Enterprise AI ROI: How to Measure Whether AI Actually Creates Value

Enterprise AI ROI is the value AI creates, in extra revenue, lower costs, faster work and lower risk, minus everything it costs to run, measured against a baseline. Most companies can't show it yet: in McKinsey's 2026 survey, only 37% reported any EBIT impact from AI. The companies that can show it measure outcomes, not activity, from day one.

This guide explains what counts as AI value, what AI really costs, what the 2025–2026 research shows — including the statistics that are usually misquoted — and how to measure ROI properly, with a worked example and a scorecard. It's written for leaders who have to decide which AI work to fund next, and which to stop.

Key takeaways

• Most organisations can't yet show AI returns. McKinsey found 37% attribute any EBIT impact to AI; PwC found 56% of 4,454 CEOs had seen neither revenue nor cost benefits in the past year.
• Payback takes years, not months. In Deloitte's 2025 survey, satisfactory ROI took two to four years; only 6% paid back within a year.
• Time saved isn't money saved. A study of 25,000 Danish workers ruled out effects on earnings or hours larger than 2%; most users spent saved time on other tasks.
• The winners redesign work and measure it. Nearly three-quarters of McKinsey's AI high performers had fundamentally redesigned workflows, and they were twice as likely to have defined ways to measure impact.
• Count the full cost. Tokens, integration, data work, change management, oversight and rework — about one in five McKinsey respondents said operating costs were already limiting their AI use.

What is enterprise AI ROI?

Enterprise AI ROI is the return an organisation earns from AI, measured the way any investment is measured: benefits minus costs, divided by costs, over a defined period — with benefits proven against a baseline rather than assumed. The hard part isn't the formula; it's agreeing what counts as a benefit and counting every cost.

Definition

AI ROI = (measured value created by AI − total cost of AI) ÷ total cost of AI, over a set period. Measured value means outcomes compared with a baseline or control group — revenue, avoided cost, redeployed capacity, reduced risk — not usage, activity or self-reported time savings.

AI creates value in five ways, and each needs a different measure:

Value type	What it looks like	How to measure it
Revenue	Higher conversion, faster sales cycles, new AI-enabled services	Revenue per lead or customer against a control group
Cost	Less outsourcing, fewer agency hours, lower cost per transaction	Cost per outcome before and after; invoices actually reduced
Capacity	More work handled by the same team	Volume per person, with quality held constant
Risk	Fewer errors, faster detection, better compliance	Error and incident rates; losses avoided
Speed and quality	Shorter cycle times, fewer defects, higher satisfaction	Cycle time end to end; defect and satisfaction scores

Why is AI ROI so hard to show?

AI ROI is hard to show because most organisations add AI to existing work, measure activity instead of outcomes, and underestimate both the time and the cost involved. The large surveys agree on the gap:

• McKinsey (2026, 1,719 respondents): 37% attribute at least some EBIT impact to AI — about the same as a year earlier — and high performers, who attribute 5% or more of EBIT to AI, stay at about 6%.
• PwC (2026, 4,454 CEOs): 30% saw more revenue from AI in the past year and 26% lower costs, but 56% saw neither.
• IBM (2025, 2,000 CEOs): only a quarter of AI initiatives had delivered their expected ROI.
• BCG (2025, 1,250 companies): about 5% are getting value at scale; 60% report minimal value.
• Deloitte (2025, 1,854 executives): satisfactory ROI typically takes two to four years; 6% saw payback within a year.

Meanwhile adoption is not the constraint. Stanford's AI Index 2026 reports the UAE leads the world in AI adoption, with 64% of its population using AI in the second half of 2025. Usage is everywhere; measured value isn't yet.

Common misconception

"95% of AI projects fail." The figure comes from MIT NANDA's 2025 report, whose authors judged that about 95% of organisations saw no measurable profit-and-loss return from generative AI, roughly six months after a pilot, based on 52 interviews, 153 conference surveys and reviews of public case studies. The authors call the results directional. Similar "85%" and "80%" failure claims trace to a 2018 Gartner forecast and a magazine article. None is a measured failure rate — but all point at the same problem: pilots without measurement.

