How to Discuss Pay or Workload When AI Speeds Up One Task

The answer: start with the workflow, then match the ask

If AI makes one part of your work faster, use that change to open a measured conversation, not to assume your whole job is lighter or a raise follows automatically. Compare ordinary cycles from work arriving to an accepted result. Count tool use, checking and correction, volume, quality, and added responsibility. Then match your request to what changed: ask for priorities or a workload adjustment if extra volume or review absorbs the time; ask for a compensation review if your scope or accountability has durably grown, or a review is due. Raise both when both conditions hold. A fast subtask is a local observation. The MIT working paper “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence” found faster, higher-quality results on bounded writing assignments in a randomized online experiment; it did not measure whole jobs or pay. The Boston Fed’s “Workers’ Perspectives on Artificial Intelligence” reports stated intentions to seek raises associated with perceived productivity gains, not raises received. Keep two records: an operating ledger for time, accepted output, checking, and workload; and a responsibility ledger for scope, judgment, and ownership. Speed is a reason to investigate, not a promised bargaining result.

Sources: Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence; Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears

What does a faster task actually establish?

A faster task establishes a limited fact: under conditions, a defined piece of work took less time. It does not show by itself that the whole role now requires fewer hours, that the result will remain acceptable in production, that the organization captured lasting value, or that compensation should change. Those are separate questions. For a pay or workload conversation, start with which task changed, what counted as finished, and what work surrounded it. The 2023 working paper Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence reports a preregistered online randomized experiment with 444 college-educated professionals. Participants completed incentivized, occupation-specific writing assignments in fields including marketing, grant writing, consulting, data analysis, human resources, and management. Tasks included press releases, reports, analysis plans, and emails. Half received ChatGPT access for the second assignment. The abstract reports that time fell by 0.8 standard deviations and output quality rose by 0.4 standard deviations. In the results, treated participants spent about 10 fewer minutes on the second task than the control group, which averaged 27 minutes; evaluator grades rose by about 0.45 standard deviations. The finding supports a bounded claim: in this setting, access improved performance on short writing tasks by reducing time and raising judged quality. The boundary matters. The study tested selected writing assignments under incentives, not every stage of a live job, a full workweek, or a firm’s adoption decision. A task that looks like drafting may also require gathering accurate inputs, protecting confidential information, checking details, getting approval, tailoring the result, and handling follow-up. The experiment assessed submitted work, but its reported time result does not establish the effort or responsibility required for all those surrounding steps at a particular workplace. Nor does an experimental improvement tell us whether a manager will add volume, preserve saved time, change staffing, or distribute value through pay. It reports what happened in its design, not an individual worker’s bargaining outcome. The paper also makes “same task, just faster” an incomplete description. Participants shifted effort away from rough drafting and toward idea generation and editing: the share of time spent rough-drafting fell by more than half, while editing’s share more than doubled. That does not mean every worker now edits more, or that editing takes longer in total. It shows that a task’s internal mix can change when one step becomes easier. In a workplace, a faster first draft could leave room for checking, tailoring, or additional assignments; alternatively, the whole cycle could shorten. The actual workflow determines which occurred. Before using speed in a conversation, define a comparable unit, such as a recurring report measured from request to approval and delivery. Note the starting conditions, what the AI-assisted step replaced, the accepted result and quality standard, and the review or correction still required. Compare several ordinary cycles where feasible; one unusually easy assignment can distort the picture. Separate first-pass time from effort through acceptance, and record whether volume, deadlines, rework, or responsibility changed. Use only records permitted by workplace policy. The aim is a precise statement: this step became faster, while these other parts stayed the same or changed. That can open a practical discussion without claiming speed alone settles workload, value, or pay.

Sources: Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence; Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence

Where can the saved time go?

