Generative AI and Automation: IT Services Unit Economics

Review period: September 30, 2008–October 3, 2026; evidence concentrated in 2023–2026, with one historical outsourcing-contract comparator. Scope: Delivery productivity, costs, utilization, margins, revenue per engagement, and pricing/value allocation across IT services and adjacent technology-enabled operations. Source policy: Prioritize original filings, contracts, rate cards, provider/client disclosures, and academic research; distinguish measured results, company-reported benefits, contract economics, projections, and synthesis.

Executive summary

Generative AI changes delivery effort more directly than it changes revenue or profit. Enterprise coding experiments report 26.08% more completed tasks, while a customer-support study finds 15% more issues resolved per hour; a separate experiment on experienced developers’ own repositories instead found completion times increased 19%. These are different settings and measures—not an industry-wide cost-saving estimate. (microsoft.com)

Pricing and capacity conversion determine who captures the gains. With unchanged hourly rates and scope, fewer billable hours reduce engagement revenue. Fixed fees, subscriptions, and accepted-output pricing can preserve revenue while reducing delivery costs—but only if labor capacity is redeployed or removed and additional AI, assurance, and transition costs remain below the savings. Providers already describe sharing productivity with clients; this is not solely a hypothetical contracting issue. (cognizant.q4cdn.com)

Public disclosures show positive operating developments, but not a clean industry-wide AI margin effect. Accenture’s FY2026 adjusted operating margin increased only 20 basis points, despite extensive AI activity. (sec.gov)

Revoqra synthesis: The most defensible alternative is usually a hybrid: a foundation/run fee, well-defined accepted-output or consumption charges, quality safeguards, and selective gainsharing—not unrestricted hourly billing or an uncapped promise to deliver business outcomes.

1. What must change for productivity to improve unit economics?

Define the outcome before measuring the hours

The appropriate denominator is an accepted, quality-adjusted client deliverable, not generated code, prompts, automated actions, or tickets marked closed. Examples include a deployed change that passes acceptance and security tests, an application migrated without material regression, a correctly processed invoice, or a support issue resolved without an avoidable repeat contact.

This distinction matters because the research measures different things: completed development tasks, issues resolved per hour, and elapsed task-completion time. None directly measures the fully loaded profit on an outsourced engagement. (microsoft.com)

Revoqra analytical framework:

Delivery cost per accepted outcome = human effort × effective labor cost + AI/nonlabor variable costs + allocated fixed costs + expected failure costs

The components should include:

• Human effort: discovery, design, implementation, review, testing, integration, deployment, and ordinary exception handling—not just initial drafting or coding.
• Effective labor cost: compensation and benefits, contractor rates, seniority/location mix, and the cost of unused capacity.
• AI/nonlabor costs: licenses, inference, retrieval, storage, orchestration, monitoring, and other platform consumption.
• Fixed costs: transition, reusable-platform development, connectors, training, and delivery overhead, without double-counting expenses already included elsewhere.
• Failure costs: additional rework, service credits, remediation, and other losses not included in ordinary delivery effort.

A useful capacity formulation is human cost per outcome = h × w / u, where h is productive delivery effort, w is cost per available paid hour, and u is the proportion of paid capacity productively allocated. Published billable-utilization measures may not exactly match this economic definition.

Labor remains economically material. Infosys reported FY2026 consolidated employee-benefit expenses equal to 53.2% of revenue and technical-subcontractor costs equal to 8.6%. These are group expense ratios, not estimates of avoidable labor cost on an individual project. (infosys.com)

Keep five measures separate

These distinctions are the basis of the analysis below, rather than claims that every provider reports these measures consistently.

2. Productivity evidence: measured gains, contradictions, and projections

Dates below are publication dates unless an observation window is explicitly stated. Results should not be averaged across rows.

