Your team finishes its weekly reports earlier. Employees say AI helps them get through the work. The finance director asks where the benefit appears in the business results.
All three observations can be reasonable. A task can become faster while the organization spends the same amount, serves the same demand and makes the same decisions. The improvement may be real; the claim about its value may simply be unfinished.
The AI productivity paradox is therefore a useful management question: what has changed between the individual task and the business outcome? To answer it, separate the work completed, the capacity released, the use made of that capacity and the benefit actually realized. Then examine the additional costs and what would probably have happened without the change.
Why the research can seem contradictory
A study of one workflow and a survey of company executives do not necessarily measure the same phenomenon.
In their 2025 paper, Brynjolfsson, Li and Raymond examined a staggered introduction of AI assistance among 5,172 customer-support agents. They found higher average productivity, measured as issues resolved per hour, with substantial variation by worker experience and skill. This is evidence about a particular deployment and an operational outcome. It does not provide a universal financial return for companies purchasing AI assistants.
The 2026 working paper Firm Data on AI, by Yotzov and colleagues, draws on executive surveys from stratified firm samples in the United States, United Kingdom, Germany and Australia. Respondents reported limited past effects on productivity and employment while expecting larger changes ahead. These are reported firm-level assessments, not an experiment isolating AI's causal effect or an audit of every investment's cash return. The working paper had not undergone peer review.
Neither finding cancels the other. Improvements in a selected process can coexist with modest reported effects across an entire business. The studies also concern different populations, periods and measures.
Measurement itself evolves. In February 2026, METR explained why selection effects and time-recording problems made its follow-up developer experiment an unreliable estimate of current productivity effects. Its older slowdown result should not be treated as a timeless verdict on coding tools.
In a May 2026 survey, METR separately asked technical workers about perceived speed and perceived value. The research used a convenience sample and acknowledged substantial uncertainty about respondents' estimates. It offers a reason to define the question carefully, rather than a company-wide ROI benchmark.
For a leadership team, the practical discipline is to ask what each source measured before importing its conclusion into a business case.
Follow one piece of work all the way through
Consider a hypothetical product team preparing an internal update for a monthly leadership meeting. This is an illustrative scenario, not a Learning Coach client case.
An AI assistant helps turn approved project notes into a first draft. The product manager now spends less time composing the document. But the draft still needs technical verification, reconciliation with the release plan and an explanation of unresolved decisions.
Several outcomes are possible.
The team might produce the same useful update with less total effort. It might create a longer update that takes executives more time to read. Or it might use the freed time to investigate a delivery risk and give leadership a better basis for a decision.
Counting drafting time alone cannot distinguish these outcomes. Nor can counting documents produced. The relevant unit might be an accurate, decision-ready update delivered when leadership needs it, including the work required to get there.
This changes the conversation. Instead of asking whether the writing tool was fast, the team asks whether the complete process became more useful, less costly or easier to sustain.
Decide what “productive” means before collecting numbers
A measurement is easier to interpret when its boundary is explicit.
For the product update, record which activities belong to the process: gathering approved information, drafting, checking, revising and obtaining sign-off. Keep active work time separate from elapsed time. A document may require less effort yet sit in a review queue for the same number of days.
Define an acceptable result too. A shorter process that introduces unsupported claims has not delivered an equivalent output. Relevant checks might include factual accuracy, omitted dependencies, clarity about uncertainty and whether decision-makers can identify the action required.
The comparison also needs care. An unusually simple reporting month is a weak baseline for a difficult release period. A particularly enthusiastic early adopter may not represent the wider team.
Where feasible, compare sufficiently similar work with and without the change, or use a planned introduction across comparable groups. Record differences in task complexity, staffing, demand and tools. A simple before-and-after comparison can support learning, but it leaves alternative explanations open.
There is no requirement to make every local evaluation a research trial. There is a requirement to keep the strength of the claim proportionate to the design.
Freed capacity needs a destination
Suppose the end-to-end process really does require less effort at acceptable quality. That is a useful finding. It creates capacity.
What happens next determines the kind of benefit the organization can claim.
Existing unmet demand can be served. The team may complete work that previously waited. The evidence concerns additional useful output, alongside its quality and the demand for it.
Work can improve. People may investigate uncertainties, provide better explanations or strengthen review. The evidence should concern those improvements, rather than assuming they occurred because time became available.
Pressure can ease. A team may reduce overtime or create space for learning. These can be legitimate objectives. Report observed changes in workload or experience directly; a claim about lower absence or better retention would require additional evidence.
