DCF Research

Enterprise GenAI ROI: How to Measure and Audit It

By DCF ResearchEdited by Peter KorpakPublished · Updated

Enterprise GenAI ROI is credible when a program's fully loaded cost is compared with benefits measured against a pre-launch baseline, using a control group and an agreed rule for turning time saved into money. Most organizations cannot yet show this: in McKinsey's 2026 survey, 80% of respondents said AI had improved their own productivity, but only 37% attributed any EBIT impact to AI.

This guide is part of our AI consulting firms research, written for the executive sponsor who must justify a GenAI program or audit one that is already running.

What does a GenAI program actually cost?

Count six cost lines: build and integration, data preparation, inference, platform and support, evaluation and governance, and change management. The recurring lines grow with usage, which is where budgets slip.

Gartner estimated in October 2024 that CIOs who do not understand how their GenAI costs scale could make a 500% to 1,000% error in their cost calculations; more than 90% of the 300-plus CIOs it surveyed said managing cost limits their ability to get value from AI. In McKinsey's 2026 survey, about 20% of respondents said operating costs, including tokens, had constrained their organization's AI use. In KPMG's Q3 2026 pulse of 314 US leaders at organizations with $1 billion or more in annual revenue, 74% included cost reviews in AI approvals and 43% had usage or token budgets.

Cost lineWhat it includesBehaviorHow to estimate it
Build and integrationEngineering, retrieval or agent design, integrations, security reviewOne-time, front-loadedTeam size × weeks × hours × quoted rate
Data preparationCleaning, document conversion, access metadata, labeled test setsOne-time, then upkeepGive it its own work stream and budget line
InferenceModel tokens, embeddings, rerankingGrows with usageRequests × tokens per request × vendor price
Platform and supportSearch or vector store, hosting, monitoring, on-call, model upgradesRecurringMonthly bill plus the share of an engineer's time
Evaluation and governanceTest sets, red-teaming, audit trail, vendor risk reviewOne-time, then a review cadenceHours for set-up, then a fixed schedule
Change managementTraining time, process redesign, a named internal championMostly one-timeHours × loaded labor cost

Use loaded labor cost, not salary. In June 2026, US Bureau of Labor Statistics data show wages and salaries at 68.5% of employer compensation cost for private-industry management, professional, and related occupations ($54.06 of $78.88 per hour worked), so total compensation was about 1.46 times wages, before facilities, software, and supervision.

Token prices differ sharply between model tiers. As of October 2026, Anthropic's pricing page lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens, a factor of 20 between the two. Whether the smaller model clears your quality bar is a testing question to put in the plan.

What benefits count, and how do you turn them into money?

Three benefit types count: time saved, revenue, and risk reduction. Only time that changes a cost or output line, such as avoided hires, lower overtime, or more units sold, becomes a financial benefit. Freed hours that disappear into slack do not.

In McKinsey's 2026 data, only 14% of respondents at organizations using AI said AI had contributed to a decline in workforce size over the past year, even though 80% reported better individual productivity. Danish researchers who linked adoption surveys to administrative earnings and hours records found no significant effect of chatbot use on either, and 85% of users said they reallocated saved time to other tasks (Humlum and Vestergaard, NBER). In the 2026 Wharton and GBK survey, hours saved was the most-used ROI measure (40% of departments), yet departments tracking at least one growth measure were more likely to report strongly positive returns than those tracking efficiency alone (35% against 13%).

BenefitExample metricsConvert to money byCommon mistake
Time savedHandling time, tickets per agent-hour, drafting timeAvoided hires on volume growth, lower overtime or contractor spend, extra output soldMultiplying freed hours by salary when staffing and output do not change
RevenueConversion, retention, upsell, new-product revenueIncremental margin against a control groupCrediting AI for growth that a pricing or campaign change caused
Risk and qualityError rate, rework, compliance findings, loss eventsFrequency × cost per event, before and afterTreating a few incidents as evidence

How do you set a baseline and attribute results to the system?

Measure the baseline before launch, keep a comparison group, and write metric definitions, data owners, and review dates into the statement of work. Without a counterfactual an ROI figure is an opinion, and people overrate their own gains: in METR's randomized trial, developers believed AI had made them 20% faster while their tasks took 19% longer.

Deloitte's interviews point to a second problem, concurrent change. Executives told researchers that AI usually arrives alongside data-quality work, team changes, or process streamlining, which makes its own contribution hard to isolate. Brynjolfsson, Li, and Raymond could measure a 15% gain in customer support because the tool was introduced in stages, which gave them a comparison group.