Does productivity from AI turn into money?

Productivity from AI turns into money only when the time it frees is redeployed to work the business values — more sales, more customers served, less outsourcing — or when it avoids a cost you'd otherwise pay. The research shows real gains at task level, and much smaller effects at company and labour-market level so far.

Study	What it found	What it means for ROI
Customer support (Brynjolfsson, Li and Raymond, QJE 2025)	About 15% more issues resolved per hour, most for newer agents	Real gains where tasks are frequent and measurable
BCG consultants (Dell'Acqua et al., 2026)	On tasks suited to AI: 12.2% more tasks, 25.1% faster, higher quality; on a task outside AI's strengths, 19 percentage points less likely to be right	Gains depend on matching AI to the right tasks
Experienced developers (METR, 2025)	19% slower with AI, while believing they were about 20% faster	Self-reported time savings are unreliable
Denmark, 25,000 workers (NBER, 2026 revision)	No measurable effect on earnings or hours, ruling out more than 2%; 85% of users reallocated saved time to other tasks	Time saved gets absorbed unless work is redesigned
Software delivery (DORA, 2025)	Throughput up with AI adoption, but delivery stability still down	Faster steps can create problems downstream

METR's own follow-up in 2026 couldn't confirm the slowdown and called its original design unreliable — which is the real lesson: measure outcomes with a comparison group, because impressions mislead in both directions. Gartner's 2026 sales survey makes the redeployment point directly: AI saved sellers 4.8 hours a week, but 72% of sales organisations reinvested little of that time, and those that did were 3.1 times likelier to beat lead-to-opportunity targets.

What does AI actually cost?

AI costs far more than its licences. The total includes usage and tokens, integration with existing systems, data preparation, change management and training, human oversight and quality checks, and the rework and incidents that come with any new system. Leaving any of these out inflates ROI on paper and disappoints in practice.

Cost	What it covers	Why it's often missed
Licences and usage	Seats, API calls, tokens	Usage grows with success; BCG notes frontier models cost 5 to 25 times more per token than simple ones
Integration	Connecting AI to CRMs, ERPs, data and identity systems	Usually the largest build cost
Data work	Cleaning, labelling, access controls	Done once badly, then again properly
Change management	Training, process redesign, communications	Treated as "soft" and unbudgeted
Oversight and testing	Human review, evaluation suites, monitoring	Ongoing, not one-off
Rework and incidents	Fixing errors, handling complaints, rollbacks	Invisible until something goes wrong
Liability	Legal exposure for what the AI says or does	Real — a tribunal held Air Canada liable for its chatbot's wrong advice in 2024

Token costs deserve a line of their own. BCG estimates a conversation twice as long costs about four times as much, because the model re-reads the growing context, and about one in five McKinsey respondents say AI operating costs, including tokens, already constrain their AI use. Agents that loop, retry and reason at length can make the cost per outcome, not the cost per call, the number to watch.

How do you measure AI ROI properly?

Measure AI ROI the way you'd test any business change: set a baseline, run the AI on part of the work while a comparable part continues without it, measure outcomes on both, count the full cost, then decide to scale, fix or stop. Repeat the measurement after scaling, because results change when volume and users change.

Measuring AI ROI

1. 01Baseline
   • Current outcomes
   • Current costs
   • Cycle times

   Measure before anything changes

2. 02Compare
   • Pilot group
   • Control group
   • Staggered rollout

   Isolate the AI's effect from everything else

3. 03Measure outcomes
   • Revenue
   • Cost per outcome
   • Quality
   • Risk

   Outcomes, not usage or self-reports

4. 04Count full cost
   • Usage
   • Integration
   • Oversight
   • Rework

   Including the people keeping it running

5. 05Decide
   • Scale
   • Fix
   • Stop

   Stopping is a valid, valuable result

6. 06Re-measure
   • Quarterly
   • After scaling

   Returns drift as usage and costs change

Four rules keep the measurement honest:

1. Measure the whole workflow, not the step. If AI drafts contracts twice as fast but legal review becomes the bottleneck, cycle time doesn't improve. AI workflow redesign covers how to find and fix those constraints.
2. Prefer comparisons to before-and-after. Markets, seasons and teams change; a control group separates the AI's effect from everything else.
3. Price time in money only when it's redeployed. An hour saved is worth an hour of the work it's moved to — or nothing, if it disappears into email.
4. Give it the right horizon. Judge pilots on leading indicators, programmes on lagging ones, over two to three years.

