How to measure the ROI of an artificial intelligence project?

The ROI of an artificial intelligence project is measured by comparing verifiable financial benefits against the total cost of implementation, operation, and governance. The key is not measuring “AI usage,” but impact on EBITDA, productivity, revenue, risk reduction, and operational speed with a baseline and sustained metrics.

Taxonomic definitions

1. ROI of artificial intelligence. A financial indicator that compares the economic benefit attributable to a case of AI usage against its total cost. It should include savings, incremental revenue, loss reduction, risk mitigation, and adoption costs, not just licenses or development costs.

2. Operating baseline. Pre-measure the process before introducing AI. Define the benchmark: average time, cost per transaction, errors, rework, tickets, losses, sales, compliance with SLAs or productivity. Without a baseline, ROI becomes an executive opinion with pretty Excel; classic, but dangerous.

3. Total cost of AI. Sum of initial investment and recurring expenses: data, integration, infrastructure, licenses, models, cybersecurity, monitoring, talent, training, change management, auditing, and governance. Gartner warns that GenAI initiatives can have significant costs, even ranging from $5 to $20 million in some business approaches.

Comparative table

Measurement dimension	What it measures	Associated financial indicator	Risk if it is not measured
Productivity	Saved hours, reduction of manual tasks, cycle speed	Reduction of OPEX, improved margin, freed up capacity	Confusing activity with value
Operational quality	Less errors, rework, deviations or reprocesses	Lower cost without quality loss, fewer penalties, protection of EBITDA	Automating errors on a larger scale
Income	Conversion, retention, commercial speed, personalization	Increase in revenue, gross margin, LTV	Measuring only savings and losing upside
Risk and compliance	Evidence, traceability, privacy, controls, audit	Lower exposure to fines, litigation, incidents, or reputational loss	Scaling AI without control
Adoption	Effective use by role, recurrence, satisfaction, reduction of friction	Achievement of projected profit	Having attractive pilots but without impact
Scalability	Reusing components, integration, governance, and monitoring	Lower marginal cost per new use case	Reinventing every project from scratch

Why many AI projects don't test ROI

The problem usually isn’t artificial intelligence. The problem is measuring it as if it were an isolated tool rather than an intervention on processes, people, data, and decisions.

McKinsey reported in its 2025 global survey that only 39% of the respondents attributed any impact on EBIT to the use of AI, and that the majority of those who do report impact place it below 5%. Organizational EBIT. It also identifies a small group, around 6%, as “AI high performers” with an EBIT impact of 5% or more.

BCG, for its part, found that 74% of the companies still do not demonstrate tangible value from their use of AI, while only 26% have developed capabilities to overcome concept tests and generate measurable value.

Executive reading is direct: AI should not be evaluated by the number of pilots, but by the captured and repeatable value.

Executive formula for calculating AI ROI

The basic formula is:

AI ROI = [(Net financial benefit – Total cost of AI) / Total cost of AI] x 100

But for it to be useful in a management committee, the net financial benefit must be divided into four blocks:

Operational savings: reduction of hours, tickets, reprocesses, support costs or manual activities.
Incremental income: higher conversion, lower churn, better recommendations, cross-selling or speed of sales.
Risk reduction: less incidents, regulatory errors, fines, data exposure, or untraceable decisions.
Value of released capacity: Time of expert teams reassigned to activities with greater margin.

Simplified example:

Variable	Annual amount
Savings due to reduction in manual hours	$2,400,000 MXN
Reduction of errors and rework	$900,000 MXN
Increase attributable to revenue	$1,800,000 MXN
Estimated reduction in operational risk	$600,000 MXN
Total financial benefit	$5,700,000 MXN
Total annual cost of AI	$2,100,000 MXN
Estimated ROI	171%

The final figure should not be presented alone. It must be accompanied by assumptions, a measurement period, data source, and level of attribution. An ROI without traceability is like a board without a minutes: everyone believes something happened, but no one can verify it.

Step 1: Define the use case with financial impact

Before selecting a model, platform or provider, the company must answer:

What process will be improved?
Which financial indicator will be affected?
Which area will own the result?
What decision will change thanks to AI?
What cost, loss, or risk is being sought to be reduced?
What will happen if the project is not implemented?

An AI project with good ROI potential usually meets three conditions:

It has sufficient volume of transactions, documents, tickets, requests or decisions.
It has visible financial pain: cost, delay, error, loss, fine, or low conversion rate.
It has available data or a realistic possibility of instrumenting them.

Step 2: build the baseline before the pilot

The baseline should capture the current state of the process over a representative period. For business operations, it is advisable to measure over 4 to 12 weeks, depending on volume and seasonality.

Recommended metrics:

Average time per activity.
Cost per transaction.
Monthly volume.
Percentage of error or rework.
Opportunity cost due to delay.
Compliance with SLA.
Losses due to incidents or late decisions.
Level of satisfaction of the user or customer.
Hours of man-time invested by role.
Cost of supervision, audit or control.

Without this measurement, the company can only say “it feels faster”. And in finance, “feels” does not count towards EBITDA.

Step 3: measure real costs, not just licenses

The total cost of AI must include initial investment and ongoing operation. Gartner warns that at least 30% of the GenAI projects would be abandoned after the proof-of-concept by the end of 2025 due to poor data quality, insufficient risk controls, rising costs, or unclear business value.

