How do you measure ROI on AI?

Measure AI ROI the same way you measure any investment: (return − cost) ÷ cost. The hard part is defining 'return' honestly. Start with a baseline of the metric you want to move (cost, revenue, time, error rate) before deploying. Then track that same metric after deployment, attributing only the change you can reasonably link to the AI. Include the full cost — licences, integration, staff time, and change management — not just the software fee. If you can't tie the AI to a baseline metric, you can't claim ROI.

Why do most AI pilots show no ROI?

Because they were never set up to be measured. MIT's NANDA report (State of AI in Business 2025) found 95% of enterprise generative-AI pilots show no measurable P&L impact. The most common reason isn't that the AI didn't work — it's that no baseline was captured, no success metric was defined, and no one could attribute a financial change to the tool. A pilot with no measurement plan defaults to 'no measurable return,' regardless of what it actually did.

What metrics should I track for AI ROI?

Track a small number of metrics with a clear money link: time saved per task (× loaded hourly cost), cost reduction (fewer hours, lower error/rework rate, reduced outsourcing), revenue impact (higher conversion, faster response time, more capacity served), and quality (error rate, customer satisfaction). Pick one or two primary metrics per use case. Vanity metrics — 'number of AI queries run' — do not prove ROI and should be ignored.

How long before AI shows ROI?

It depends on the use case, but you should define the measurement window up front. Simple automations (scheduling, drafting, routing) can show time and cost savings within weeks. Larger workflow changes take a quarter or more because the return is gated by adoption — staff have to actually change how they work. Per BCG's 10/20/70 framing, roughly 70% of the outcome depends on people and process, which is also the part that takes longest to land.

Should I include change-management cost in AI ROI?

Yes. Leaving it out is the most common way AI ROI gets overstated. The software licence is often the smallest line. The real costs are integration, data cleanup, staff training, and the time spent redesigning workflows and driving adoption. Per BCG's 10/20/70, the people-and-process work is around 70% of the effort — so excluding it makes your ROI look far better than it is and sets up a nasty surprise later.

AI ROI: How to Measure the Return on AI in Your Business

A practical framework for measuring AI ROI — tied to why most enterprise pilots show no P&L impact. Baseline, attribute, and track the metrics that actually prove return.

AI ROI is measured the same way as any other investment — (return minus cost) divided by cost — but the reason most AI projects can't prove a return isn't the math. It's that they were never set up to be measured. MIT's NANDA report found 95% of enterprise generative-AI pilots show no measurable P&L impact (State of AI in Business 2025, August 2025). For a large share of those, the problem wasn't the technology — it was the absence of a baseline, a defined metric, and honest cost accounting.

This is the practical framework to avoid that: baseline, attribute, track, decide.

Why "No Measurable ROI" Is Usually a Measurement Failure

Read the MIT finding carefully. The bar is "no measurable P&L impact." A pilot can do useful work and still land in that 95% if nobody captured what changed.

Three failures put projects there:

No baseline. Nobody recorded the cost, time or conversion rate before the AI. So there's nothing to compare to.
No defined success metric. "Improve productivity" can't be proven. "Cut quote-drafting time from 40 minutes to 10" can.
No attribution. A number moved, but it's impossible to say the AI moved it versus seasonality or other changes.

Fix these three and you've already separated yourself from the majority of failed pilots — before touching the model.

The AI ROI Formula

ROI = (Return − Cost) ÷ Cost

Simple. The discipline is in defining both sides honestly.

Return = the measured change in a baseline metric you can attribute to the AI, converted to money: time saved × loaded hourly cost, cost avoided, or incremental revenue.

Cost = everything, not just the licence:

Cost line	Often forgotten?
Software / licence / API usage	No
Integration and setup	Sometimes
Data cleanup and preparation	Often
Staff time to learn and adopt	Almost always
Workflow redesign and change management	Almost always

That last block is the killer. Per BCG's 10/20/70 rule, roughly 70% of AI effort is people and process. If you cost only the software, your ROI looks great on paper and collapses in reality.

The Metrics That Actually Prove Return

Pick one or two primary metrics per use case. The four categories that carry a money link:

Time saved — minutes per task × volume × loaded hourly cost. The most reliable starting point for service businesses.
Cost reduced — fewer hours, lower error/rework rate, reduced outsourcing or overtime.
Revenue impact — higher conversion, faster response time (which lifts close rates), more capacity served without adding headcount.
Quality — error rate, rework rate, customer satisfaction.

Ignore vanity metrics. "We ran 4,000 AI queries this month" proves activity, not return. Activity is not ROI.

A Worked Example (UAE Service Business)

A Dubai clinic deploys an AI tool to draft patient follow-up messages and handle first-line scheduling queries.

Baseline (before):

Reception spends ~3 hours/day on follow-ups and routine queries.
Loaded cost of that time: ~AED 80/hour → ~AED 5,280/month.

After (measured over 8 weeks):

Time on those tasks drops to ~1 hour/day → ~AED 1,760/month.
Measured monthly saving: ~AED 3,520.

Full cost:

Tool + integration: ~AED 1,200/month equivalent.
One-off setup, training, workflow redesign: ~AED 6,000 (amortised).

First-quarter ROI: positive once the one-off cost is recovered — and, crucially, provable, because a baseline existed. (Figures illustrative; the point is the method, not the numbers.)

Note what made this work: a single use case, a captured baseline, full costing including training, and an attributable metric. That is the entire game.

Tie ROI to the Adoption Problem

ROI on AI is gated by adoption. A tool nobody uses returns nothing regardless of how good it is. This is why the timeline matters: simple automations can show savings in weeks; workflow changes take a quarter or more because staff have to actually change how they work — the 70% in BCG's 10/20/70.

So bake adoption into measurement: track usage rate alongside the financial metric. If usage is low, your ROI problem is a change-management problem, not a technology one — and adding a better model won't fix it.

The Five-Step Measurement Plan

Pick one P&L-linked use case.
Capture a baseline before deploying.
Cost it fully — including training and change management.
Track the same metric post-launch over a defined window, attributing conservatively.
Calculate and decide — scale if positive and adopted; trigger the kill-switch if not.

Write the success metric and the kill-switch before you start. A project with no defined target cannot demonstrate ROI — which is precisely how pilots end up in the 95%.

How This Connects to Failure Rates

Measurement is the difference between the 5% that succeed and the 95% that don't show impact. Our companion post, Why Do AI Projects Fail?, breaks down the failure statistics and their sources; 70/20/10 for AI explains why people and process dominate the outcome. For the full delivery model, see the AI Consulting in Dubai pillar guide.

To build a measurement plan for a specific AI initiative before you commit budget, book an AI consultation via evolvxai.com, or read more about how we work.

Sources

MIT NANDA — The GenAI Divide: State of AI in Business 2025 (Aug 18 2025), 95% of enterprise GenAI pilots no measurable P&L return: finance.yahoo.com; report mirror mlq.ai
BCG / Ernesto Pagano — 10/20/70 people-centric AI (70% people and process): pmi.org/blog/ai-transformation-people-insights-bcg