How to measure AI ROI in an enterprise: baseline 3 to 5 KPIs before the AI touches the workflow, capture 8 to 12 weeks of time, cost, volume and error data, convert every KPI movement into cash, then let payback and NPV decide whether the next tranche gets funded. Most AI projects do not die in production. They die in a board paper that says "transform the business" where a number should be. McKinsey's latest GCC survey found that 84% of organisations use AI but only 11% are value realisers, meaning they scale deployment and attribute at least 5% of earnings to AI. The gap is measurement, not technology. Here is the method: a 7-step stage-gating framework, the financial formulas your finance team already uses, and a worked UAE example with numbers you can adapt this quarter.
Why do AI business cases die at the funding gate?
The pattern is consistent. A CTO presents a pilot that cut document handling time by 40%, the board nods, and then the CFO asks one question: "What did that add to earnings, and what does the next phase cost?" Silence follows, and the second tranche goes to a project with a spreadsheet.
The GCC data makes the gap concrete. McKinsey's survey of 139 senior executives and board directors, conducted in August and September 2025 in partnership with the GCC Board Directors Institute, found that 84% of organisations had adopted AI in at least one business function, up from 62% in 2023, but only 31% had scaled beyond pilots and only 11% qualified as value realisers: organisations that scale deployment and attribute at least 5% of earnings to AI. Widely reported MIT research from 2025 found that around 95% of enterprise generative AI pilots produced no measurable impact on profit and loss. Whether that is a technology failure or a measurement failure is arguable. Either way, it is a business case failure.
Three things separate the 11%. In the same survey, value realisers were markedly stronger on talent and operating model, on technology and data fundamentals, and on change management. They tied deployment to measurable KPIs, and one interviewee described requiring the foreseen financial benefit to appear in profit and loss forecasts before a use case was approved. They did the boring accounting work first.
That matters to you directly. If you are a CTO or head of engineering asking a board to green-light a second or third AI project, you are not competing with other AI projects. You are competing with a warehouse expansion, a new branch and a hiring plan, all of which arrive with payback periods attached. Yours needs one too.
How do you measure AI ROI in an enterprise?
Before the steps, fix the vocabulary. ISACA's AI ROI model evaluates AI investments across three categories. Most business cases present only the first one, and badly:
| Category | What it captures | Time horizon | Where it appears in the board pack |
|---|---|---|---|
| Measurable ROI | Cost savings, revenue increases, productivity gains | 6 to 24 months | The NPV, payback and IRR table |
| Strategic ROI | Progress towards long-term goals such as digital transformation and competitive position | 3 to 5 years | The strategy slide, with a named milestone |
| Capability ROI | Skills, new roles, and the data and platform maturity the project leaves behind | Ongoing | The closing slide, as an explicit by-product |
A case that claims strategic ROI but shows no measurable ROI will not fund tranche two. A case that shows measurable ROI but leaves no capability behind makes every later project more expensive. The framework below produces all three, but it makes the measurable category carry the funding decision.
For the financial maths, use the metrics your finance team already trusts. You.com's enterprise AI ROI guide sets out the chain plainly: pick 3 to 5 KPIs, benchmark the current state over the last 8 to 12 weeks, set a future-state target, measure weekly against that baseline, convert KPI movement into money (hours saved multiplied by hourly cost, conversion lift multiplied by traffic and order value, incidents avoided multiplied by average incident cost), then compute payback, net present value (NPV), internal rate of return (IRR) and simple ROI. Lead with NPV. It is the number CFOs use to green-light projects.
What do the 7 steps look like, from workflow map to funding gate?
This is IBM's 7-step stage-gating framework, adapted with the financial instrumentation above. The core idea: rather than funding a portfolio of AI solutions at once, fund one capability, prove it against KPIs, then let the measured savings or revenue unlock the next tranche of work.
- Map workflows and workflow segments. List your high-level workflows (for an airline: ticketing, ground operations, onboard experience). Break each into discrete segments, then generate AI use cases per segment with domain experts. Those use cases become your candidate capabilities.
- Prioritise five candidates. Pick five workflow segments and run high-level financial projections on each. IBM's standard is a projection that is 50 to 60% accurate, which means you must know the revenue and cost levers in each segment before you model.
- Select the MVP. Plot the five on impact (vertical) against feasibility (horizontal). Weigh reach (a small, engaged team versus the whole company) and synergies (shared users, shared integration patterns). For a first capability, financial impact is the crucial factor, because its savings fund the next stage.
- Define KPIs and the experiment plan. How many sites, users or transactions? Is the sample large enough for a statistically significant read? How will you isolate the AI's effect from everything else changing? How long must you run?
