Measuring AI ROI

Frameworks to Achieve 3X–5X Returns

Artificial intelligence (AI) budgets are rising, but board confidence is not automatic. The reason is simple: many organisations are still measuring AI as a technology deployment rather than as a business outcome.

The warning signs are visible. MIT Project NANDA’s 2025 State of AI in Business research reported that, despite an estimated $30–40 billion of enterprise generative AI investment, 95% of organisations in its sample were getting no measurable return, while only 5% of integrated pilots were extracting material value. The same report found that most enterprise-grade systems failed because of brittle workflows, poor contextual learning and misalignment with day-to-day operations. MIT NANDA, via SlideShare

McKinsey’s 2025 global AI survey tells a similar story from another angle: 88% of respondents said their organisations were using AI in at least one business function, yet most had not embedded it deeply enough into workflows to realise material enterprise-level benefits. The high performers, around 6% of respondents, were more likely to redesign workflows, scale faster, secure senior leadership ownership and use AI for growth and innovation as well as efficiency. McKinsey

That is the core lesson for executives: AI return on investment (ROI) is not proved by access to a tool, model usage, prompt volume or a successful demo. It is proved by a measured change in cost, revenue, risk or productivity against a signed-off baseline.

The difference between AI activity and AI value

AI activity is easy to count:

• number of users with access to a tool;
• number of prompts submitted;
• number of documents summarised;
• number of agents prototyped;
• number of pilots launched.

AI value is harder, but more useful:

• fewer hours spent on a process;
• lower error and rework rates;
• faster response or cycle times;
• higher conversion, retention or revenue per employee;
• reduced exposure to compliance, operational or security risk;
• measurable customer or employee experience improvement.

The first set of measures can show adoption. The second set shows whether adoption is worth anything.

This is why Virtco®’s AI readiness approach starts with the business outcome, not the tool: find the real problem, measure the baseline, test a controlled pilot and scale what works. The public AI readiness guidance is explicit that organisations struggle to prove the “after” if they have not measured the “before”, and that a controlled pilot should have success criteria, adoption planning and risk controls from the start. Virtco® AI Readiness and Exploitation

The core AI ROI formula

The simplest ROI formula is still useful:

AI ROI (%) = ((Total realised benefits - Total AI costs) / Total AI costs) × 100

For example, if an AI initiative creates £300,000 of measured annual benefit and costs £100,000 to design, implement, run and support, the net ROI is:

((£300,000 - £100,000) / £100,000) × 100 = 200% ROI

However, many boards prefer to talk in return multiples:

Return multiple = Total realised benefits / Total AI costs

So:

• £300,000 benefit from £100,000 cost = 3× gross return;
• £500,000 benefit from £100,000 cost = 5× gross return.

That means a 3–5× return multiple is equivalent to 200–400% net ROI, assuming the multiple is calculated on gross benefits divided by cost.

Be clear which definition you are using. Confusing gross return multiple with net ROI percentage is one of the fastest ways to lose financial credibility.

The four pillars of modern AI value

Modern AI ROI should not be limited to labour cost reduction. The stronger business cases usually combine four types of value: cost avoidance, revenue generation, risk mitigation and productivity gains. Elvex uses the same four-pillar structure in its AI ROI guidance, and it is a useful way to stop AI cases becoming too narrow. Elvex

1. Cost avoidance

Cost avoidance is the money the organisation does not have to spend because AI removes friction, prevents rework or delays the need for additional capacity.

Typical examples include:

• avoiding additional hires by absorbing growth with existing teams;
• reducing external agency, business process outsourcing or contractor spend;
• lowering rework caused by inconsistent data entry or manual checking;
• reducing support tickets through better self-service;
• avoiding overtime or temporary labour during peak periods.

A practical formula is:

Cost avoidance = Avoided unit volume × Cost per unit

For headcount-related avoidance, use a cautious formula:

Avoided hiring value = Avoided full-time equivalent capacity × Fully loaded annual employment cost × Confidence factor

Use a confidence factor because saved time does not automatically become cashable saving. If the same people are still employed, the benefit may be productivity capacity rather than reduced spend.

2. Revenue generation

Revenue generation measures new or accelerated income enabled by AI.

Typical examples include:

• faster lead response improving conversion;
• better personalisation increasing average order value;
• sales teams spending more time on qualified opportunities;
• AI-assisted product or proposal development reducing time to market;
• new AI-enabled services or premium support tiers.

A practical formula is:

Revenue benefit = Incremental revenue × Gross margin

Do not count revenue at 100% unless the business case is deliberately measuring turnover rather than profit contribution. For ROI, gross margin is usually more defensible.

