What Three IT Services Firms Reveal About AI Productivity

The signal

The observed businesses are growing. These public figures do not establish a sudden, company-wide AI productivity breakthrough.

Start with the business, then choose the worker denominator

An employee count is a point-in-time stock. Revenue and profit accumulate throughout a period. Dividing quarterly revenue by the final day’s workforce can misstate the relationship when hiring or departures occur during the quarter. A worker hired near the end contributes little to the earlier revenue but appears fully in the closing denominator.

We use the average of opening and closing headcount as a transparent approximation for Infosys and Cognizant. It is better aligned with the quarter than a closing-only count, but it is not a time-weighted employee census. It also includes non-billable personnel. That makes it a broad company productivity proxy, not a measure of output per programmer or hours saved by an AI tool.

The panel is small and explicit: two companies with paired workforce observations and financial results, plus Accenture as a third case testing the profit baseline. These companies are not a statistically representative sample of global services employment.

The purpose is to establish what can be reproduced now, then identify which interpretation the evidence can bear.

Infosys: a modest improvement after averaging the workforce

Infosys’s filings provide the necessary employee endpoints. The September 2025 quarter’s opening and closing employee counts are 323,788 and 331,991. For the September 2026 quarter, they are 328,062 and 332,542. The resulting averages are 327,889.5 and 330,302. ^[1][2]

Reported US-dollar revenue increased from $5.076 billion to $5.145 billion, while gross profit rose from $1.560 billion to $1.598 billion. Dividing by those average workforces gives approximately $15,481 and $15,577 of quarterly revenue per employee, and $4,758 and $4,838 of gross profit per employee. The calculated gains are 0.62% and 1.69%, respectively. ^[2]

That is progress, but it is materially smaller than a narrative of a step-change in output might suggest. It is also compatible with several causes. The measure does not distinguish better software tooling from higher realized prices, a different service mix or a shift in where work is performed.

The denominator itself changes the impression. Closing headcount was almost flat year over year, but the approximate average workforce increased about 0.74%. A closing-only ratio would therefore assign more of the revenue increase to productivity. Neither calculation reveals actual billable hours; the average is simply the more appropriate first approximation for this question.

Cognizant: revenue productivity improves while profit depends on the measure

Cognizant reported 359,900 employees at the end of September 2026, up 3,200 from June. Its prior-year release reported 349,800 at September 2025, up 6,000 from June. Those disclosures imply approximate quarterly averages of 358,300 and 346,800—an increase of 3.32%. ^[3][4]

Against reported revenue growth of 4.6%, revenue per average employee increased approximately 1.24%. This is calculated from the rounded growth rate, so it should not be mistaken for an exact internally reported productivity measure. ^[3]

Profit complicates the picture. The current release reports a 15.4% GAAP operating margin versus 16.0% a year earlier, while adjusted margin increased to 16.3% from 16.0%. Combining those rounded margins with the same revenue and workforce changes produces an approximate 2.55% decline in GAAP operating profit per employee and a 3.14% increase on the adjusted basis.

Both measures are useful if their purpose is clear. GAAP includes costs borne during the period; the adjusted measure is management’s view after specified exclusions. An investor should inspect those exclusions rather than choose whichever measure best supports a pre-existing AI story. The difference is part of the evidence, not a reason to abandon the analysis.

Accenture: a strong recovery with a difficult comparison

Accenture’s fiscal fourth-quarter release reported revenue growth of 6% in US dollars and 7% in local currency. Its GAAP operating margin was 15.3%, against 11.6% a year earlier. But the prior-year adjusted margin, excluding business-optimization costs, was 15.1%. The comparison is therefore a 3.7-percentage-point GAAP recovery or a 0.2-point improvement against that adjusted baseline. ^[5]

The company’s fact sheet reports approximately 814,000 employees at fiscal year-end. That single point does not provide a matched quarterly average workforce, so we do not manufacture an Accenture profit-per-average-worker figure from it. ^[6]

The financial result remains positive. The analytical issue is attribution: a cost that affected the prior-year comparison should not be relabelled as a sudden improvement in AI-driven delivery. The smaller baseline-adjusted comparison may still include useful operational progress, but the headline margin change is an unreliable estimate of it.

This is why a cross-company leaderboard based on reported profit per employee can go wrong. A restructuring year, an acquisition or a change in outsourcing can move the ratio sharply even when the underlying production process changes much less.

What the figures say about shrinking demand

The two matched companies report revenue growth, and both approximate average workforces increased. Accenture also reports growth. Within this limited panel, the evidence does not support a simple picture of collapsing service demand accompanied by a shrinking workforce.

That observation does not settle what happens to particular roles. A company can hire in one practice, reduce another, shift work geographically and change the skills it needs while total employment grows. The aggregate panel cannot identify which jobs were replaced, which tasks became easier or whether entry-level career paths narrowed.

There is also an economic difference between doing existing work more efficiently and creating demand for new work. A services firm might use automation to deliver a contract with fewer hours, then deploy the freed capacity to another client. Alternatively, customers may negotiate lower prices because the work takes less effort. The same technological capability can therefore produce different revenue and margin outcomes.