Spending can change. The business may avoid an external purchase, reduce paid overtime or make a different resourcing decision. A cash-saving claim needs evidence that expenditure actually changed, with attention to other effects.
Capacity can also remain fragmented or unused. Minutes scattered across separate jobs do not automatically become a block of time that another team can employ. Coordination and work design matter.
The Learning Coach perspective is that teams need to learn how to recognize and use the opportunity they have created. Tool fluency helps people complete a task; shared judgment helps them decide which task deserves the saved attention. This connects to the broader human learning advantage in companies.
Keep benefits, costs and assumptions visible
A salary-based estimate of time released can describe the approximate value of capacity under stated assumptions. It does not establish an equivalent reduction in payroll expenditure.
Be particularly careful about counting the same opportunity twice. If the released time is used to produce additional work, do not also describe all of it as eliminated labor cost. Explain which benefit occurred and how it was valued.
The cost boundary should be equally clear. Depending on the deployment, relevant additions can include subscriptions or usage charges, integration, evaluation, training, review, support and ongoing maintenance. Some are one-off costs; others recur or grow with usage.
Include changes in other people's work. A faster first draft may shift effort to a specialist reviewer. A team-level saving can disappear when the full process is examined.
An ROI estimate also needs a time period, an agreed definition of incremental benefit and a credible comparison with what would otherwise have happened. Finance and operational owners should be able to trace the estimate back to evidence. A reassuring percentage cannot repair a missing benefit definition.
Some valuable outcomes resist immediate monetary valuation. Better documentation, more manageable work or learning opportunities can be reported in their own terms. Keeping them visible is more useful than assigning an unsupported cash figure.
Match the claim to the proof
The following table is an editorial discussion aid, not a validated assessment instrument or accounting standard.
| Proposed claim | Evidence to examine | What remains unproven |
|---|---|---|
| Drafting became faster | Comparable task records and acceptable output quality | Lower total process cost |
| The team released capacity | Net effort across drafting, review and rework | Useful redeployment or cash savings |
| More useful work was completed | Accepted outputs, demand and quality checks | Additional profit attributable to AI |
| Expenditure fell | Actual spending changes and associated effects | Whether AI caused the entire change |
| Decisions improved | Decision criteria and relevant follow-up outcomes | A general company-wide return |
A practical review can end with one of several decisions: continue the current use, change the process, gather better evidence or stop a use that does not justify its costs. An uncertain result is information. It should not be converted into a success story merely because adoption has already been announced.
Make measurement a conversation people can trust
Employees often see work that a dashboard misses: checking an answer, repairing a draft, helping a colleague or deciding that an AI suggestion should be rejected.
Invite that knowledge into the evaluation. Ask people where effort disappeared, where it moved and which new work became possible. Explain how the information will be used. If employees believe every reported minute will immediately become a higher workload target, leaders should consider how that incentive might affect the reports they receive.
For an executive learning event, this makes a useful shared exercise. Give operations, finance and HR the same fictional workflow and ask each group what would count as improvement. Then compare their definitions and the evidence needed to support them.
A keynote can introduce the distinctions; a facilitated session can help teams apply them. Neither establishes a financial return from the event itself. The work afterward must test whether people can make better judgments about their own processes.
For organizers developing that discussion, the guide to designing a corporate AI learning summit explores how to connect an event's central question to workplace application.
The business question worth answering
Faster work can be valuable before it becomes a visible financial return. It can also create output that nobody needs, move effort downstream or leave capacity without a purpose.
The research discussed here is a selected set of studies checked on 28 October 2026, not a systematic review. It does not establish a universal AI effect or a standard payback period. The workflow example and management questions are proposed ways to reason about evidence, rather than demonstrated results.
The useful question for a leadership team is specific: which worthwhile outcome changed, what did achieving it cost, and how confidently can we connect that change to the way we used AI?
Bring that question to your next executive or workplace learning event. Explore Learning Coach's international keynotes and request a tailored proposal.
References
- Brynjolfsson, Li and Raymond (2025), Generative AI at Work. Quarterly Journal of Economics; evidence from a customer-support deployment.
- Yotzov and colleagues (2026), Firm Data on AI. Atlanta Fed Working Paper 2026-3, posted 24 March; executive survey evidence.
- METR (24 February 2026), developer experiment design update. Selection and measurement limitations in a follow-up study.
- Becker, METR (11 May 2026), technical-worker productivity survey. Self-reported speed and value, with sampling and interpretation limits.