A faster step can produce several outcomes. The minutes might reduce total effort; they might be spent checking or correcting the result; they might support more accepted work; or they might be absorbed by a larger queue without changing work hours. The clock alone cannot tell you which occurred. For a workload conversation, ask where capacity went across the whole cycle, not only how fast the first pass became. “Shifting Work Patterns with Generative AI” reports a field experiment across 66 firms and 7,137 knowledge workers, randomly selected for access to a generative AI tool integrated into workplace applications. In the second half of the six-month experiment, the 80% of treated workers who used it spent two fewer hours per week on email and reduced work outside regular hours. Researchers did not detect shifts in task quantity or composition from individual tool provision. Under those conditions, measured time savings accompanied less out-of-hours work without a detected task-mix or quantity change. This does not show that every user got the same result, that all saved minutes became leisure, or that other employers will reduce workload. The reported outcomes were time use and task mix, not universal workload relief or compensation. The OECD’s “Generative AI and the SME Workforce” examines a different unit: employer reports from small and medium-sized enterprises across seven countries. In its 2024 survey, SMEs said generative AI was used more for simple, one-off, trivial tasks than for complex, recurring, important ones, though many used it for a combination. Among SMEs using it, 65% reported improved employee performance; fewer reported scaling up, competing with larger companies, or increasing revenue. These are perceived organizational benefits, not workers’ time logs or proof that a particular employee produced more. The pattern cautions against treating one easy use case as a description of a whole role. Individual access in a field experiment, business-level self-reports, and a worker’s own experience answer different questions; they cannot be collapsed into one claim about saved time. A practical capacity-flow map can make the distinction concrete. Start with time removed from the assisted step. Subtract time added for checking, corrections, approvals, and follow-up. Then ask whether the queue changed, another task was displaced, accepted output or quality changed, and any time was actually released. This map is a reasoning aid, not a validated instrument. It distinguishes an apparent saving from a net change in the working day. If a first draft takes less time but each draft needs careful checking, effort may have shifted. If the queue grows, the same saving may become more throughput or tighter deadlines. Observe the workflow before assuming either result. Bring that account to a manager: identify what became faster, what checking or new work followed, and what happened to the queue and accepted results. If capacity is consistently released, ask whether some can remain protected for focused work or recovery. If it is absorbed, ask for an output target that includes quality checks, and clarify which work should move when demand exceeds it. The evidence cannot decide the employer’s answer; it can make your request specific. Describe the observed flow, name the outcome you seek, and agree how to review it.

Sources: Shifting Work Patterns with Generative AI; Generative AI and the SME Workforce

What should count in a fair before-and-after comparison?

A fair comparison begins with a matched unit of work: the same kind of assignment, similar complexity, quality bar, and endpoint, such as a result accepted by a reviewer. Compare ordinary cycles before and after AI assistance, rather than selecting the quickest success. Measure from receipt to acceptance, not only the minutes spent producing a first draft or running a prompt. Include preparation, inputs, prompting, checking, corrections, approvals, and follow-up. If the assignment changes materially, record that difference instead of treating the cycles as equivalent. Speed and task composition can move together. In “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” a preregistered online experiment assigned occupation-specific, incentivized writing tasks to 444 college-educated professionals. For these bounded tasks, ChatGPT access reduced time by 0.8 standard deviations and raised evaluator-rated quality by 0.4 standard deviations; the paper also reports a shift away from rough drafting toward idea generation and editing. This establishes results in the experiment, not a general estimate for a whole job. It suggests a practical question: which stages shrank, which expanded, and did the accepted result remain comparable? A faster first pass is important, but it is one line in the account. An operating ledger can make the workload question concrete. For each representative cycle, note: full-workflow time; outputs accepted at the agreed quality bar; review, correction, or rework effort; queue, deadlines, or handoffs that changed; and what happened to apparently freed time. Was it released, used for more accepted work, absorbed by new requests, or spent verifying outputs? Keep the measures simple enough to update consistently. Show the workflow direction without claiming precise attribution for every minute. Add a separate responsibility ledger so the efficiency comparison does not hide role changes. Record who chooses the method, checks results, resolves exceptions, owns the final decision, handles escalations, and is accountable when something is wrong. Note whether these duties are occasional or regular. More output with unchanged responsibility differs from becoming the person who sets a process, approves its use, or owns quality. These observations can inform a role review, but do not establish a particular salary value or entitlement. This two-ledger approach is a practical synthesis, not a validated instrument. Operating changes can inform priorities or capacity; durable changes in ownership and accountability can inform a scope discussion. “Generative AI and the SME Workforce” reports that its 2024 survey analysis found SMEs using generative AI used it more for simple, one-off, trivial tasks than complex, recurring, important ones. This employer-reported SME evidence does not measure your role or predict your manager’s response. It is a reason to describe the actual task bundle rather than generalize from one easy use. Use a few normal cycles across a representative period, allowing for variation in difficulty, volume, and seasonality where feasible. Agree on the window and quality standard before a short trial, then review the measures together. Keep aggregate or anonymized records when sufficient, and follow workplace rules for confidential information and performance tracking. The useful evidence may be accepted-work counts, approximate time bands, review burden, and responsibility changes, without retaining confidential content. A bounded account might be: “For comparable assignments, the first pass took less time, while review and incoming volume changed in these ways. Can we agree on output and checking standards, and what work should move?” If the record is incomplete, say so and propose a trial with an owner, measures, and review date. That lets both sides examine whether capacity was released, quality held, or responsibilities expanded before treating one fast task as proof of a job-wide change.