A worked example: AI in a customer enquiry team

A hypothetical 20-person enquiry team handles 10,000 customer enquiries a month at a fully loaded cost of AED 25 per enquiry, or AED 250,000 a month. The numbers below are illustrative — replace every one with your own baseline.

Item	Before AI	With AI (after six months)
Enquiries a month	10,000	10,000
Resolved by AI end to end	0	3,500 (35%)
Handled by people	10,000	6,500
People needed at the same workload	20	13
Monthly people cost	AED 250,000	AED 162,500
Monthly AI cost (usage, platform, monitoring)	—	AED 30,000
Monthly oversight and quality review	—	AED 15,000
One-off build and integration	—	AED 360,000, spread over 36 months: AED 10,000
Total monthly cost	AED 250,000	AED 217,500

The gross saving is AED 32,500 a month — about 13% of the original cost, and roughly AED 390,000 a year. The return only becomes real in one of two ways: the seven freed roles are redeployed to work that earns money (outbound sales, account management) or the team absorbs growth without hiring. If the freed capacity simply disappears into slower days, the measured ROI is zero, however good the AI is.

Two adjustments complete the picture. Risk-adjust the benefit for the share of AI resolutions that later need rework or cause complaints. And track quality — customer satisfaction and repeat contact rates — alongside cost, because a cheaper service that loses customers isn't a saving.

What do companies that get AI ROI do differently?

Companies that get AI ROI redesign work around AI, commit leadership attention and budget, define how they'll measure impact before they start, and keep successful systems running long enough to pay back. The research on high performers is consistent on each point.

• They redesign workflows. Nearly three-quarters of McKinsey's 2026 high performers had fundamentally redesigned workflows because of AI, up from 55% the year before.
• They measure. High performers were twice as likely to have defined processes to measure the impact of AI initiatives; Gartner found 63% of high-maturity organisations implement metrics.
• They commit. High performers were more than twice as likely to spend over 15% of their technology budget on AI, and twice as likely to report visible senior-leadership commitment.
• They persist. In Gartner's 2025 survey, 45% of high-maturity organisations kept AI projects in production for three years or more, against 20% of low-maturity ones.
• They pick core processes. BCG estimates about 70% of AI's potential value sits in core business functions rather than support activities.

What this means

The pattern is the same one every earlier technology followed: tools produce potential; process change produces returns. The companies ahead on AI aren't the ones with the most pilots. They're the ones that chose a few core workflows, rebuilt them, and measured the result honestly.

An AI ROI scorecard to use

Use a scorecard per use case, reviewed quarterly, with leading indicators that show whether value is coming and lagging ones that prove it arrived.

Dimension	Leading indicators	Lagging indicators
Adoption	Active users, share of eligible work routed through AI	Work volume handled per person
Quality	Accuracy on test sets, human override rate	Complaints, rework, customer satisfaction
Speed	Cycle time of the AI step	End-to-end cycle time
Financial	Cost per outcome	Revenue, avoided costs, redeployed capacity
Risk	Policy violations caught, incidents logged	Losses, fines, legal claims
Cost	Usage and token spend per outcome	Total cost of ownership against plan

For the risk dimension, established frameworks help: the NIST AI Risk Management Framework organises the work into governing, mapping, measuring and managing AI risk, and the AI governance guide shows how controls fit around agents. For the use cases worth measuring first, the executive guide to AI agents covers how to choose them.

The most useful number on the scorecard is often the one that tells you to stop. A programme that measures honestly will retire some use cases — and fund the ones that work with the money it saves.