The early negative METR result should therefore not be frozen into a claim about all 2026 tools. Conversely, the follow-up does not validate a universal positive productivity number. METR’s February–April 2026 survey of 349 technical workers found median self-reported changes in work value of 1.4–2×, but those are perceptions of value, not measured hours or cash costs. (metr.org)

Throughput gains are not the same percentage as hours saved

Revoqra-derived algebraic conversions—not additional empirical findings: At fixed hours, identical task mix, and unchanged quality, a 26.08% throughput increase would imply approximately 20.7% fewer hours for a fixed output: 1 − 1/1.2608. A 15% increase in resolutions per hour similarly implies approximately 13.0% fewer hours per resolution. The inputs are the source-reported estimates above; these conversions do not establish end-to-end delivery or cash savings. (microsoft.com)

The experiments also involve different tools, worker experience, tasks, and adoption patterns. A favorable task-level estimate must survive integration, review, security, deployment, and subsequent maintenance before it can justify a contractual productivity commitment.

3. Provider disclosures: economic progress, but incomplete AI attribution

The following is a comparison of evidence, not a ranking of comparable margins. Fiscal periods differ; adjusted operating margin, EBIT margin, segment profit margin, and cost-of-services ratios are different measures.

Provider-defined AI revenue categories should not be treated as comparable measures of incremental revenue or profit. They also do not reveal how much new AI work offsets lower fees for existing services.

Existing pricing is already more varied than a pure labor-arbitrage description suggests

Cognizant’s 2025 Form 10-K provides an unusually useful revenue breakdown:

Calculation note: Shares divide each reported category by $21.108 billion total revenue. Sample: one issuer, one fiscal year, three contract categories; figures are rounded. This is not an industry estimate. (sec.gov)

Similarly, Genpact said approximately 45% of total 2024 revenue and 70% of Advanced Technology Solutions revenue was non-FTE based. Its definition includes fixed fees as well as outcome- and consumption-based contracts. Therefore, non-FTE revenue is not synonymous with outcome-based revenue. (investorday2025.genpact.com)

Price competition concerns the number of units as well as the rate

Cognizant’s CEO described competition as delivering the same output with fewer units, rather than simply lowering the billing rate. He said productivity could be shared with clients while partly retained, and argued that cheaper delivery could unlock previously uneconomic legacy modernization. This is management’s commercial interpretation—not a measured estimate of demand elasticity or market-wide value creation. (cognizant.q4cdn.com)

4. Contracts, public pricing rules, and client examples

A. Contract language already allocates productivity and volume risk

Cognizant’s 2025 filing says many contracts tie compensation to performance standards, productivity improvements, or milestones. Failure can reduce fees, increase delivery costs, delay payment, or create damage claims. Productivity gains are therefore not necessarily an unrestricted provider windfall even under an existing contract. (sec.gov)

A historical primary example is the September 30, 2008 Health Net–Cognizant master services agreement. It provides no general minimum or maximum service-volume commitment, subject to a specified schedule exception; permits Health Net to reject changes increasing its own or other suppliers’ costs; and requires cooperation with client-performed or third-party services. This predates GenAI and is not evidence of current AI pricing. It illustrates why volume guarantees, change approval, and insourcing/transition rights matter to value capture. (sec.gov)

B. Public commercial rules show viable alternatives—but different definitions of value

These platform examples demonstrate workable pricing mechanisms. They do not establish that traditional IT outsourcing portfolios have already converted to those mechanisms.

Salesforce’s own architecture guidance explicitly distinguishes per-action and per-conversation economics and calls for understanding actions per business transaction. That is the relevant bridge between a rate card and the cost of a useful client result. (architect.salesforce.com)

C. Genpact: directly relevant contract-cohort economics, with important omissions

At its June 26, 2025 investor day, Genpact associated migration of existing accounts from FTE-led to agentic delivery with 3% net revenue growth after productivity commitments and 300 basis points of gross-margin improvement. The presentation ties the figures to agentic contracts signed through May 31, 2025. (investorday2025.genpact.com)

Classification: Company-reported account/contract economics—not a randomized productivity experiment or a demonstrated company-wide AI effect. The relevant slide does not disclose cohort size, a complete cost bridge, or separate realized P&L results from signing economics. Nevertheless, it is more directly informative about monetization than a stand-alone coding-speed benchmark.

