Measured against actual revenue and employment data, the number was 0.6%. The dashboard saw three times the gain reality did.
What your dashboard cannot see is what is eating your margin
Three additional measures help explain the gap: verification effort, decision quality and alignment between teams.
The first is verification effort. GitClear analyzed 211 million changed lines of code from Google, Microsoft, Meta, and enterprise repositories across five years. Code churn — code discarded within two weeks of being written — rose from 5.5% in 2020 to 7.9% in 2024. Duplication in blocks of five or more lines rose eightfold. Copy-pasted lines now exceed moved lines. Velocity dashboards looked spectacular. Quality dashboards did not exist.
A Workday survey found around 40% of AI time savings is offset by extra work fixing AI-generated content. For every 10 hours gained, four go to cleanup. Review and correction costs need to be included in the productivity calculation.
A case reported by CFO.com illustrates this bottleneck. An AI tool sped up document generation by 240 times. End-to-end throughput hit 89% of target. The bottleneck moved from drafting to validation, exactly where the dashboard was not looking.
The second thing your dashboard cannot see is decision quality. Not the count of decisions made — the reversal rate inside a defined window, the escalation frequency, the outcome attribution over time. A decision’s quality becomes clearer when you can see whether it held up in practice or was overturned within 90 days. Activity metrics report before that verdict; decision quality requires waiting for it.
The third is alignment between teams. As work accelerates, misalignment compounds. Two teams moving fast in different directions create more friction than one team moving slowly in the right direction. In slow systems, misalignment shows up early. In fast systems, divergence expands several layers before it becomes visible — and the rework has already happened. Speed amplifies the cost of pointing the wrong way.
The Faros results also show downstream costs. The same Faros AI study found code review time up 91% on high-AI teams, PR size up 154%, bugs per developer up 9%. None of which shows up on the activity chart.
The dashboard is answering the question it was built to answer. When execution is no longer the constraint, the people making investment decisions need it to answer a different one, showing review costs and outcomes too.
But can you measure decision quality at all?
The case for activity metrics is not that leaders are naive. They are cheap, consistent, available at scale. Decision-quality metrics need longer time horizons, counterfactuals, subjective attribution. The LSE concedes that quality and novelty are “substantially harder to observe and aggregate than time or cost savings.” Under complexity and time pressure, imperfect-but-available metrics beat theoretically-superior-but-unavailable ones. Tell a leader to stop measuring what they have without giving them something usable, and you are asking them to manage blind.
A second objection. Swap volume for “decision quality” and you trade an objective proxy for a subjective one. Who decides what counts as a good decision? Over what horizon? If outcomes depend on factors outside the decider’s control, measurement penalizes bad luck and rewards good luck. Alignment metrics risk becoming conformity pressure — teams that legitimately disagree with strategy get flagged instead of heard. Gaming a subjective metric is harder to detect than gaming a ticket count.
These are real measurement problems, but activity metrics have their own weaknesses.
Activity metrics that AI can game without human intervention are already the worst form of Goodhart’s Law. Agents inflate commit velocity, ticket throughput, and feature counts on their own. The system you are trying to measure is now optimizing the measurement. Rising counts alone would not demonstrate improved outcomes.
And the implementation cost argument cuts both ways. The cost of measuring the wrong thing is not imprecision — it is systematic misallocation. 6% productivity gains while win rate flatlines. 1.8% perceived growth that turns into 0.6% measured growth. Those gaps can lead to misplaced investment.
The third objection is the historical one. Solow’s paradox resolved itself once organizations learned to integrate IT. Wait, and the AI productivity paradox will resolve the same way.
It misses a structural difference. IT automated execution inside existing workflows, so output per hour and cost per unit stayed directionally valid through the adoption lag. AI does not just accelerate execution — it makes output abundant while shifting the constraint to verification, judgment, and alignment. The meta-analysis from California Management Review reinforces this: AI’s productivity dividend is real “in specific contexts, for specific users, and under specific workflow designs,” not as an aggregate. Aggregate output stopped being where the value is created.