- Fund and execute the MVP. Take it through your governance process with the measurement plan attached. IBM cites PepsiCo's governance board, which assesses, validates and approves use cases against responsible AI principles before funding.
- Measure against baseline. Track the KPIs from step 4 from the first week of rollout, not from month three.
- Gate. If KPIs met expectations, fund full-scale implementation and open the next tranche. If not, diagnose: adoption failure (fix with change management), partial KPI movement (re-scope or redefine KPIs), or technology failure (get vendor support or change solution). Stop. Do not quietly extend the pilot.
Steps 1 to 3 produce the pipeline. Steps 4 to 6 produce the evidence. Step 7 is where funding actually gets decided.
Worked example: what does AI ROI look like at a Dubai insurer?
Take an illustrative regional scenario and run it end to end. A Dubai motor insurer processes 2,400 claims a month with a 12-person intake team. Handling one claim (document checking, data entry, triage routing) takes 45 minutes on average. The business wants to grow volume without growing headcount.
Steps 1 and 2: map and shortlist five. The workflow is claims. Its segments are first notification of loss (FNOL), document intake, triage, assessment and settlement. Five candidate capabilities with rough projections:
| Candidate capability | Estimated annual impact | Feasibility | Key risk |
|---|---|---|---|
| Document extraction and triage assistant | AED 210,000 | High | Extraction accuracy on Arabic scans |
| FNOL chat in Arabic and English | AED 150,000 | Medium | Call transcripts needed for training |
| Fraud signal scoring | AED 400,000 claimed | Low | Little labelled fraud data |
| Renewal churn prediction | AED 250,000 | Medium | Customer data quality |
| Policy Q&A for brokers | AED 90,000 | High | Small ceiling on impact |
The projections are the 50 to 60% accurate kind. They are for ranking, not for the accounts.
Step 3: select. Document extraction and triage wins on impact multiplied by feasibility, has a clear measurable KPI (handling time) and shares its integration pattern (document ingestion) with the later FNOL candidate. Fraud signal scoring is the biggest claimed number and the worst first project, because it cannot pass a gate you cannot measure.
Steps 4, 5 and 6: baseline, instrument, run. The team captures 10 weeks of baseline before go-live, inside the 8 to 12 week window the You.com guide recommends:
| KPI | Baseline | Target | Floor (stop below) | Data source |
|---|---|---|---|---|
| Handling time per claim | 45 min | 32 min | 37 min | Claims system timestamps |
| Rework rate | 8% | 4% | 6% | QA audit sample (200 claims a month) |
| Cycle time, intake to triage | 3 days | 1.5 days | 2.2 days | Claims system |
| Assistant adoption (% of claims touched) | n/a | 85% | 70% | Assistant usage logs |
| Human override rate | n/a | 5% or less | 8% | Assistant usage logs |
Turning KPIs into money. Start with a defensible payroll assumption: AED 6,250 a month fully loaded for an intake specialist, which is base pay plus visa, medical insurance and end-of-service provision, or about AED 36 an hour at 176 working hours a month. Substitute your own numbers; the method is what matters.
With the target met, handling time falls by 13 minutes per claim. At 2,400 claims a month that is 520 hours released a month, 6,240 hours a year, roughly 3 full-time equivalents. Now apply the rule that saves most ROI cases from the bin: hours only count when you name where they go.
- Volume is forecast to grow 15% over the next 12 months (360 extra claims a month). The released hours absorb that growth without two hires, at AED 75,000 a year fully loaded each: AED 150,000 a year.
- Rework falls from 8% to 4% on 28,800 claims a year: 1,152 rework cases avoided at 30 minutes each and AED 36 an hour is about AED 21,000 a year.
- An overtime budget of AED 8,000 a month falls 40%: AED 38,400 a year.
Total measured benefit: about AED 209,400 a year. That is the number that goes in the business case, not 6,240 hours. The remaining capacity covers faster cycle times and peak-season buffer; leave it out of the cash case.
The cost side. Build (two engineers for three months, integration, evaluation harness): AED 200,000 one-off. Run: AED 65,000 a year for monitoring, support and a part-time product owner. Inference is the smallest line, and you should say so. At 25,000 input tokens and 1,500 output tokens a claim, 2,400 claims a month is 60 million input and 3.6 million output tokens a month. On OpenAI's published pricing, gpt-5-mini at $0.25 per million input tokens and $2 per million output tokens costs about $22 a month; the larger gpt-5 ($1.25 and $10 per million) costs about $111 a month, roughly AED 5,000 a year. The money is in people and integration, not tokens.
