AI / PRODUCTIVITY
ADOPTION → TASK
TASK → PROCESS
PROCESS → CAPACITY
CAPACITY → VALUE
Most organizations measure AI adoption. The harder question is whether AI changes the economics of work.
Licenses, prompts and active users can show diffusion. They do not prove that an institution can produce more, faster or better with the resources it already has.
AI adoption is not productivity. Productivity begins when the economics of work change.
By José Ñáñez
Technology Advisor · Board Member
Published April 7, 2026 · Updated August 24, 2026 · 11 min read
THE MEASUREMENT PROBLEM
Generative AI creates an unusual measurement problem. Usage appears immediately: licenses are activated, prompts are executed and employees begin experimenting. Productivity appears later, and only if the technology changes how work is performed.
A worker may complete a task faster without the organization processing more cases. A team may save hours without reducing backlog. A process may improve without changing cost-to-serve. And a business may release capacity without deciding whether that capacity will absorb growth, improve service, avoid future hiring or reduce cost.
The measurement system therefore has to follow the causal chain from use to work, from work to process capacity, and from capacity to economic outcome.
The metric that matters is not how much AI is used. It is how much more the institution can produce with the resources it has.
WHAT THE EVIDENCE ALREADY SHOWS
Empirical studies already show meaningful productivity gains. They also show why a single benchmark is a weak management tool: gains vary by worker, task and position in the production chain.
01 · HETEROGENEITY
AI-assisted issues resolved per hour
NBER research on roughly 5,000 customer-support agents found a 13.8% average productivity increase and much larger gains among less experienced workers.
NBER — Generative AI at Work →02 · TRANSLATION
Autonomous coding-agent cumulative effect
A 2026 NBER study covering more than 100,000 GitHub developers found that large gains in coding activity attenuated sharply before reaching final releases.
NBER — Writing Code vs. Shipping Code →Time saved can remain local
Across 66 firms and 7,137 knowledge workers, users of a workplace GenAI tool spent about two fewer hours on email per week, yet researchers did not detect broader shifts in task quantity or composition. Individual time savings did not automatically become organizational redesign.
NBER — Shifting Work Patterns with Generative AI →THE PRODUCTIVITY STACK
Each layer answers a different management question. The mistake is stopping at the first layer because it is easiest to instrument.
Most AI measurement programs stop where the interesting management problem begins.
Are people using AI?
Licenses, active users, assisted work
Did the unit of work improve?
Minutes per task, quality, rework
Did the operating system improve?
TAT, throughput, backlog, cases/FTE
Did resource productivity change?
Clients/FTE, revenue/FTE, cost-to-serve
Was released capacity deliberately captured?
Growth, avoided hiring, cost, margin, service
INTERACTIVE ANALYSIS
Move three assumptions. The chart shows a 12-month illustrative rollout and the widening or narrowing gap between theoretical capacity release and organizationally captured productivity.
ASSUMPTIONS
Model mechanics
Potential capacity release
AI-exposed work × task efficiency gainCaptured productivity
potential capacity release × value capture rateUncaptured capacity
potential capacity release − captured productivityThe time curve uses an illustrative rollout profile and a slower organizational capture curve. At month 12 the model converges to the assumptions selected above.
Potential capacity release
15.0%
of baseline work
Captured productivity
6.0%
of baseline work
Uncaptured capacity
9.0%
of baseline work
12-month productivity capture curve
Target exposure rolls out progressively. Organizational capture is assumed to lag technical efficiency.
Interpretation
At the selected assumptions, AI can potentially release 15.0% of baseline work capacity. If the organization captures 40% of that released capacity, observable productivity reaches 6.0%, leaving 9.0 percentage points not yet translated into operating output.
The management challenge is therefore not only increasing AI efficiency, but improving how released capacity is reallocated and absorbed by the operating model.
Where did the uncaptured capacity go?
This is a deterministic scenario model, not a benchmark, forecast or accounting estimate. Workload is normalized to 100 units so the model can isolate the operating logic.
TIME SAVED IS NOT VALUE
The phrase “hours saved” hides several operating questions. Are those hours concentrated enough to change staffing or throughput? Do they occur in a bottleneck? Can demand absorb the capacity? Has management changed workload expectations? Is the capacity being redirected toward growth, quality, service or cost?
This is why theoretical efficiency and captured productivity should be reported separately. The gap between them is not necessarily failure. Some capacity may deliberately be retained for quality, resilience, learning or control. But management should know where it went.
A productivity program becomes credible when it can reconcile the technical gain with the operating outcome.
Time saved is an input. Capacity released is an operating effect. Value is a management decision.
ENTERPRISE PRODUCTIVITY
If AI becomes material to the operating model, its effect should eventually become visible in ordinary business ratios. The exact metric depends on the institution, but the principle is stable: more useful output from the same resource base.
This is a stronger standard than asking whether employees like an AI tool. It also prevents local efficiency claims from being mistaken for enterprise productivity.
True AI productivity should eventually appear in enterprise productivity ratios — not only in AI dashboards.
Can the same team absorb more operational work?
Can the institution support more customers without proportional headcount?
Can complexity and product breadth grow without the same operating burden?
Does capacity translate into economically valuable output?
Does the operating cost of a customer or process structurally change?
Does released capacity remove bottlenecks customers actually feel?
HOW TO MEASURE IT
Case, transaction, document, decision, service interaction or workflow.
Time, quality, rework, throughput, queue and resource cost before AI.
Measure actual AI-assisted execution rather than intended use.
A faster task matters only if it changes the process bottleneck or capacity.
Growth, service, quality, avoided hiring, cost or retained resilience.
Confirm whether process gains become resource productivity and economic outcomes.
PRODUCTIVITY WITH CONTROL
Productivity should be measured net of rework, errors, adverse outcomes, human review and new control costs. A process that moves faster but creates more remediation has not necessarily become more productive.
The same evidence discipline used for AI governance should support productivity measurement: which system changed the work, where a human intervened, what quality threshold applied and what outcome followed.
MEASUREMENT DISCIPLINE
Net productivity = useful output gained − rework − control burden − adverse operational effects
RESEARCH CONTEXT
The studies below support the measurement logic. Their results should not be copied into a business case without observed workflow data and institution-specific baselines.
THESIS
The real question is not whether employees use AI. It is whether the institution can produce more, faster or better with the same — or fewer — resources.
Measure adoption, but do not confuse it with productivity. Measure task gains, but follow them into the process. Measure released capacity, but make explicit how it is captured. That is where AI becomes operating leverage.
José Ñáñez