For sales conversion use cases:

Incremental gross profit = Additional qualified opportunities × Conversion uplift × Average deal value × Gross margin

3. Risk mitigation

Risk mitigation is the value of reducing the likelihood or impact of adverse events.

Typical examples include:

• earlier detection of fraud, anomalies or operational failures;
• fewer compliance breaches;
• reduced data leakage from uncontrolled AI use;
• fewer customer-impacting errors;
• reduced reliance on unsupported manual workarounds.

A practical formula is:

Risk mitigation value = (Baseline risk exposure - Residual risk exposure) × Attribution factor

Where:

Risk exposure = Probability of event × Financial impact of event

This is not perfect, but it creates a disciplined conversation. It also helps boards compare AI controls with other risk treatments.

For AI specifically, the risk case should include both reduced existing risk and new AI-related risk. Virtco®’s AI Risk Framework™ treats AI risk management as a continuous loop, including ownership, risk indicators, controls, monitoring, incident response and performance tracking.

4. Productivity gains

Productivity gains measure how much more useful work people can complete with the same or fewer resources.

Typical examples include:

• summarising documents faster;
• reducing manual data entry;
• drafting reports, emails or proposals more quickly;
• accelerating research and knowledge retrieval;
• shortening software, analysis or content production cycles.

A basic formula is:

Productivity value = Hours saved × Fully loaded hourly cost × Recapture rate

The recapture rate matters. If AI saves 1,000 hours, but only 40% of that time is redirected into valuable work, the measured productivity value should be based on 400 hours, not 1,000.

A more defensible version is:

Productivity value = Hours saved × Fully loaded hourly cost × Recapture rate × Quality factor

The quality factor adjusts for whether AI output is as good as, better than or worse than the previous process after human review.

The total cost side of the equation

AI ROI is often overstated because the cost side is too thin. A serious AI business case should include total cost of ownership, not just licences.

Include:

• discovery, process mapping and use case prioritisation;
• data cleansing, data access and integration work;
• model, platform, licence and usage costs;
• infrastructure, compute and storage;
• security, privacy, legal and compliance review;
• workflow redesign and target operating model changes;
• testing, evaluation and quality assurance;
• staff training, communications and adoption support;
• human-in-the-loop review effort;
• monitoring, maintenance, incident handling and continuous improvement;
• internal management time and opportunity cost.

A practical formula is:

Total AI cost = One-off implementation costs + Ongoing run costs + Change and adoption costs + Risk and governance costs

If the cost model excludes adoption, support and governance, the ROI number is not board-ready.

IBM’s 2025 CEO study reinforces this point. It found that only 25% of AI initiatives had delivered expected ROI over the previous few years and only 16% had scaled enterprise-wide, while half of surveyed CEOs said rapid investment had left their organisation with disconnected, piecemeal technology. IBM

A 90-day AI ROI template

The first 90 days should not attempt to prove every possible benefit. They should prove whether the use case is real, measurable, adoptable and worth scaling.

Step 1: Define the value hypothesis

Use this template before any build begins.

Field	Template	Example
Business problem	What measurable problem are we solving?	Customer service triage takes too long.
Current baseline	What is happening today?	1,200 tickets per month, average first response 18 hours.
Target outcome	What must improve?	Reduce average first response to 6 hours.
Value pillar	Cost, revenue, risk or productivity?	Productivity and customer experience.
Benefit owner	Who owns the business result?	Head of Customer Operations.
Measurement owner	Who captures and reports the data?	Service operations analyst.
Data source	Where will the evidence come from?	Service desk timestamps and quality review sample.
Adoption group	Who must use the new workflow?	Tier 1 support team.
Control group	What will we compare against?	One support queue remains on the existing process for four weeks.
Go/no-go threshold	What result justifies scaling?	40% cycle-time reduction with no quality degradation.

The important point is ownership. A technology lead can own delivery, but a business leader must own the benefit.

Virtco®’s 3-Bees Framework formalises this discipline. Bee 1, Benefits Capture, creates a typed, baselined and owned benefits log, identifies dis-benefits, captures constraints and risks, and assigns named benefit owners before the organisation moves into delivery.

Step 2: Capture the baseline

A useful baseline contains both volume and quality.

Metric	Baseline value	Source	Owner	Notes
Monthly volume
Average handling time
Average cycle time
Error or rework rate
Escalation rate
Customer satisfaction
Employee effort or pain score
Current cost per unit

Do not automate a process you have not measured. If the “before” picture is vague, the “after” claim will be contested.