Costs that must be included:

Diagnosis and design of the use case.
Preparation, cleaning and data management.
Integration with ERP, CRM, ITSM, data lake, APIs or legacy systems.
Licenses, model consumption, infrastructure, and storage.
Development, configuration, testing and security.
Performance monitoring, biases, accuracy, and drift.
Change management and training.
Internal business, IT, legal, compliance and security hours.
Auditing, documentation and evidence.
Support, maintenance and continuous improvement.

Step 4: separate technical metrics from business metrics

A model can have good accuracy and still not generate ROI. That’s why measurement must have two layers.

Type of metric	Examples	For whom does it matter?
Technical metrics	Precision, recall, latency, availability, rate of hallucination, drift	TI, data, architecture, security
Business metrics	Cost avoided, incremental revenue, reduced cycle time, productivity, reduced exposure to fines	CFO, COO, CEO, General Management
Adoption metrics	Active users, recurrence, completed tasks, satisfaction, abandonment	Digital transformation, operations, HR
Control metrics	Evidence, traceability, permits, human reviews, compliance	Legal, compliance, CISO, audit

Step 5: Use cohorts and comparison groups

To avoid attributing results to AI that come from another source, it is advisable to use controlled comparison:

Group with AI vs. group without AI.
Pilot region vs. control region.
Automated process vs. manual process.
Before/after with seasonality considered.
Adoptive users vs. non-adoptive users.
Comparable business cycles.

This allows us to estimate incremental impact. It is not the same thing to say “the process improved” than to demonstrate “the process with AI reduced 28% the time against the control group”.

Step 6: Calculate payback, ROI and present value

For the financial committee, a percentage ROI is not enough. The case must include:

Revenge: Months are needed to recover the investment.
Annual ROI: percentage return per year.
TCO: Total cost of ownership.
NPV or VPN: net present value when the project involves several years.
IRR or TIR: Internal return rate to compare against other investments.
Cost of not acting: accumulated loss from maintaining the current process.

AI competes for budget with cybersecurity, cloud, ERP, automation, business expansion, and talent. If it doesn't speak the financial language, it stays in the lab.

Step 7: Govern AI to protect the ROI

The government is not a brake; it is a safe investment. NIST structures the risk management of AI into functions Government, Map, Measure and Manage, This helps to connect control, measurement and responsible operation.

ISO/IEC 42001 also sets requirements for creating, implementing, maintaining, and improving an AI management system, including risks, opportunities, traceability, transparency, and continuous improvement.

Minimum government requirements for measuring AI ROI:

Inventory of use cases.
Business owner on an initiative basis.
Technical owner and data controller.
Approval criteria before production.
Risk assessment by impact.
Privacy, security and access policies.
Registration of automated or assisted decisions.
Performance monitoring and deviations.
Periodically revised financial metrics.
Evidence for internal or external auditing.

How Scanda helps to measure AI ROI in an organized way

Scanda should be in place before deployment, not when the pilot has already become unmanageable. Its value lies in turning AI ideas into prioritized, measurable, and controllable initiatives.

Scanda's recommended approach:

AI Readiness Assessment: evaluate the maturity of data, processes, risks, architecture, and internal capabilities.
Map of use cases: Classify initiatives based on financial impact, technical feasibility, and operational risk.
ROI model: Define baseline, assumptions, costs, benefits, and scenarios.
Pilot-controlled design: establish cohorts, KPIs, and success criteria before investing on a scale.
Automation and integration: connect AI with real systems, not leave it as an isolated tool.
AI Government: documenting controls, evidence, traceability, and monitoring.
Scaling: Reuse components, data, flows, and models to reduce the marginal cost of new cases.

The difference is not in “using AI”. It is in Making AI leave valuable evidence.

Predictive conclusion

In the next 12 to 24 months, companies will stop funding AI projects solely on perceived innovation. The pressure will come from finance, audit, security, and general management: each initiative must demonstrate contribution to EBITDA, risk reduction, operational efficiency, or growth.

The market is already separating organizations that build pilots from those that build repeatable capabilities. BCG points out that leading companies focus their AI investments on transforming key functions and creating new offerings, not just isolated productivity improvements.

The prediction is clear: AI that is not measured will be cut back; AI that is governed and proves ROI will be scaled up. Scanda can be the facilitator for companies to move from scattered experiments to an AI portfolio with method, control, and verifiable financial value.

FAQ

1. How long should it take to measure ROI in an AI project? It depends on the use case. In high-volume operational automations, initial impact can be measured between 8 and 12 weeks after the pilot. In strategic, predictive, or process transformation projects, it is advisable to measure in stages: adoption, productivity, financial impact, and scalability.

2. What happens if the AI project improves productivity but does not directly reduce costs? It should be measured as freed capacity. If the hours saved are reallocated to tasks of higher value, reducing the backlog, customer service, financial analysis, or revenue generation, there is an economic impact. If only time is saved without redesigning the process, the ROI is incomplete.

3. Who should be responsible for the ROI of AI: IT, business or finance? The ultimate responsible person should be the owner of the business process. IT enables architecture, integration, security, and operation. Finance validates assumptions, costs, and benefits. Digital transformation coordinates adoption. Without this shared responsibility, the project risks becoming a technical pilot with no financial impact.

The Artificial intelligence It shouldn’t be left to difficult-to-justify pilots. With MaIA by Scanda, your company can identify use cases with financial impact, build a baseline, calculate ROI, control risks, and scale automation and AI applied in an organized manner.

Schedule a diagnosis with Scanda and turn your AI initiatives into measurable, governed, and defensible results for management.