AI Productivity Gains Do Not Tell You Who Keeps the Time
A faster task creates a distribution decision, not an automatic redundancy or raise. Audit where the saved minutes go.
AI-exposure identifies the task a tool changes; productivity records whether the released time becomes capacity, quality, growth, pay, prices, or fewer workers.
A measured gain can strengthen an employer's bargaining position without delivering a worker benefit, especially where workers cannot negotiate its distribution.
The dashboard is open on a shared screen.
Tickets close faster than they did before the tool arrived.
The break room is quieter.
One person is answering another queue. One vacancy has stayed vacant. Someone who used to train new hires is checking generated replies. The manager calls the graph productivity.
Maybe you write code. Maybe you draft documents. Maybe you talk to customers while a system offers a sentence before you do.
The conditions differ.
The question does not.
Where did the saved minutes go?
Do not let a faster graph answer that for you.
The Speed Claim
Start with the result that is real.
In a staggered deployment at one large business-process software company, a GPT-3-based assistant raised issues resolved per hour by 15% among 5,172 customer-service agents.
Keep the conditions attached.
The system monitored customer chats and suggested responses and relevant guidance.
Agents could edit or ignore the suggestion.
They stayed responsible for the conversation.
This was not a job-loss count.
It was a within-job productivity result, in a stable support workflow, with the data needed to make the assistant useful.
The gain was not equal across the workforce.
Lower-skilled and less-experienced agents increased issues resolved per hour by about 30%.
Higher-skilled and more-experienced agents saw little effect.
Treated agents with 2 months of tenure performed as well as untreated agents with more than 6.
That is not proof that expertise disappeared.
It is evidence that an experienced worker’s accumulated patterns can be handed to a novice as a suggestion at the moment of work.
Study the consequence on both sides.
Your employer may train people faster.
It may also decide it needs fewer people in the first place.
A Faster Assignment
The coding result needs the same discipline.
GitHub randomly split 95 professional developers working on a JavaScript HTTP-server task.
The Copilot group finished in 1 hour 11 minutes.
The group without Copilot took 2 hours 41 minutes.
That is enough to reject the claim that the tools merely make people feel faster.
It is nowhere near enough to say a software department needs half as many people.
The experiment measured one common assignment with a test suite.
Your working day contains requirements that arrive incomplete, design trade-offs, integration, code review, security checks, unfamiliar failures, operations, and responsibility when a production system harms somebody.
Some of those tasks get faster.
Some become the review that the faster task requires.
Some become more valuable because the first draft is cheap.
Call the two situations by their names.
Exposure asks whether a task is likely to change when a tool is available.
Productivity asks whether a defined workflow produces more quality-adjusted output per hour.
Neither one tells you who receives the gain.
OpenAI’s task assessment estimated that roughly 80% of U.S. workers could have at least 10% of tasks affected by GPT access.
That is a potential task-change figure.
It does not report a pay increase, a reduced price, a new headcount, or a redundancy.
Microsoft’s observed-use study found information gathering and writing prominent in how people sought assistance.
That identifies a surface where your work may change.
It does not reveal the payroll destination of the minutes.
The claim on your time begins at the task.
The distribution decision begins after it.
The Missing Minutes
Take out a sheet of paper.
Choose one recurring task that the tool touches.
Do not choose “customer service,” “writing,” or “software development.”
Those are occupations and departments, not measurable work.
Write a task with a beginning and an end.
“Turn an incoming account question into a resolved case.”
“Turn a set of source documents into a draft that a named person approves.”
“Turn a defined bug report into a tested change.”
Record the baseline minutes for one completed task before assistance.
Record the post-tool minutes for the same completed task with assistance.
Then record the minutes a dashboard prefers to lose.
Review.
Correction.
Rework.
Escalation.
Customer recovery.
Write the calculation in full.
Net minutes saved equals baseline minutes, minus tool-assisted minutes, minus review minutes, minus rework minutes.
Do not call the difference between baseline and tool-assisted time a saving until the other minutes are present.
A system that moves 20 minutes of checking from staff to customers has not made the work disappear.
It has moved the work and changed who pays for the error.
Use a second column for volume.
Count completed cases, drafts approved, or changes accepted.
Then record errors, revision time, repeat contacts, escalations, safety where it applies, and the time another team spends cleaning up the result.
Quality-adjusted output means a completed unit with a defect does not count as identical to a completed unit that works.
A call centre moving from 12 to 15 cases per agent-hour may still have an incomplete gain if repeat contacts rise or unresolved work moves elsewhere.
The numerator is not innocent.
Do not let closed mean finished when it means passed to someone else.
Now record the distribution fields.
Staffing.
Open roles.
Backfills.
Overtime.
Pay.
Price.
The gain becomes legible only when these sit beside the time record.
Two Months
Do not classify the gain after a good week.