Step 3: Build a controlled pilot

The 90-day pilot should be narrow enough to measure and safe enough to learn from.

A practical sequence is:

Period	Focus	Output
Days 1–15	Confirm problem, baseline, risk and data readiness	Signed value hypothesis and baseline pack
Days 16–30	Design workflow, controls and measurement events	Pilot design, success criteria and test plan
Days 31–60	Build and run controlled pilot	Working pilot with human-in-the-loop review
Days 61–75	Measure results and adoption	Benefits evidence, user feedback and risk log
Days 76–90	Decide whether to stop, iterate or scale	ROI report and scale recommendation

This mirrors Virtco®’s practical AI readiness guidance: assess the current position, prioritise the right use case, build a safe pilot, train the team and then scale what works. Virtco® AI Readiness and Exploitation

Step 4: Calculate 90-day pilot ROI

Use this template for the first ROI calculation.

Input	Formula	Value
Baseline process volume	Units per period
Baseline time per unit	Minutes or hours
AI-assisted time per unit	Minutes or hours
Time saved	Baseline time - AI-assisted time
Fully loaded hourly cost	Salary, on-costs and overhead
Recapture rate	% of saved time converted into useful work
Productivity value	Hours saved × hourly cost × recapture rate
Cost avoidance	Avoided cost items
Revenue contribution	Incremental revenue × margin
Risk mitigation	Reduced expected loss
Total measured benefit	Sum of all benefits
Total pilot cost	Build + licences + data + change + governance
Net ROI	(Benefit - cost) / cost × 100
Return multiple	Benefit / cost
Payback period	Cost / monthly benefit

The pilot ROI should be presented with confidence levels:

• confirmed benefit: evidenced in operational data;
• probable benefit: supported by pilot data but requiring longer observation;
• potential benefit: plausible, but not yet evidenced.

Only confirmed and carefully qualified probable benefits should be used for the scale decision.

A beyond-90-days ROI template

After 90 days, the question changes. The pilot has either failed, needs redesign or has earned the right to scale.

At this point, ROI measurement should move from a pilot spreadsheet into a benefits register and operating cadence.

Quarterly benefits register

Benefit ID	Benefit description	Pillar	Owner	Baseline	Target	Actual	Variance	Confidence	Action
B1		Productivity
B2		Cost avoidance
B3		Revenue
B4		Risk

Every benefit should have a named owner, a baseline, a target, a measurement method, dependencies, risks and a realisation timeline. That is exactly the discipline described in Bee 2 of Virtco®’s 3-Bees Framework, where benefit profiles are built and tied to measurement plans, key performance indicators and the infrastructure needed to capture them in production.

Annualised ROI

Once the use case is in production, calculate annualised ROI:

Annualised AI ROI (%) = ((Annualised realised benefit - Annualised run cost - Amortised implementation cost) / (Annualised run cost + Amortised implementation cost)) × 100

For multi-year investments, add discounted cash flow:

NPV = Σ (Net benefit in year t / (1 + discount rate)^t) - Initial investment

Use net present value (NPV) when:

• benefits ramp up over several years;
• costs are front-loaded;
• the programme competes with other capital investments;
• risk and uncertainty are material.

Scale-readiness checklist

Before scaling, answer these questions:

• Has the pilot improved a business metric, not just a usage metric?
• Is the result repeatable across teams, locations or workflows?
• Is the data source stable and trusted?
• Are users adopting the new way of working without excessive support?
• Are human review points clear?
• Are privacy, security, bias and compliance risks controlled?
• Is there a benefit owner for the scaled process?
• Is there a sustainment owner after the project team steps away?
• Have dis-benefits been identified and managed?
• Is there a dashboard or review cadence for ongoing performance?

Scaling without this evidence turns a small unproven experiment into a larger unproven cost base.

Why outcome measurement must be embedded from the start

Outcome measurement cannot be bolted on at go-live. By then, it is often too late.

If the team did not capture the baseline, they cannot prove improvement. If the workflow was not instrumented, they cannot collect the right data. If no one owns the benefit, no one will chase it. If adoption was not planned, the system may be technically correct and commercially irrelevant.

This is why Virtco®’s 3-Bees Framework is useful for AI ROI.

Bee 1: Benefits Capture

Bee 1 answers: What value are we trying to create, and who owns it?

It captures:

• the business problem;
• typed benefits and dis-benefits;
• quantified baselines;
• named benefit owners;
• readiness and impact assessments;
• constraints, risks and dependencies;
• the outline business case.

The result is not a vague AI ambition. It is a measurable benefits hypothesis with accountable ownership.