Sources: Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence; Generative AI and the SME Workforce

How should I frame a workload request?

Frame the request around the decision your manager can make: what result must be protected, what capacity is available, and which deadline, scope, or review step should change. If the queue or expected volume has grown, or checking now takes substantial time, explain that and ask which work should move. The goal is to agree what will be delivered, to what standard, and who owns the tradeoff. Separate a temporary surge from sustained redesign. A short spike may call for temporary coverage or a moved date; recurring extra work, a larger queue, or a standing review duty may require a durable change. Bring a brief forecast of key commitments, due dates, remaining effort, and the risk if everything stays in place. Mark estimates as estimates, and focus on the few items creating the conflict rather than presenting a long task list. Translate those facts into choices. If volume rose, state the expected output and capacity gap, then ask whether the target, deadline, or another commitment should change. If review grew, ask who checks the work, what standard counts as acceptable, and how much time is allowed. If faster work is followed by new assignments, name what those assignments displace and ask what should be deferred, reduced, or stopped. For a temporary change, propose an end date; for a recurring one, ask to review normal workload and ownership. The Work Coach playbook, “How to Tell Your Boss You Have Too Much Work,” recommends forecasting commitments, offering tradeoffs, asking for a decision, and recording owners and dates. This is practical guidance, not controlled research. The field experiment reported in “Shifting Work Patterns with Generative AI” offers a useful boundary: across 66 firms and 7,137 knowledge workers, users in the experiment’s second half spent less time on email, but researchers detected no shift in task quantity or composition from individual tool provision. That finding does not predict your team’s workload; it is a reason to ask what changed in your own queue. A concise opening might be: “The first pass is faster, but the review queue and incoming volume have grown. I want to keep the agreed quality standard. With these deadlines, I can finish A and B this week, or include C if we move A or reduce C’s scope. Which outcome should I plan for, and who owns the additional review?” Use only records you are allowed to share. If the evidence is incomplete, say so and propose a bounded check-in rather than overstating the result. Close with one decision, an owner, and a date to revisit the arrangement. Confirm what will be delivered, what moved, who handles review, and when you will check whether the plan worked. The Work Coach playbook advises recapping the decision and changing the calendar to match it. If the manager cannot decide in the meeting, ask who can and by when. A useful result is observable: a priority, deadline, scope, or review duty changes, or a follow-up is scheduled. That makes workload relief a plan both sides can inspect without assuming less work is guaranteed.

Sources: How to Tell Your Boss You Have Too Much Work; Shifting Work Patterns with Generative AI

When is a pay request supported by more than speed?