Step 7: the financial gate. Assume go-live in month 5 and run the numbers:
def payback_month(build, monthly_benefit, monthly_run, go_live_month):
cumulative = -build
for month in range(1, 37):
if month >= go_live_month:
cumulative += monthly_benefit - monthly_run
if cumulative >= 0:
return month
return None
def npv(rate_annual, cashflows):
return sum(cf / (1 + rate_annual) ** t for t, cf in enumerate(cashflows))
build = 200_000 # AED one-off
benefit = 209_400 / 12 # AED a month, from named redeployments
run = 65_000 / 12 # AED a month
print(payback_month(build, benefit, run, go_live_month=5))
# 21
# Cashflows: year 0 build, then 8 live months in year 1, full years after
print(npv(0.10, [-200_000, 96_267, 144_400, 144_400]))
# 115343.8... AED
Base case: payback in month 21, three-year NPV of about AED 115,000 at a 10% discount rate. Now model the sensitivities, because a CFO will:
| Case | Annual benefit | Build cost | Payback | 3-year NPV at 10% |
|---|---|---|---|---|
| Conservative (20% less benefit, build overruns) | AED 167,500 | AED 240,000 | ~33 months | about AED 16,000 negative |
| Base | AED 209,400 | AED 200,000 | ~21 months | ~AED 115,000 |
| Optimistic (25% more benefit, build on plan) | AED 261,750 | AED 200,000 | ~17 months | ~AED 230,000 |
Presenting the conservative case as slightly negative NPV is not weakness. It tells the board exactly which assumption decides the outcome (benefit realisation), which is what the gate in the next section is designed to test.
How do you instrument AI ROI before launch?
Two rules from the sources bear repeating, because they are where measurement usually breaks.
First, capture the baseline before the AI touches anything: time, cost, volume, error rate, 8 to 12 weeks of it. Retroactive baselines are possible but weaker, and a claims system's timestamp history is usually enough if you start now. Decide in advance where reclaimed hours go and set measurable targets for them (more claims absorbed, faster cycles, more proposals a month), or the CFO will rightly discount your time-savings figure.
Second, make the gate enforceable in config, not in a slide. Something like this in your deployment repo keeps the argument honest:
gate:
kpi: handling_time_minutes
baseline: 45
target: 32
minimum: 37 # stop and diagnose below this
window: 8_weeks
sample: 2400_claims
decision:
pass: fund_full_rollout_and_next_tranche
fail: change_management_review
kill: adoption_below_70_pct_for_4_weeks
One more regional caution. McKinsey found that 53% of GCC respondents named output inaccuracy a top barrier to AI adoption. That argues for human-in-the-loop design (in our example, a 5% override target) and for putting override and correction rates in the KPI table from day one, not adding them later.
How do you set the gate: pass, iterate or stop?
The gate meeting is short, and the decision is one of three, agreed before launch:
- Pass. KPIs at or above target for the full measurement window. Fund full rollout and open the next tranche. In our example, extend the assistant from intake into assessment.
- Iterate. Some KPIs moved, some did not. One re-scope cycle with a fixed budget and a new gate date. Adoption between 70% and 85% is the classic iterate signal: the technology works, the workflow does not yet.
- Stop. KPIs below floor, or adoption below the kill threshold. Close the project, write up what you learned (that is capability ROI, and it counts), and redirect the budget.
Add the two non-financial columns from the ISACA model to the gate scorecard so they are on record. Strategic: did this move a named 3 to 5 year goal? Capability: what did we leave behind (evaluation harness, prompt library, trained team, cleaner data)? An initiative that leaves no capability behind has to earn its funding twice over on the next project.
The discipline compounds. Each verified win funds the next small bet, and the conversation shifts from "should we invest in AI?" to "where else will this work?" That is the posture of the 11% in the McKinsey survey.
What should you do next?
If you are preparing a second or third AI business case for a GCC board, do these five things this month:
- Map your top three workflows into segments and shortlist five candidate capabilities with 50 to 60% accurate projections.
- Choose the MVP on impact multiplied by feasibility, and name where every released hour goes.
- Start a 10-week baseline on 3 to 5 KPIs now, before vendor selection is final.
- Build the one-page case: objective, KPI table, costs (build, run, inference), NPV, payback, sensitivities and kill criteria.
- Write the gate definition into the deployment repo before go-live, not after.
Mapping how your business runs and saying plainly where AI will pay off, what it takes and what to do first is what our AI strategy and readiness engagements do. We are a UAE company based in Dubai, and the same team that does the consulting does the deployment and the support. If you want a second pair of eyes on the numbers before they go to your board, start there.
Originally published on Azrty.
Top comments (0)