HFS’s original research also describes an anonymized US health-insurer arrangement paying the vendor 17% of net adjudication savings above baseline, with a reported $38 million annual benefit. Without a named client, underlying contract, or auditable cost bridge, this is evidence of a reported gainsharing structure—not a verified benchmark for achievable savings or an appropriate universal sharing percentage. (hfsresearch.com)

D. IBM as its own client: cost reduction beyond a laboratory task

IBM reports AskHR resolving 94% of common inquiries, a 75% reduction in support tickets, and a 40% reduction in the HR operating budget over four years. The transformation began in 2016. Its broader Client Zero program combines AI, hybrid cloud, automation, consulting, and process changes. (ibm.com)

Classification: Provider-reported operational and budget results, without a control group. The budget reduction is closer to economic value than hours saved, but cannot be assigned wholly to generative AI. Nor does an internal cost reduction establish the fee or margin achievable when selling the same capability to another organization.

E. Klarna: client value capture and the difference between work equivalence and jobs

Klarna’s SEC prospectus excerpts report approximately $39 million of customer-service cost savings in 2024 and AI handling 69% of customer-service chats in the twelve months ended June 30, 2025. The reported equivalence to more than 700 full-time agents was estimated from reductions in human-handled conversations—not a verified count of 700 layoffs. (sec.gov)

A later annual-filing excerpt reports 31 million AI conversations and 80% of customer-service chats during 2025. These claims demonstrate continuing adoption, not proof that all support can dispense with humans. Validation here was limited to SEC-indexed filing excerpts; no complete AI cost bridge was reviewed. (sec.gov)

Independent AP reporting describes the continuing need for skilled human assistance in complex cases such as identity theft. This is a quality and exception-handling constraint, not evidence that the company abandoned AI. (apnews.com)

Revoqra inference: A client that substitutes technology for previously purchased human handling can capture cost savings while reducing an outsourcer’s addressable labor revenue. The model/platform provider captures a technology fee, but the public disclosures do not establish the division of net value among every participant.

F. Adoption forecasts must remain forecasts

HFS’s 2025 research covering 40 service providers reported an average 30% of 2024 GenAI engagements containing outcome, consumption, or risk-sharing components. Providers expected the share to exceed 50% by 2026. The latter remains a forecast in that source—not a verified 2026 industry adoption rate. The engagement-level sample size and market weighting were not disclosed in the cited summary. (hfsresearch.com)

5. Worked example: the same productivity improvement can halve profit—or increase it

Revoqra illustrative scenario—not a forecast or provider result. Hold scope and quality constant at 100 synthetic accepted work units:

• Before automation: 1,000 human hours, billed at $100/hour.
• Avoidable or redeployable delivery labor cost: $60/hour.
• Other delivery costs: $10,000.
• After automation: 750 hours, plus $5,000 incremental AI/assurance costs.

Baseline delivery cost is $70,000. Post-automation cost is $60,000: $45,000 labor + $10,000 other + $5,000 AI/assurance.

Three conclusions follow from the assumptions:

1. Time-based pricing can reduce gross profit despite better delivery efficiency. Here revenue falls $25,000 while costs fall only $10,000.
2. A higher margin percentage does not necessarily mean higher profit dollars. At $90,000 revenue, gross profit remains $30,000 despite a higher margin rate.
3. Output pricing does not automatically let the provider retain all savings. Competitive negotiations determine the fee and sharing rule.

The example assumes labor capacity can actually be redeployed or removed. If employees remain paid and idle, the modeled reduction in labor cost is not a company-level cash saving.

Utilization and demand are the conversion mechanisms

Another illustrative calculation: if required delivery hours fall 25%, with unchanged paid capacity and demand, economic utilization falls from 80% to 60%. To preserve hourly-billed revenue at unchanged rates, the provider would need 33.3% more comparable output volume, because each outcome now requires only 75% as many billable hours.

This volume response is not guaranteed. Providers can benefit through additional scope, new engagements, reduced contractor use, or smaller capacity—but must demonstrate which mechanism occurred.

Published utilization also requires careful interpretation. Infosys’s June 2026 measure was 84.9% excluding trainees but 82.1% including them. (infosys.com) Accenture separately explained that faster mobilization meant starting new contracts sooner, rather than increasing utilization. Consequently, faster revenue conversion and greater productivity need not show up in the same operating metric. (investor.accenture.com)

6. Pricing-model comparison

The following is Revoqra synthesis: conditional economic mechanisms and plausible risks, not observed industry-wide frequencies.