A pay request has a stronger basis when a faster task accompanies a lasting change in the job: broader scope, higher-level judgment, process ownership, responsibility for quality or risk, or a contribution the organization recognizes. A scheduled compensation review is also a reason to raise the question now. Separate those questions before deciding what to ask. An increase in accepted output may matter to the organization, but without evidence of changed scope or an established review process, its connection to this worker’s pay remains uncertain. Productivity and compensation are different outcomes. A worker may finish a first draft sooner while the organization retains the time saving, assigns more work, or relies on the worker to check and approve the result. The task may be faster in all three cases; only some also change durable responsibilities or support a compensation review. Ask: “What do I now own, decide, or remain accountable for, and is that change expected to continue?” Ongoing responsibility for quality checks, workflow design, escalation decisions, or more accepted work can make the scope question concrete. “Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears,” a 2026 Boston Fed policy perspective, offers limited context. Its two survey waves, in December 2024 and 2025, each drew on the New York Fed’s rotating internet panel of approximately 1,300 U.S. household heads. Among AI-exposed workers in the regression sample, the 6 percent who strongly agreed that AI had increased their productivity had an estimated 14 percent likelihood of saying they were more likely to ask for a raise in the next 12 months because of AI; the neutral group’s estimate was 3.1 percent. These are intentions associated with self-perceived productivity, not verified requests or raises received. The survey does not establish that AI caused an individual to ask, that an employer granted a raise, or that this pattern applies to your workplace. That limit argues against a pay-first pitch based only on speed: “I finished this sooner, so I should earn more.” Connect the evidence to durable duties: “I now own the review and exception process as well as the initial work. Can we review whether this ongoing scope fits my level and compensation, and agree on criteria and timing?” If the responsibility is not yet permanent, ask what would make it part of the role and when to revisit it. Raise pay directly when scope has already changed permanently or a review is due; you need not wait for a long productivity record. Name the added work, decisions, accountability, and how long the change has lasted. “How to Tell Your Boss You Have Too Much Work,” a practical Work Coach playbook, recommends outlining commitments, deadlines, remaining effort, and tradeoffs, then asking the manager to choose among options. Adapt its clarity: ask which compensation process applies, what evidence is considered, who decides, and when you can expect an answer. If no decision is possible in the meeting, request a next step, owner, review date, or criteria for reconsideration. If no date or criteria are offered, weigh your own constraints and alternatives rather than treating an indefinite deferral as assurance. Pay bands, contracts, collective arrangements, and the manager’s authority may shape the available process, so ask about the route that actually governs your role.

Sources: Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears; How to Tell Your Boss You Have Too Much Work

Should I raise pay, workload, or both at once?

There is no single best opening for every worker. Match the ask to the clearest change, keeping the other question visible. When the issue is a fuller queue, more checking, or pressure to deliver the same quality faster, lead with workload and priorities. When the role has acquired durable higher-level duties, decision authority, or accountability, raise compensation directly, especially if a formal review is due. If both capacity and responsibility have changed, put both on the agenda; a workload discussion should not erase a pay question already warranted by the role. Compare four factors. **Net effort:** after preparation, AI use, review, correction, and follow-up, is the cycle easier, or has effort moved elsewhere? **Accepted output and quality:** has finished work changed, and are standards and rework visible? **Responsibility:** who now decides, checks, owns the process, handles exceptions, or answers for failure? **Timing and process:** is this a temporary surge, a sustained change, or the right point in the compensation calendar? A strong workload case can coexist with an uncertain pay case; a clear scope change can justify discussing pay even when total time savings are unknown. The Federal Reserve Bank of Boston’s *Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears* helps explain why perceived gains may affect willingness to ask, but it does not settle what anyone should receive. The two December waves (2024 and 2025) surveyed a nationally representative rotating internet panel of about 1,300 US household heads. In regression estimates, respondents who strongly agreed that AI had increased their productivity had a 14 percent likelihood of saying they were more likely to ask for a raise in the next 12 months due to AI, compared with 3.1 percent among neutral respondents. The strongest-gain category was 6 percent of the sample. These are self-reported perceptions and intentions, not observed requests, raises granted, or proof that productivity caused wage growth. It shows stated intention, not entitlement from a fast task. Sequencing can help when the record is incomplete: establish the work pattern, agree what should be prioritized, then revisit scope and compensation on a date. If you now own higher-level work, say so and ask which role or pay review process applies, while also addressing workload. The Work Coach playbook recommends a short forecast, naming the constraint, offering tradeoffs, and asking the manager to choose. It is practical guidance, not evidence of manager agreement. It makes the request actionable: which date, deliverable, review step, or competing task should move? The request’s size and timing should fit actual constraints. Salary floor, caregiving, health limits, probation, contract terms, pay bands, collective arrangements, local rules, or retaliation concerns can affect what is feasible and through which channel. They may favor private preparation, a representative or HR process, or asking for transparent criteria before negotiating directly. Check policies and seek qualified local guidance where rights or contract terms matter. Before the meeting, choose an opening and fallback. If workload is the clearest change, request a specific tradeoff or bounded review of volume and checking time. If responsibilities have durably expanded, request a formal scope and compensation review, with criteria and an owner if no decision is available immediately. If both are supported, name both and sequence only what needs more evidence. A later date helps only when specific and attached to what will be reviewed. Decide which observable change needs action now and record a next step for the other question.