  1. Choose one primary metric per use case and write its definition: unit, source system, calculation, owner.
  2. Capture the pre-launch baseline over a period long enough to include normal swings in volume and seasonality.
  3. Roll out in stages or hold back a comparison group, and keep the mix of tickets or tasks comparable.
  4. Track a quality guardrail beside every speed metric (error rate, escalation rate, customer satisfaction), because faster and worse is a loss. List other changes in the same period, such as reorganizations, pricing, and hiring, so they are not credited to AI.
  5. Agree stop rules before launch. In Gartner's October 2026 survey of 2,002 CIOs and technology executives, organizations with criteria for stopping pilots that should not reach production were 2.3 times more likely to be in what Gartner calls the vanguard group (CIOs reporting better business outcomes and high CEO-perceived AI value), yet only 31% of CIOs had such criteria. Just 17% had criteria for retiring low-value systems already in production.
  6. Have finance sign off the final number.

What does a worked ROI estimate look like?

The estimate below is illustrative, not a measured result. It shows that the result depends mostly on two assumptions, the productivity gain and the share of freed time that becomes cash, rather than on the build cost.

DCF Research estimate: an AI assistant for 200 customer-support agents. Every input is an assumption unless a source is named.

  • Agents: 200, each working 1,800 hours a year.
  • Loaded labor cost: $36.73 per hour worked (BLS, June 2026, total compensation for private-industry office and administrative support occupations; excludes facilities, software, and supervision).
  • Labor cost in scope: 200 × 1,800 × $36.73 = $13,222,800 per year.
  • Productivity gain: 8% more tickets per agent-hour, about half the 15% average that Brynjolfsson, Li, and Raymond measured across 5,172 support agents, because this is a first release.
  • Capacity value: 8% × $13,222,800 = $1,057,824 per year.
Cost lineArithmeticAmount
Build and integration (one-time)4 people × 12 weeks × 40 hours × $175 per hour$336,000
Knowledge-base cleanup (one-time)2 × 6 weeks × 40 hours × $175$84,000
Evaluation and security review (one-time)1 × 4 weeks × 40 hours × $175$28,000
Change management (one-time)Training: 200 agents × 6 hours × $36.73 = $44,076; change lead: 20 hours a week × 26 weeks × $78.88 (BLS, management, professional, and related occupations) = $41,018$85,094
One-time total$533,094
Inference (per year)1.44 million tickets (200 × 1,800 × 4 per hour) × 17,500 input and 1,500 output tokens at $2 and $10 per million = $0.05 per ticket$72,000
Search and hosting (per year)$2,000 per month$24,000
Support and upkeep (per year)0.5 engineer × 1,800 hours × $73.69 (BLS, professional and related occupations)$66,321
Run total (per year)$162,321

The $175 hourly rate is an assumed blended consulting rate for illustration, not a market quote; replace it with your own quotes. The benefit depends on how much of the freed capacity becomes cash, so the table below varies that share at the 8% gain.

Share of freed capacity that becomes cashAnnual benefitAnnual net after run costsPayback on one-time costs3-year net
38%$402,000$240,0002.2 years$186,000
60%$635,000$472,0001.1 years$884,000
100%$1,058,000$896,0000.6 years$2,153,000

The 38% row borrows Bain's October 2026 finding that GenAI is realizing 38% of its expected value across functions; it is an anchor, not a forecast. Three-year net is 3 × annual benefit − one-time costs − 3 × run costs, undiscounted, with full benefits from day one. At a 4% gain and 38% conversion, annual net falls to about $39,000 and payback stretches to nearly 14 years. Run costs alone need about 15% of the capacity value to turn into cash, and a three-year payback needs about 32%.

The sponsor's task is to name the hires, overtime, or output lines that will actually change. Deloitte's 2025 respondents typically expected satisfactory ROI within two to four years, so a payback near two years in the 38% case is in line with that.

What do published surveys actually report about GenAI ROI?

Published surveys put the share of respondents reporting value from AI anywhere from 13% to 80%, depending on the threshold and on who answers. Read each figure with its definition, its sample, and whether it covers GenAI or AI in general.