Kyndryl’s official Bridge Service Catalog documentation illustrates a digital interface for subscribing to services and partner solutions. Such a catalogue is evidence of a delivery/commercial channel—not, by itself, proof of software-like margins or a particular recurring-revenue model. (docs.kyndryl.com)

Output outcomes versus business outcomes

An accepted migration, correct invoice, or resolved issue is a delivery outcome. Revenue growth, working-capital release, or enterprise-wide cost reduction is a business outcome, with more external dependencies. Charging for a delivery outcome can align incentives without making the provider insure results it cannot control.

Pure business-outcome pricing is most credible when the counterfactual, causal contribution, decision rights, and payment measurement can all be specified. Otherwise a base fee plus limited at-risk component is a more defensible allocation of uncertainty.

7. Fit across service lines

These are proposed contracting approaches, not claims about the prevailing pricing mix in each service line.

The evidence supports particular components rather than complete replacement of human delivery: coding assistance improves some task throughput; agent augmentation improves some support workflows; and IBM’s internal transformation combines automation with process and operating-model changes. (microsoft.com)

8. Who captures the value—and when?

Clients

Clients capture reduced invoices under actual-hours billing, negotiated price reductions, internal operating-cost reductions, and the value of faster or better delivery. But released employee time is not automatically reduced expense: it may instead fund additional work.

Services providers

Providers capture retained fixed-fee or subscription economics, contribution from redeployed capacity, additional scope, and returns on reusable assets. Those benefits compete with productivity commitments, renewal pressure, training/transition costs, and increased AI assurance obligations. Infosys’s margin-expansion program explicitly combines value-based selling, staffing/location efficiencies, automation/GenAI, portfolio changes, and indirect-cost optimization—illustrating the multiplicity of levers. (infosys.com)

Model, cloud, and platform suppliers

Suppliers capture seat, consumption, and outcome fees. Public GitHub and Salesforce terms demonstrate why an IT provider’s own subscription or fixed-price promise can sit above a variable-cost technology layer. The resulting value split depends on actual usage and negotiated terms, not just the advertised seat or action price. (docs.github.com)

The contract lifecycle changes the split

Revoqra synthesis: An existing fixed fee can leave early net savings with the provider; a renewal or rebid can transfer more of them to the client. Gainsharing makes the division explicit but introduces baseline and verification costs. Minimum commitments, scope expansion, technology ownership, and termination/transition rights can be as important as the nominal pricing label.

9. Risks that can defeat otherwise plausible economics

1. Local acceleration, downstream bottlenecks. Faster drafting or code generation may increase review, testing, integration, or remediation work. Measure the entire accepted-delivery cycle and subsequent quality—not just the assisted step. The conflicting developer experiments make this a consequential measurement issue. (microsoft.com)
2. Capacity stranded rather than monetized. Savings require redeployment or removal of paid capacity. Staffing/location changes and contractor reductions should be distinguished from AI-induced effort reductions.
3. A costly residual workload. After simple cases are automated, humans may handle a more complex mix. Average human handling cost can rise even while total service cost falls; segment results by complexity rather than assuming the residual cases resemble the original average.
4. Unreliable outcome verification. Intercom’s billable procedures can include handoffs, while Zendesk distinguishes contained from model-verified resolutions. Contractual definitions therefore require inspection; neither a closed conversation nor a model’s approval should automatically be equated with lasting client value. (intercom.com)
5. Consumption and dependency exposure. Track actions and other usage per successful business transaction, rather than unit prices alone. This is also the approach Salesforce’s architecture guidance recommends. (architect.salesforce.com)
6. Baseline disputes and apparent rather than net savings. Gainsharing should adjust for volume, case mix, inflation, client process changes, and transformation/AI costs. Separate realized expense removal, avoided future expense, released capacity, and estimated business benefit.
7. Risk transferred faster than control. Outcome guarantees are fragile when the provider does not control data quality, client decisions, integration dependencies, or demand. Define decision rights, exclusions, change control, and liability boundaries before setting the at-risk fee.
8. Accounting and cash timing obscure economics. Infosys’s audited disclosures describe effort/cost-based progress recognition and significant judgment in estimated completion costs for certain fixed-price contracts. AI-related estimate changes can therefore affect reported progress as well as physical delivery economics. Track realized contribution and cash collection alongside reported margins. (infosys.com)

Items framed as mechanisms above are analytical risks, not estimates of their incidence across the industry.