Record the same task for two monthly observations.
The first month may contain training, a temporary backlog, an unusually easy case mix, or staff working harder because management is watching the pilot.
The second month does not make the result permanent.
It does force the first claim to meet a second operating record.
After the second observation, classify the destination.
Use one primary category.
Write down secondary effects rather than pretending a mixed result is clean.
Capacity. The same staff handle more work, reduce overtime, or shorten a queue. Headcount and pay may stay unchanged.
Quality. Time returns as checking, better resolution, fewer errors, or more reliable service. A workflow can become slower at the visible step and better at the whole service.
Growth. The organization uses lower cost or faster service to serve more demand. Volume rises, and the labour response may be hiring, retention, or a larger workload.
Labour removal. Jobs, paid hours, backfills, hires, or contractor work actually fall while output is held or redesigned.
These are not four moods of the same number.
They are four different uses of released time.
Do not use capacity as a polite word for labour removal.
A vacancy left unfilled is a staffing fact.
A team asked to carry more cases is a capacity fact.
Both can happen at once.
The contact-centre study itself describes channels that do not fit a simple cut story: shorter chats, more chats handled per hour, and a small increase in successful resolution.
It also reports improved customer treatment and lower attrition, though the attrition finding has weaker identification because a worker can leave only once.
That is why the audit records more than minutes.
Minutes are the beginning of the accounting.
They are not its verdict.
The Pay Ledger
Your employer can say the team is more productive and still leave your compensation unchanged.
That sentence is not cynical.
It is a measurement fact.
Productivity expands the capacity to pay, invest, lower prices, or increase operating surplus.
It does not choose among them.
National productivity releases will not settle it either.
They cover a defined sector aggregate, and a nominal pay figure is not comparable with a real productivity figure until prices are handled.
An index describes its own time path.
It is not a wage, a bonus pool, or a labour-share promise.
Ask for the local equivalent, then keep the units honest.
What is the quality-adjusted output per paid hour?
What are average hourly compensation, typical-worker compensation where available, and the price measure used for each?
What share of the gain becomes lower prices, higher quality, investment, retained earnings, or compensation?
The OECD’s producer-side comparison uses real gross value added per hour for labour productivity, and average compensation per hour deflated by the same price index for labour income.
That is not a rule any one employer must follow.
It is the minimum discipline needed before anyone calls a pay change a share of a productivity gain.
The Distribution Is Not Automatic
The longer record is less tidy than the dashboard.
OECD analysis finds that real average hourly labour income failed to keep pace with labour productivity since the mid-1990s in around a quarter of countries with available data.
Average pay can also move with productivity while a typical worker falls behind.
The OECD’s own median-compensation counterfactuals put the shortfall in double digits for the typical worker over two decades.
Those are descriptive comparisons, not a bill owed to you.
They show why “the company got more productive” cannot settle “who got the money.”
Do not let a company compare a nominal wage increase with a real-output claim and call the arithmetic complete.
Ask which price index sits under each number.
Ask whether the comparison is average compensation or median compensation.
Ask whether benefits, hours, and the covered population are the same.
If you cannot negotiate the distribution, a real gain in output per hour can strengthen your employer’s bargaining position without creating a gain for you.
That is the point of the audit.
It replaces the word productivity with a record you can contest.
The Compressed Ladder
Do not miss the complementarity result because it is inconvenient.
The support study suggests an assistant can spread embedded best practice to less-experienced workers.
That can make supervised practice faster and improve the service a new worker provides.
It can also change why an employer hires juniors at all.
The tool may lower the time required to reach ordinary output.
Your employer may preserve the entry route and use the gain for quality or growth.
Or it may reduce junior hiring and rely on a smaller number of people to supervise exceptions.
That is not the same as saying a junior worker was fired.
It is a change in the path into the occupation.
Keep the indicators separate.
Record new-hire counts.
Record separations.
Record tenure mix.
Record training time.
Record who now performs review and who holds authority for exceptions.
The human loop is not a decorative phrase.
It needs an action, the information needed to take it, authority to override, and enough time to do the review.
A clinician asked to approve 60 model-produced notes in an hour has a different control from a clinician who sees only uncertain cases with linked evidence.
Count the review workload before you call your role augmented.
The person carrying the liability may have received a worse job, not a better one.
Minimum Viable Action
Before you sleep tonight, time one recurring task.
Write down its baseline minutes.
Run the task with the tool.
Write down the tool minutes, review minutes, and correction minutes the tool leaves behind.
Do not calculate a staffing conclusion.
Start a record.
Tomorrow, add volume, errors, and the first observable sign of where the saved time went.
What you will be able to say afterwards
- Net time saved equals baseline task time minus tool time, review time, and rework time.
- A task-level productivity gain does not by itself establish a staffing reduction, a pay increase, or a price reduction.