SurveySampleWhat it reportsHow to read it
McKinsey, The state of AI in 2026 (Aug 2026)1,719 respondents, 97 countries, fielded May 4 to June 8, 202637% attribute at least some EBIT impact to AI; about 6% are high performers (5% or more of EBIT, impact described as significant); 80% report individual productivity gainsAI broadly, including agents; self-reported
Gartner CIO survey (Oct 2026)2,002 CIOs and technology executives, May to July 202613% report significant or transformational value from AI so far; likelihood of use-case success is one in fiveCIO view of enterprise-level value
Wharton and GBK, The Adaptive Advantage (Oct 2026)1,007 US managers and above at firms with 1,000+ employees and $50 million+ revenue, fielded July 9 to 29, 202680% report positive returns and 28% strongly positive; hours saved is the most-used KPISelf-reported, often by the people who own the programs
PwC 29th Global CEO Survey (Jan 2026)4,454 CEOs in 95 countries and territories, Sept 30 to Nov 10, 2025In the last 12 months 30% saw higher revenue from AI, 26% lower costs, 12% both, and 56% neitherCEO view; AI broadly
BCG Applied AI Index (Sept 2026)1,330 CxOs and senior leaders7.5% are in BCG's future-built tier and a further 41% are scaling and outperforming; AI spend is about 3.3% of revenueA maturity classification, not a measured ROI
Bain Automation and AI Pathfinder (Oct 2026)951 companies, 63 enterprise processesGenAI realizing 38% of expected value across functions, against 52% for traditional automation; nearly 40% of companies that measured savings landed below 10% when targeting 11% to 20%Closest to a realization rate
Deloitte, AI ROI: the paradox of rising investment and elusive returns (Oct 2025)1,854 senior executives in Europe and the Middle East, Aug 15 to Sept 5, 2025Most expect satisfactory ROI on a typical use case within two to four years; 6% report payback in under a year; 15% of GenAI users report significant, measurable ROIPayback expectations
MIT NANDA, The GenAI Divide (Jul 2025, preliminary)52 organization interviews, 153 survey responses, 300+ public initiativesHeadline: 95% of organizations getting zero return. Its own funnel: 60% evaluated enterprise-grade tools, 20% piloted, 5% reached productionInterview-based, with success judged six months after pilot

What to take from the spread:

  • Definitions drive the numbers. Any EBIT impact (37%), a positive return (80%), significant or transformational value (13%), and a strongly positive return (28%) are different thresholds. GBK, which ran the Wharton survey, notes that self-reported ROI tends to be generous because respondents often grade programs they own.
  • BCG's 2025 report of 1,250 executives found 5% of companies achieving AI value at scale and 60% reporting minimal gains. Its 2026 index describes a shift to nearly half capturing meaningful value, but the samples and tiers differ, so compare with care.
  • Treat the MIT figure as a preliminary study. The authors call their numbers directionally accurate, and simple arithmetic on their funnel (5% ÷ 20%) gives about one in four organizations that reached pilot going on to production, not one in twenty.
  • Practices are the most consistent finding. McKinsey reports that nearly three-quarters of high performers fundamentally redesigned workflows, against about one-quarter of other respondents, and that high performers were about twice as likely to have defined processes to quantify AI impact. In the Wharton data, 67% of leaders whose department met all three transformation conditions reported strongly positive ROI, against 19% of those that did not.

What do task-level studies say about time saved?

Controlled studies show real gains on some tasks, no gain on others, and losses when tools are used beyond what they do well. Test a time-saved assumption in your own workflow before it goes into a business case.

StudySettingResult
Brynjolfsson, Li, Raymond, Quarterly Journal of Economics (2025)5,172 customer-support agents at one firm; tool introduced in stages15% more issues resolved per hour on average; about 30% for less experienced agents; little effect for the most experienced
Dell'Acqua et al., Organization Science (2026)758 knowledge workers in a study run with Boston Consulting Group; 18 realistic tasks plus one task outside the tool's abilityWithin its ability: 12.2% more tasks completed, 25.1% faster, higher quality. On the task outside it: 19% less likely to be correct
METR (2025)16 experienced open-source developers, 246 tasks in their own repositories, early-2025 toolsTasks took 19% longer with AI; developers had expected a 24% speedup and afterwards believed they had been 20% faster
Humlum and Vestergaard, NBER (2025, revised 2026)Danish adoption surveys linked to administrative earnings and hours recordsNo significant effect on earnings or recorded hours, ruling out effects larger than 2%; 85% of chatbot users reallocated saved time to other tasks

METR's February 2026 update says its newer data are unreliable because many developers declined to work without AI, so the 2025 slowdown is a dated snapshot. The lesson for a buyer is the gap between perceived and measured time, not the sign of the effect.

How do you structure an engagement so ROI can be audited?

Make measurement a contract deliverable: baseline, metric definitions, evaluation set, run-cost forecast, stop rules, and hand-over, each with an owner and a date. Consider tying part of the final payment to the measured result rather than to go-live.

  1. A measurement plan on the first page of the statement of work: metrics, baselines, targets, owners, and review dates, including a named internal owner with time allocated.
  2. Separate budget lines for data preparation, evaluation, and change management.
  3. A monthly run-cost forecast at pilot and at target volume, with token assumptions shown.
  4. Phase gates with written exit criteria and the right to stop.
  5. A production plan covering security review, compliance sign-off, monitoring, and support ownership, so a prototype does not end as a demo.
  6. A hand-over plan the internal team can execute, or a managed-service option priced separately.

Strong proposals show a measurement plan from a comparable engagement and what was measured afterward. Weak ones lead with tool certifications, define success as a delivered system, skip adoption, assume the data is ready, or promise a return out of proportion to the cost. For pricing models see AI consulting pricing; for phasing see AI strategy vs. implementation consulting and GenAI consulting proof of concept.

Sources

More in AI Consulting

← Back to Insights