AI Productivity Multiplier by Role: What Studies Actually Show (2026)

What "productivity" actually means in these studies

Most studies measure one of three things: (1) tasks completed per unit of time in a controlled lab, (2) tickets, pull requests, or contracts closed per week in a live workplace, or (3) self-reported time savings on surveys. These are not interchangeable. A controlled-task speed-up of 55.8% — the famous GitHub Copilot result from Peng et al. — does not mean developers ship 55.8% more features per quarter. It means that when 95 developers were assigned an identical HTTP server task, the Copilot group finished a median of 71 minutes faster than the control group's 161 minutes.

The most rigorous real-world number we have for developers comes from a 2024 randomized controlled trial across Microsoft, Accenture, and an anonymous Fortune 100 firm, with 4,867 developers. Copilot users completed 26.08% more tasks per week on average. That is a real, deployed, field-study number — and it is less than half the controlled-task gain. The gap between lab and field is the single most important fact on this page. When a vendor cites a study number, check whether it was measured in a lab or in production. The honest planning multiplier is the field number, not the lab number.

Self-reported survey numbers (Microsoft Work Trend Index, RAND educator survey) tend to land in the 25–40% time-savings range. They are useful as a cross-check, not as primary evidence. Workers consistently overestimate AI time savings on diaries and surveys versus what objective task-completion data shows.

Why novices gain more than experts (and what that means for hiring)

The single most replicated finding across the 2023–2025 literature is that generative AI compresses the skill distribution. In the Brynjolfsson, Li, and Raymond NBER study of 5,179 customer support agents at a Fortune 500 software firm, the overall productivity gain was 14%. But that average hid a stark pattern: novice and low-skilled agents gained 35%, while the most experienced agents showed essentially no measurable gain. The AI was teaching the bottom of the distribution to perform like the top.

The Noy & Zhang ChatGPT writing study (Science, 2023) found a similar compression: workers with weaker baseline writing ability gained the most. So did the BCG/HBS consulting study — below-average consultants moved up to or past the control group's average performance when using GPT-4 on in-frontier tasks. This has clear implications for staffing models: AI is most valuable as a floor-raiser for less experienced staff, not as a ceiling-raiser for senior experts. If you are budgeting AI tools by seat, the highest ROI per seat is at the junior end of your org chart.

There is one important counter-finding. In the BCG study, when consultants used GPT-4 on tasks outside the model's competence frontier — problems requiring numerical reasoning the model gets wrong — they performed 19 percentage points worse than the control group. Junior staff are the most likely to miss when a task is outside the frontier. That is the trade-off: AI raises the floor for in-frontier work and lowers it for out-of-frontier work.

Software engineering: the 55.8% number, in context

The 55.8% Copilot speed-up from Peng et al. (2023) is the most-cited AI productivity statistic in the world. The study itself is small (95 developers) and narrow (a single HTTP server task in JavaScript). Treat it as an upper bound for what tightly-scoped boilerplate work looks like with AI assistance, not as a forecast of weekly throughput.

The 2024 GitHub / Microsoft Research / Accenture three-firm RCT is the better number for planning. With 4,867 developers in production, Copilot users completed 26% more tasks per week, opened more pull requests, and showed no significant drop in code quality across the metrics studied (build success, code review pass-through, defect rates). GitHub's Octoverse 2024 report adds context: developers using Copilot accept roughly 30% of suggestions, and AI-generated code now accounts for a meaningful share of new commits in observed repos.

For a fuller breakdown of which AI coding tool gives the best return — and whether Cursor's larger context window beats Copilot's tighter IDE integration — see Copilot vs Cursor for developers ROI. For the workday-level view of where the time actually goes, see how much time AI saves by task in 2026.

Customer support: the 14% headline and what's underneath

Brynjolfsson, Li, and Raymond's NBER paper "Generative AI at Work" remains the gold-standard real-world deployment study. Over a year, they observed 5,179 customer support agents at a Fortune 500 enterprise software company. Agents using a generative-AI-based conversational assistant resolved 13.8% more customer issues per hour. Customers reported better experiences and were less likely to ask for a supervisor. Employee attrition fell among agents who used the tool.

The breakdown by skill level is what changed the field: novice agents (in their first two months) gained 34% productivity and reached the performance level of agents with more than six months of tenure. The most experienced agents — top quartile — showed no statistically meaningful gain. The AI was effectively codifying and distributing the implicit knowledge of senior agents to newer ones. For workforce planning, this means a contact center can ramp new hires roughly twice as fast, but does not get much extra leverage out of its veterans.

If you're trying to translate this into a staffing or savings number for your own contact center, see AI tool cost vs salary savings.

Knowledge workers: consulting, writing, and the "Jagged Frontier"

The Dell'Acqua et al. "Navigating the Jagged Technological Frontier" study (HBS / BCG / Wharton / MIT, 2023) is the most important paper on the limits of AI productivity. 758 BCG consultants were randomized to use GPT-4 (some with light training, some without) or no AI on 18 realistic consulting tasks. On tasks inside the model's competence frontier — creative brainstorming, structured writing, persuasive memos — consultants using GPT-4 completed 12.2% more tasks, did them 25.1% faster, and produced output rated 40% higher in quality.