- A credible workflow record includes volume, errors, review, rework, staffing, pay, and price alongside task time.
- Two monthly observations are needed before classifying a gain as capacity, quality, growth, or labour removal.
- Early AI gains in customer service and coding came from bounded settings and are not job-loss counts.
- Less-experienced workers can gain more from AI assistance when the tool transfers embedded best practice.
- Productivity and compensation can move apart because distribution, price measures, and wage inequality matter.
Questions this raises
Does faster work mean people will lose their jobs?
No. Faster completion measures a change in a workflow, not a staffing decision. An employer can use the released time for more volume, shorter queues, quality checks, training, fewer overtime hours, unfilled vacancies, wage changes, lower prices, or fewer positions. Staffing requires evidence about headcount, hiring, hours, backfilling, and demand after the workflow changes.
How should net time saved from an AI tool be calculated?
Net time saved equals baseline task minutes minus tool-assisted task minutes, review minutes, and rework minutes. The record also needs task volume, errors, escalations, staffing, pay, and price. A gross speed claim that omits checking, correction, customer recovery, setup, or transferred work does not show whether the organization actually uses fewer labour hours.
Why should a team wait for two monthly observations before naming an AI gain?
Two monthly observations reduce the chance that a temporary backlog, novelty effect, unusual case mix, or one-off training period is mistaken for a new operating condition. The first observation identifies a possible change. The second tests whether the same pattern persists while volume, errors, review, staffing, pay, and price are recorded. The classification remains descriptive rather than predictive.
What did the customer-service AI study actually find?
A staggered deployment at one Fortune 500 business-process software company gave 5,172 customer-service agents access to a GPT-3-based assistant. Issues resolved per hour rose 15%. Agents could edit or ignore suggested responses and remained responsible for customer conversations. The result measures a defined support workflow, not layoffs, economy-wide employment effects, or a guaranteed gain at another employer.
Why did newer customer-service agents gain more from the AI tool?
In the customer-service study, lower-skilled and less-experienced agents increased issues resolved per hour by about 30%, while higher-skilled and more-experienced agents saw little productivity effect. Treated agents with 2 months of tenure performed as well as untreated agents with more than 6 months. The likely mechanism is that the tool delivers patterns from successful historical conversations at the moment of work.
Does GitHub's 55% result mean every software team will work 55% faster?
No. GitHub's controlled exercise split 95 professional developers working on a JavaScript HTTP-server task. Copilot users completed it in 1 hour 11 minutes, compared with 2 hours 41 minutes without Copilot, a reported 55% faster rate. Production software work also includes requirements, design, integration, review, security, operations, and responsibility for failures, which the exercise did not measure.
Can an AI tool improve output without cutting staff?
Yes. A faster workflow can become capacity, quality, or growth instead of labour removal. Capacity means the existing workforce handles more work or less overtime. Quality means more checking, better resolution, or fewer errors. Growth means the employer serves more demand. Labour removal means jobs, hours, hiring, or backfilling actually fall. The time measure cannot select among those destinations by itself.
What should be measured besides minutes per task?
A complete workflow record includes completed volume, error rate, revision time, repeat contacts, escalations, customer time, safety where relevant, staffing, hours, pay, and price. Quality-adjusted output counts meaningful defects or service failures rather than treating every completed unit as equal. Moving checking work to customers or another team does not make the labour disappear; it changes where the work is recorded.
Why does a productivity gain not automatically produce a raise?
Productivity expands the capacity to pay, invest, lower prices, or increase margins; it does not allocate the gain. Pay comparisons also require consistent price measures, hours, benefits, and population definitions. OECD evidence finds that real average hourly labour income has failed to keep pace with labour productivity since the mid-1990s in around one quarter of countries with available data.
What is the difference between average pay and what a typical worker receives?
Average compensation can rise with productivity while median compensation lags because income growth concentrates toward the top of the wage distribution. In an OECD 24-country analysis, median compensation in 2013 would have been about 13% higher had it tracked productivity since 1995 under the total-economy measure. The comparison is descriptive, not a debt owed to each worker.
Does high AI exposure tell a worker where the saved time will go?
No. Exposure identifies tasks where a language-model tool may change performance under a stated method. It does not determine adoption, task redesign, demand, staffing, pay, or the distribution of output gains. About 80% of U.S. workers were estimated to have at least 10% of tasks affected by GPT access, but that estimate describes potential task change rather than observed use or employment loss.
What should workers watch when an employer says AI has made the team more productive?
Workers should ask for the task, baseline and post-tool minutes, review and rework minutes, volume, error and escalation rates, staffing, vacancies, backfills, overtime, pay, price, and the destination of saved time. A statement that output per hour improved answers only one part of the question. The distribution becomes visible in the records for hiring, hours, quality, customer service, compensation, and margins.