10. A practical evidence and contracting standard

Before committing to productivity-linked pricing, construct a matched engagement-level baseline and post-adoption ledger:

• Outputs: accepted units, scope, complexity, quality, and observation window.
• Effort: human hours by task and role, including review, exceptions, integration, and post-release work.
• Technology costs: licenses, compute, credits/actions, retries, storage, monitoring, and evaluation.
• Capacity: billable and delivery utilization, contractors, redeployment, bench time, and genuinely avoided hiring.
• Commercial results: invoice reductions, unit prices, scope/volume expansion, recognized revenue, gross profit dollars, and cash collection.
• Client value: realized cash savings, avoided expense, released capacity, and business benefit, reported separately.
• Quality: acceptance failures, reopens, repeat contacts, incidents, service credits, and downstream remediation.

Use randomized or matched comparisons where practical; disclose N, task mix, observation period, tool/model versions, and excluded costs. Do not attribute simultaneous changes in geography, staffing pyramid, acquisitions, or client scope wholly to AI.

A workable hybrid architecture:

1. A fixed fee for discovery and establishing the data/integration foundation.
2. A recurring fee for availability, governance, security, and core run operations.
3. Volume-banded charges for accepted deliverables or successful transactions.
4. Transparent, capped or approval-controlled technology consumption.
5. Limited gainsharing on independently verifiable net benefits.
6. Predefined rates/change control for genuinely exceptional work.

Bottom line

The evidence establishes that GenAI can materially improve particular delivery tasks and that non-headcount commercial models are practical. It does not establish a universal net cost reduction, an industry-wide AI margin uplift, or a single optimal pricing model.

For providers, the key condition is that all-in delivery cost declines outrun fee compression—or that new scope and reusable capabilities fill the gap. Cognizant’s CFO articulated a similar distinction between revenue pressure and maintaining margins through internal productivity. (cognizant.q4cdn.com) For clients, the critical question is whether the agreement purchases genuine, verified value at a lower total cost, rather than merely renaming labor hours as actions, credits, or outcomes.

Sources and evidence limitations

Principal sources

• Productivity research: Microsoft Research’s three-enterprise experiment summary; Generative AI at Work in the Quarterly Journal of Economics; METR’s original developer experiment, February 2026 follow-up, and May 2026 survey. (microsoft.com)
• Provider financial and commercial disclosures: Accenture FY2026 results and call; Infosys FY2026 annual report and June 2026 facts/call; Cognizant’s 2025 Form 10-K and April 2026 call; Kyndryl FY2026 disclosures; IBM 2025 results; HCLTech June 2026 results. (sec.gov)
• Contract and cohort evidence: Health Net–Cognizant executed MSA; Genpact’s June 2025 investor-day presentation and transcript; HFS’s original provider research and anonymized contract examples. (sec.gov)
• Public commercial rules: Intercom pricing; Zendesk resolution-tier documentation; Salesforce pricing, rate card, and architecture guidance; GitHub organizational billing documentation; UK G-Cloud vendor service definition. (intercom.com)
• Client and operational examples: IBM Client Zero; Klarna SEC-indexed prospectus/annual-filing excerpts; AP reporting on complex human-support requirements. (ibm.com)

Limits of the conclusions

The reviewed materials do not provide a sufficiently complete, matched set of engagement-level hours, accepted outputs, prices, technology costs, quality, and realized cash margins to calculate an industry-wide effect. Vendor/client operational cases are generally uncontrolled; anonymized contract examples cannot be independently reconstructed; public rate cards do not reveal negotiated enterprise prices. The report uses source-reported experimental estimates, explicitly labeled algebraic conversions, and a synthetic commercial scenario. No Revoqra significance tests or pooled industry productivity estimate were performed.