On tasks outside the frontier — a quantitative business problem requiring numerical reasoning the model handled poorly — GPT-4 users were 19 percentage points less likely to produce a correct answer than the control group. The model gave confident, fluent, wrong answers, and consultants trusted them. The lesson is not that AI hurts; it is that AI shifts the bottleneck from execution to judgment about when to use it.

The Noy & Zhang Science paper (2023) on writing tasks landed in a cleaner part of the frontier. 453 college-educated professionals wrote business documents — press releases, short reports, analysis plans. The ChatGPT group finished 37% faster and produced output rated 18% higher in quality by blinded evaluators. As with customer support, the productivity distribution compressed: weaker writers gained the most. For a broader view of which tool to reach for in each profession, see best AI tools by profession in 2026.

Roles where AI helps less (or not at all, yet)

Not every role gains 25–55% from generative AI. The gains shrink when the work is highly physical, deeply tacit, heavily regulated, or dominated by judgment under uncertainty. McKinsey's "Economic potential of generative AI" (2023) and follow-up "State of AI in 2024" estimate the smallest near-term productivity impact in roles like skilled trades, construction supervision, surgical and physical-procedure medicine, on-site field service, and frontline manufacturing. The 2024 McKinsey report estimates generative AI could add 0.1–0.6 percentage points to annual labor productivity through 2040, with the bulk concentrated in knowledge-worker functions.

Within white-collar work, the studies suggest AI helps least where the bottleneck is interpersonal judgment (senior management, complex client negotiation), highly regulated work where output must be verifiable line-by-line (audit sign-off, FDA submissions, certain legal opinions), and any task where hallucination cost exceeds drafting savings (medical diagnosis without a clinician in the loop, financial planning advice). For all of these, AI can still draft, but the post-AI review time can equal or exceed the time saved.

If you want a longer view of which roles eventually shrink rather than just speed up, see AI-replaceable jobs by 2030.

Methodology: how to read these numbers without getting fooled

When you encounter an AI productivity claim, ask four questions. First, is the gain measured in a controlled lab task or in production over a multi-week window? Lab numbers are roughly 1.5–2x production numbers. Second, what skill level is the worker? A 35% novice gain and a 0% expert gain average to 14% — and that 14% is the headline. Third, is the task inside the model's frontier? A study that handpicked in-frontier tasks will look better than your real workload. Fourth, who paid for the study? Vendor-funded studies (some Microsoft 365 Copilot numbers) tend to land higher than independent academic RCTs.

A reasonable planning rule: take the largest peer-reviewed gain for a role, halve it for real-world weekly throughput, and halve it again if your team is mostly senior. A development team using Copilot might plan for 13% more shipped tasks per week — half of the 26% field-study gain — even though the lab study says 55.8%. A senior consulting team using GPT-4 might plan for 6–12% throughput gain, not the 25% headline. These are conservative numbers, but they are the ones that survive contact with reality.

How AI productivity gains translate into pay and pricing

A 25% productivity gain does not automatically translate into a 25% pay raise or a 25% lower rate. The split between worker, employer, and customer is set by market structure. In tight labor markets where AI-fluent workers are scarce (senior ML engineers, AI-native product managers), more of the gain accrues to the worker through higher salary. In commoditized markets (basic copywriting, level-1 support), the gain often accrues to the buyer through lower prices.

For freelancers and contractors, the practical question is whether to keep your hourly rate and finish faster (more clients, more income), or lower your rate and capture more market share. The right answer depends on whether your bottleneck is demand or supply. See the freelance AI rate calculator for a model that walks through both strategies, and the AI tool cost vs salary savings calculator for the employer-side view.

How to estimate your own AI productivity multiplier

1. 1. Pick the closest peer-reviewed role from the table

   Start with the row in the table above that most closely matches your day-to-day work, not your job title. A "product manager" who mostly writes specs and emails is a knowledge worker in the Noy & Zhang writing study (37% faster). A PM who runs SQL queries is closer to the data analyst row (30–40% faster). A PM who negotiates contracts is closer to the lawyer drafting row.

2. 2. Halve the lab number to get a planning number

   Controlled-task gains are roughly 1.5–2x larger than field-deployment gains. If the headline for your role is 50% faster, plan for 25%. If it's 30%, plan for 15%. This is the conservative number that survives a quarter of real-world use.

3. 3. Adjust for your skill percentile

   If you are in the top quartile of your role, halve the planning number again — the Brynjolfsson NBER pattern of near-zero gain for experts applies broadly. If you are in the bottom half, keep the full planning number. AI is a floor-raiser more than a ceiling-raiser.

4. 4. Subtract the time you spend reviewing AI output

   For any high-stakes deliverable (legal, medical, financial, customer-facing copy), measure the time you spend verifying AI output. Subtract it from the gross time savings. For a contract draft, this can erase 30–50% of the gain. For an internal Slack summary, it's negligible.

5. 5. Multiply by the share of your week the task covers

   Your productivity multiplier is task-weighted. If AI makes you 40% faster at writing but writing is only 30% of your week, your weekly productivity gain is roughly 12%, not 40%. Run the math for each major bucket of your week and sum.

6. 6. Track real output for 4 weeks

   Pick a measurable unit (tickets closed, PRs merged, pages drafted, decks shipped). Track weekly output for two weeks without AI, then two weeks with AI on the same workload. The delta is your actual multiplier. Most teams find it lands 30–50% below their self-reported survey estimate.

Frequently asked questions