Bee 2: Benefits Mapping and Profiling

Bee 2 answers: How will the solution create the benefit?

It connects:

AI capability → business change → intermediate benefit → end benefit → strategic objective

This is where benefit profiles, key results, key performance indicators and measurement infrastructure are defined. Virtco®’s framework specifically warns that teams often define good KPIs and then discover at go-live that they cannot collect them. Measurement design must therefore happen during solution design, not after deployment.

Bee 3: Benefits Realisation

Bee 3 answers: Did the value actually land, and will it stick?

It includes:

• pilot results;
• training records;
• hypercare logs;
• adoption and KPI dashboards;
• benefits realisation reviews;
• comparison of actual versus forecast benefit;
• sustainment and business-as-usual handover.

The framework is explicit that adoption alone is not ROI. Benefits realisation reviews must measure actual versus forecast benefit against profiles and baselines, attribute the result, report it to the sponsor and update the benefits register.

The 3–5× AI return pattern

BCG’s 2025 research found that only 5% of firms were “AI future-built”, but those firms achieved five times the revenue increases and three times the cost reductions of other companies from AI. BCG

That does not mean every AI project should promise a 3–5× return. It means the organisations getting outsized returns tend to share a pattern:

• they focus on business value, not novelty;
• they redesign workflows rather than dropping AI onto broken processes;
• they combine productivity, revenue, risk and cost measures;
• they secure senior ownership;
• they build adoption and reinforcement into delivery;
• they scale proven use cases rather than multiplying disconnected tools;
• they keep measuring after go-live.

A credible 3–5× AI business case should therefore show:

Clear baseline + accountable owner + controlled pilot + measured benefit + governed scale path

Without those elements, the number is aspiration.

With them, it becomes a managed investment case.

Common AI ROI mistakes

Mistake 1: Counting time saved as cash saved

If AI saves ten hours per week, that is valuable. But it is not automatically a cash saving unless cost is actually removed or additional valuable work is completed.

Use recapture rates and explain the assumption.

Mistake 2: Measuring the model instead of the workflow

Accuracy, latency and prompt quality matter, but they are not the business outcome. Measure the end-to-end workflow: cycle time, throughput, rework, conversion, customer satisfaction and risk.

Mistake 3: Ignoring adoption

A technically capable AI solution creates no value if people avoid it, mistrust it or work around it. Virtco®’s adoption material treats leading indicators such as communication reach, training completion, pilot feedback and sponsor activity as early warnings, while lagging indicators such as active usage, proficiency, reduced workarounds and business outcomes show whether value is being realised.

Mistake 4: Forgetting risk-adjusted value

AI may reduce one risk while creating another. A customer-facing assistant might reduce service workload but introduce hallucination, privacy or conduct risk. ROI should be risk-adjusted, not risk-blind.

Mistake 5: Scaling too early

A promising demo is not a production case. Scale only when the benefit is evidenced, the operating model is clear, controls are in place and ownership has transferred into the business.

A practical executive scorecard

Use this scorecard at steering group or board level.

Question	Red	Amber	Green
Is the business outcome defined?	Vague efficiency goal	Outcome defined but weak metric	Specific metric and target agreed
Is the baseline measured?	No baseline	Partial or anecdotal baseline	Signed-off baseline with data source
Is there a benefit owner?	No owner	Technology owner only	Named business owner accountable
Are costs complete?	Licences only	Some delivery costs included	Full total cost of ownership included
Are risks controlled?	No AI risk review	Risk review in progress	Controls, owners and monitoring agreed
Is adoption planned?	Training after go-live	Basic training plan	Role-based adoption and reinforcement plan
Is the pilot measurable?	Demo only	Some operational metrics	Controlled pilot with go/no-go criteria
Is scale justified?	Based on enthusiasm	Based on limited usage	Based on realised benefit and readiness

If any row is red, the initiative is not ready for a confident ROI claim.

Conclusion: ROI is engineered, not discovered

AI ROI does not appear at the end of a pilot. It is engineered from the start.

That means choosing a valuable problem, capturing the baseline, assigning benefit ownership, designing the measurement infrastructure, managing adoption, controlling risk and reviewing actual benefit after deployment.

The organisations achieving 3–5× returns are not simply buying better tools. They are better at connecting AI capability to business change and business change to measurable outcomes.

If your organisation is experimenting with AI but cannot yet prove value, start with the measurement system. Virtco® can help you assess readiness, prioritise use cases, define baselines and build a practical route from pilot to measurable return through its AI Readiness and Exploitation accelerator.