According to Deloitte's 2025 State of AI survey, only 29% of executives can confidently measure the return on their AI investments. The other 71% are making multi-million-pound decisions on faith, gut feeling, or competitor pressure. This is not acceptable for any other category of capital expenditure, and it should not be acceptable for AI.
The difficulty is real. AI's benefits are often diffuse, delayed, and distributed across multiple business functions. A customer service AI that reduces call handling time also improves customer satisfaction, reduces churn, and frees agents for higher-value interactions. Capturing the full return requires a more sophisticated framework than simple cost savings divided by investment.
This article provides that framework — built from practical experience deploying technology at scale, not from theoretical models.
The standard ROI formula — (Net Benefit / Total Cost) x 100 — works well for investments with clearly bounded costs and easily quantified benefits. A new manufacturing line produces measurable units at measurable cost. AI investments rarely work this way.
AI costs are layered: infrastructure, data preparation, model development, integration, training, ongoing maintenance, and opportunity costs. Benefits are similarly layered: direct cost savings, indirect efficiency gains, quality improvements, risk reduction, and strategic optionality that may not materialise for years.
Applying a simple ROI calculation to this complexity either understates returns (by capturing only direct cost savings) or overstates them (by including speculative future benefits without discounting).
Beware of AI vendor ROI claims that cite only the most favourable metric while ignoring total cost of ownership. A vendor claiming "300% ROI" based on labour savings alone is hiding infrastructure costs, integration effort, maintenance burden, and the opportunity cost of the engineering time required for implementation.
Comprehensive AI ROI measurement requires evaluating returns across four distinct dimensions. Each has different measurement methods, time horizons, and confidence levels.
This is the most straightforward dimension and where most organisations start — and stop. Cost reduction from AI typically appears in three forms:
Establish a pre-AI baseline for each process targeted. Measure the same metrics 3, 6, and 12 months post-deployment. Calculate the difference, subtract AI-related costs (infrastructure, licensing, maintenance, support), and the remainder is net cost reduction.
| Cost reduction category | Measurement method | Confidence level | Time to measure |
|---|---|---|---|
| Labour hours saved | Time tracking pre/post | High | 3-6 months |
| Error rate reduction | Defect tracking pre/post | High | 6-12 months |
| Resource optimisation | Utilisation metrics pre/post | Medium | 6-12 months |
| Process cycle time | Workflow analytics | High | 3 months |
A mid-market insurance company deployed AI-assisted claims processing. Pre-AI baseline: 45 minutes per claim, 12% error rate. Post-AI (6 months): 18 minutes per claim, 4% error rate. With 50,000 annual claims and a fully loaded processor cost of £35/hour, the direct labour saving was £787,500 annually. Error reduction saved an additional £180,000 in rework costs. Against a total AI investment of £350,000 (including integration and first-year licensing), the direct cost reduction ROI was 177%.
AI-driven revenue growth is harder to isolate but often represents the largest component of total ROI. Revenue impacts typically appear through:
The challenge is attribution. When revenue increases after AI deployment, how much is due to AI versus market conditions, sales effort, or product changes?
The most reliable approach is controlled experimentation: A/B testing where one group receives AI-enhanced treatment and a control group does not. Where experimentation is not possible, use difference-in-differences analysis comparing AI-treated business units against comparable untreated units.
Revenue attribution is the hardest part of AI ROI measurement. Never attribute all revenue improvement to AI without a control group or counterfactual analysis. The most common mistake is conflating correlation (revenue grew after AI deployment) with causation (revenue grew because of AI deployment).
AI-driven risk reduction is a legitimate component of ROI but is frequently excluded because it represents costs avoided rather than costs saved. Three categories deserve measurement:
Calculate risk mitigation ROI by estimating the expected value of losses prevented: probability of adverse event multiplied by financial impact, compared before and after AI deployment.
AI investments create strategic options — the ability to pursue opportunities that would be impossible without AI capability. These include entering new markets, launching new product categories, building proprietary data assets, and creating competitive moats.
Strategic optionality is the most difficult dimension to quantify but often the most valuable. Real options analysis provides a framework: value the AI investment as a call option on future business opportunities, using Black-Scholes or binomial models adapted for strategic decisions.
A practical AI ROI measurement system requires three components:
Before deploying any AI system, document the current state of every metric you plan to track. Without baselines, ROI measurement is impossible. Include process times, error rates, conversion rates, customer satisfaction scores, and cost per transaction.
Decide in advance how you will attribute improvements to AI versus other factors. A/B testing is gold standard. Where not possible, use difference-in-differences or interrupted time series analysis. Document your methodology so stakeholders can evaluate the rigour of your claims.
Present AI ROI across cost reduction, revenue growth, risk mitigation, and strategic optionality — with confidence levels for each. A single ROI number is misleading. A four-dimension view gives the board the context to make informed decisions about continued investment.
Several common mistakes systematically distort AI ROI calculations:
Ignoring total cost of ownership. The licensing fee for an AI tool is typically 20-30% of the total cost. Integration, data preparation, training, change management, and ongoing maintenance account for the rest. Any ROI calculation that uses only the licensing cost as the denominator is overstated.
Measuring too early. The J-curve effect means that AI investments often show negative returns in the first 6-12 months. Measuring ROI at 3 months and concluding "AI doesn't work" is a timing error, not an investment error.
Failing to measure what matters. If an AI system improves decision quality — better hiring decisions, better pricing decisions, better investment decisions — but you only measure process speed, you will miss the most valuable returns.
The Opagio Growth Platform helps organisations track AI-related intangible asset accumulation alongside financial metrics, providing a more complete picture of AI value creation. The AI ROI framework for boards provides additional governance-level guidance.
AI ROI is measurable — but only if you measure the right things. The four-dimension framework (cost reduction, revenue growth, risk mitigation, strategic optionality) captures the full spectrum of AI value creation. Start with baselines, define your attribution methodology, and report with confidence levels. The 71% of executives who cannot measure AI ROI are not facing an impossible problem — they are using the wrong framework.
Ivan Gowan is Founder and CEO of Opagio. He spent 15 years as a senior technology leader at IG Group (LSE: IGG), where he oversaw engineering growth from 4 to 250 people during the company's rise from £300m to £2.7bn market capitalisation. Learn more about the Opagio team.
Only 29% of executives can measure AI ROI confidently. The problem is not that AI fails to deliver value — it is that traditional ROI frameworks were not designed for investments that create intangible assets. Here is the 4-layer framework that connects AI spending to board-level outcomes.
Read more →AI vendors promise dramatic cost reductions, but actual savings depend on total cost of ownership, implementation quality, and measurement rigour. This article provides a practical methodology for measuring real AI-driven operational savings — net of all hidden costs.
Read more →
The SEC has made AI-washing an enforcement priority, and for good reason: billions in capital are being allocated based on AI claims that range from exaggerated to fabricated. After 15 years evaluating technology claims at IG Group, here is the 7-point checklist that separates genuine AI capability from performative marketing.
Read more →Get the latest insights on intangible asset growth and productivity delivered to your inbox.
Book a free consultation to see how the Opagio Growth Platform can help your business.