Five moves to push your metrics upstream
Start with measures that help the people doing and reviewing the work explain what has changed. These five moves can begin this quarter.
Audit what you already track. Tag every KPI as activity, output, or outcome. Most organizations find 80% or more sits in the activity layer — counts of work done, not work that mattered. The ratio is the diagnosis. Assign the audit to a named owner, not a committee, or it will never finish.
Instrument decision quality where decisions are expensive. Start with three high-stakes categories — capital allocation, hiring, product prioritization. Track reversal rate inside 90 days, escalation frequency, outcome attribution. Do not measure every decision; that creates governance overhead that mirrors the productivity theater it is supposed to replace.
Make verification visible. Implement verification-to-generation ratios on AI-augmented workflows. Track review cycles, rejection rates, time spent validating. Pilot on two teams — one heavy adopter, one light — so the overhead pattern shows up. Without verification time, the dashboard gives an incomplete account of throughput.
Track alignment coherence. Three lightweight indicators. Priority consistency: do team top-3 priorities map to organizational top-3? Cross-team dependency friction: count blocked work, conflicting requirements, rework cycles caused by misalignment. Strategic traceability: pick a random sample of in-progress work and trace each item to a stated objective in two steps. Untraceable items are alignment drift, regardless of how fast they move.
Write outcome contracts before work begins. State the intended outcome, the measurement method, and the review timeline upfront. After delivery, measure whether the outcome materialized. Outcome realization rate — the share of completed work that achieved its stated purpose — becomes your primary health metric. Without it, outcome measurement becomes another process artifact that generates reports and changes no decisions.
One note on what to keep. Activity metrics still earn their place in regulated environments, internal benchmarking, and small teams where alignment is maintained through direct conversation. Keep those measures where they are useful, and assess them alongside outcomes and review costs.
Only 44% of organizations currently use financial guardrails or AI FinOps practices for measuring AI value. The other 56% run dashboards not designed for the question they are being asked.
The dashboard is not lying. It is answering the wrong question.
When AI removed the execution constraint, your dashboard did not become useless. It became specific. It answers a question you already solved: how much work can the system produce?
The new question — the one no dashboard was designed to ask — is whether the right work is being done, whether teams are pulling in the same direction, and whether the cost of verifying AI output is showing up anywhere visible.
You do not need a metrics transformation program. You need three small audits. Tag this quarter’s KPIs by layer. Pick one decision category and track reversal rate. Instrument verification-to-generation ratio on one workflow. Compare the results with the gains reported by the existing dashboard.
Use those results to decide which workflows to expand or redesign. Include the time of the colleague who checks the output and the outcome for the person who needs it. Both belong beside the count of work produced.
References
- E001 — NBER / PwC 29th CEO Survey (2026). “Thousands of CEOs Just Admitted AI Had No Impact on Employment or Productivity.” Fortune.
- E002 — LSE Business Review (2026). “AI Productivity Gains Should Be Measured in More Than Minutes Saved.”
- E003 — Artur Markus (2025). “The AI Productivity Measurement Crisis.” AI Unfiltered.
- E004 — California Management Review, UC Berkeley (2025). “Seven Myths about AI and Productivity.”
- E005 — Joe Procopio (2025). “AI Productivity Is Finally Real — But AI Productivity Theater Is About to Ruin It.” Inc.
- E006 — InfoWorld (2025). “AI Agents and Bad Productivity Metrics.”
- E007 — GitClear (2025). “AI Copilot Code Quality 2025 Research”; Lars Janssen on verification debt; Workday rework data via CIO.
- E008 — CFO.com (2026). “Your AI Business Case Is Built on the Wrong Math.” (240x doc / 89% throughput case.)
- E009 — Fortune / NBER (2026). Executive AI usage and CFO perception gap.
- E010 — Faros AI (2025). “AI Software Engineering Productivity Paradox.” (1,255-team study.)
- E011 — Gartner (2026). “5 AI Metrics That Actually Prove ROI to Your Board.”
- E012 — Faros AI (2025). 1,255-team study; InfoWorld commentary.