Sources: Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears; How to Tell Your Boss You Have Too Much Work

How can I make the conversation specific and follow through?

Make the request specific by moving from observation to decision: say what changed in the workflow, what it means for capacity or responsibility, what you need decided, and when you will check the result. For example: “The first pass is faster, but checking and the incoming queue have grown. Can we agree on the output and review standard, what should move, and a date to revisit my scope and compensation?” Keep only clauses your record supports. If the queue did not grow, do not say it did; if compensation is not yet the issue, ask first about priorities.

Before the meeting, select a small, representative sample of ordinary work rather than one unusually easy success. Note the task, accepted endpoint, time spent preparing and checking, any correction or approval, and what happened to waiting work. Add responsibility changes that matter: whether you now set the review standard, own the final decision, or handle escalations. A short comparison is easier to discuss than a complicated spreadsheet, and less likely to confuse speed on one step with a change in the whole role.

Choose the decision you want. If the immediate problem is a growing queue or review burden, ask about capacity or priorities: “At this volume, I can meet the deadline for A or B with the required checks. Which should move, or what support is available?” If duties have become durably broader or a compensation review is due, name that: “These responsibilities are now part of my regular scope. What is the process and timing for reviewing my level and compensation?” If both apply, separate the immediate workload decision from the pay process. Do not hide a pay request inside a general efficiency conversation.

Offer two feasible tradeoffs, not a list of frustrations. Propose a bounded trial with a fixed quality bar, or the same output target with review time protected. Ask who owns checking and which existing task will be deferred if volume rises. The Work Coach playbook “How to tell your boss you have too much work” recommends making commitments and effort visible, proposing choices, and asking the manager to decide what moves. This is practical guidance, not evidence that a manager will accept a particular option; it makes the request answerable.

Close by agreeing on the measure, owner, and review date. For a trial, track accepted output, correction or review effort, and whether the queue or deadline changed. Confirm the decision in a brief written recap, including what moves and when you will follow up. Reopen the discussion if quality slips, workload changes, or responsibility expands; do not let a temporary trial silently become permanent. If your manager cannot set pay, ask which compensation cycle, HR process, representative, or collective route applies. Boston Fed’s “Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears” reports raise-seeking intentions associated with perceived gains, not raises received. That evidence cannot promise an outcome. Keep notes within workplace policy, protect confidential material, and scale the request to your circumstances and available time.

Sources: How to Tell Your Boss You Have Too Much Work; Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears

What should I do before the next meeting?

Before the next meeting, choose one workflow and record a few ordinary cycles, not just the smoothest example. Note how many results were accepted, how much checking or rework they needed, and where the time saved on the first pass went. That is enough to make the next conversation more specific; it is not a forecast of what your manager will decide. The Boston Fed’s survey reports an association between workers’ perceived AI productivity gains and saying they were more likely to ask for a raise. It measured intention, not requests made or raises received, so use it as context for why the question may arise, not evidence that your own pay should change. Then choose the ask that fits your record. If checking or incoming volume absorbs the apparent time saving, request a clear priority, workload, or quality-standard decision. If your duties have already expanded on a lasting basis, or a compensation review is due, put scope and pay on the agenda now. Agree on who will do what and when to revisit the result. If you still need a wider task-by-task view, the free [AI job risk checker](/ai-job-risk-checker) offers change-pressure signals; those signals are not a validated probability of displacement. A measured workflow gives you a grounded opening, while the employer’s response and any pay outcome remain uncertain.

Sources: Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears

Questions readers ask