The observed company ratios are where those effects eventually meet. They should be read alongside contract mix, client demand and costs, rather than used as a shortcut to claims about every worker.

The bullish case and the measurement lag

The strongest reason not to dismiss AI productivity is that implementation can precede financial recognition. Firms may spend on training, redesign delivery processes and absorb transition costs before better economics emerge in renewals. A modest current improvement does not rule out larger benefits later.

It can also be rational for a provider to pass some efficiency gains to customers. Lower prices may support retention or open a previously uneconomic market. In that case, slower revenue per employee could coexist with improved output or customer value. Company financial statements do not directly measure either quantity.

The skeptical case is that investors may capitalize an expected productivity gain before the provider has shown an ability to retain it. If competition transfers most of the savings to customers, a genuine technological improvement need not produce the margin expansion implied by an optimistic valuation.

DN’s test is therefore joint: comparable revenue, profit, average workforce and customer demand should be considered together. A rising ratio caused solely by a contracting denominator is weaker evidence than profit per worker rising while demand and delivery quality remain healthy. Even the latter remains consistent with more than one cause.

Use the tool to separate scale from productivity

The illustrative calculator takes revenue, a consistently defined profit measure, and opening and closing employees for two equal periods. Its hypothetical defaults show revenue rising from 100 to 110, profit from 20 to 23 and an unchanged average workforce of 1,000. That produces 10% revenue-per-worker growth and 15% profit-per-worker growth.

Change the workforce to 1,100 and the revenue-per-worker improvement disappears. Change only the profit definition between periods and the calculation becomes economically invalid even if the arithmetic runs. The calculator can reject zero denominators and invalid numeric inputs; it cannot discover a changed accounting definition, acquisition perimeter or subcontractor boundary.

Readers must therefore confirm those conditions before interpreting the output. For the observed Infosys and Cognizant cases, we preserve reported-dollar figures and label the approximate workforce basis. We do not claim to remove all acquisition or foreign-exchange effects, and we do not convert the ratios into estimated hours saved.

The practical decision is whether the latest disclosure warrants a higher expected margin, a different contract price or simply continued observation. These three cases support modest operational progress and a demand picture more resilient than the replacement narrative suggests. They still leave the share captured by shareholders—and the share attributable specifically to AI—to be demonstrated over subsequent periods.

The evidence, side by side

Average Workforce Productivity Bridge

Hypothetical inputs. Use equal periods, a consistent profit definition and a matched business perimeter.

Numeric guards require positive workforce, prior revenue and prior profit; current revenue cannot be negative. User preconditions: equal periods, consistent profit/worker definitions, matched perimeter and currency. These semantic conditions are not automatically verified.

Scenarios to watch

Productivity with demand

Conditional DN interpretation

Comparable profit per average worker rises alongside stable or growing demand. This is stronger evidence of durable economics, while attribution still requires deployment evidence.

Savings passed to customers

Conditional DN interpretation

Tasks take less effort but contract prices decline. Customer value improves without equivalent vendor margin gains.

Ratio improved by denominator or accounting

Conditional DN interpretation

Employment cuts, outsourcing or a costly comparison period lift the ratio without comparable growth in underlying delivery output.

Method and limitations

Research cutoff: 13 November 2026 UTC. Six primary documents cover Infosys, Cognizant and Accenture. Infosys averages are (323,788+331,991)/2=327,889.5 and (328,062+332,542)/2=330,302. Revenue inputs are $5,076 million and $5,145 million; gross profit is $1,560 million and $1,598 million. Cognizant opening counts are derived from disclosed quarter-end counts and quarterly increases: 343,800 and 356,700; averages 346,800 and 358,300. Its productivity estimates use rounded revenue growth 4.6% and margins 16.0%, 15.4% and 16.3%, so the derived percentages are approximate. We use same-quarter year-over-year periods, do not annualize, and keep Accenture’s fiscal-quarter baseline comparison separate. This is two matched workforce cases plus a third accounting case, not a completed sector panel.

Opening/closing averages are approximations and can miss uneven hiring, acquisitions or departures within the quarter. Employees are not equivalent to hours worked, billable capacity or full economic labor input; subcontractors can change the boundary. Reported-dollar results retain currency and business-mix effects. No causal AI estimate, occupation-level displacement estimate or cross-company productivity ranking is claimed. GAAP and adjusted profit outputs must remain separately labelled.

Primary sources and evidence register

Source dates and reporting periods differ. Several documents from one issuer are one source family, not independent confirmations. The evidence download contains the observations used in this article.

Frequently asked questions

Why use average rather than closing employees?

Revenue and profit accumulate through the quarter. Averaging the opening and closing workforce reduces one timing mismatch, although it remains an approximation rather than a time-weighted count.

Does revenue per employee measure AI productivity?

It is a broad financial proxy. Prices, utilization, business mix, currency, acquisitions and outsourcing can all change it without any change in AI-driven task efficiency.

Why do Cognizant’s two profit measures disagree?

GAAP and adjusted operating margins include different costs. Applying the same workforce denominator preserves that difference rather than choosing one definition without explanation.

Do these results show consulting demand is collapsing?

No. The three examined businesses report revenue growth. The small panel cannot establish the whole industry’s outlook or what happens to particular occupations.