Free resource · Word Kit

AI Measurement and ROI Kit: Numbers That Hold Up for the Board

Measure the uses, value and return on investment of generative AI with defined indicators, a before/after protocol and a conservative ROI calculation.

📄 Word, 27 pages📊 4 families of indicators💶 ROI in 3 scenarios🆓 Free

Download the kitSee the four families

The question always comes up, and usually at the worst possible time: after six to twelve months, when the cost of licenses becomes apparent and the initial enthusiasm has faded. "What's the return on investment for us with AI?" The organizations that answer this well aren't necessarily those that have spent the most; they're the ones that measured from the outset, using few indicators, written definitions, and a baseline established before deployment.

This page explains what to measure, how to obtain defensible figures, how to calculate a return on investment that a CFO will accept, and which calculation errors destroy the credibility of a balance sheet. It comes with a complete toolkit, in Word format, with ready-to-use tables and protocols.

1. Why measure before you need numbers?

The measurement is the only aspect of a deployment that cannot be recovered. A charter can be written in three months, training scheduled in six months, an ambassador network established in nine months. A starting point, however, can only be established beforehand. Once the deployment is complete, only approximate reconstructions remain, which cannot withstand the first scrutiny of management control.

The second reason is political. A project leader who arrives at a management meeting with a single, flattering, and unadjusted figure is torn apart in ten minutes. The one who arrives with a conservative figure, an acknowledged discount, a range, and a list of what they couldn't measure gets the discussion they were looking for: not "Is this figure accurate?" but "What are we doing next year?" Caution isn't a weakness in communication; it's the best argument available.

Finally, measurement is a powerful driver of adoption in its own right. Publishing usage rates by department, without ranking individual performance but with support offered to those lagging behind, has a greater impact than an internal communication campaign. Departments compare themselves, and the leading teams showcase their use cases.

2. The four families of indicators

The correct approach is not to export what the administration console offers, but to write down the questions the management committee will want answered, and then find the simplest figure that answers them. Four questions, four categories, one main figure per category.

Committee question	Family	Main figure
"Do people use it?"	Adoption	Percentage of employees within the scope who use the tool at least once a week.
"To do what?"	Usage	Number of documented and reused use cases, and their distribution by department.
"What does this bring us?"	Value	Time saved per regular user per week, plus a process indicator measured before and after.
"What does it cost, and what are the risks?"	Costs and risks	Total annual cost, including internal time, and number of incidents or deviations observed.

These categories are interconnected. A high adoption rate without a measurable value indicates superficial use; a high value within a narrow scope suggests a use case that should be scaled up; and a rising cost without adoption indicates licenses that need to be redistributed. The toolkit provides, for each indicator, its precise definition, source, and collection frequency, including the definition of "active user," which is the most debated and the most cost-effective to define definitively.

3. The three methods of measuring the value

They are not mutually exclusive: the first establishes the trend, the second provides the evidence, and the third enables long-term monitoring. A serious organization uses all three, albeit within different scopes.

Method	What she gives	What it costs
The framed declaration	A median time saved by regular users, and especially its change from one quarter to the next. Biased upwards, therefore discounted by half before any economic conversion.	A question in the quarterly barometer.
The before/after measurement	The only truly defensible figure: the difference in production time for a deliverable, on the same people and the same task, with quality control.	Two to three days of work, on a pilot.
The process indicator	A figure already held by the profession (response time, volume processed, completion rate), which relates to the final result and not to an isolated task.	Almost nothing: the figure already exists.

The before/after protocol is what convinces a management committee, and the toolkit details it in seven steps: choose the task, define the unit of measurement, record the "before" data on at least fifteen deliverables, train and equip the team, allow for a learning curve, record the "after" data, perform blind quality control, and write the results in one page. It also highlights the four pitfalls that invalidate the measurement: the novelty effect, changes in scope, self-selection of enthusiastic volunteers, and the hidden time spent on verification and correction.

4. Calculate the ROI without making a mistake

The return on investment calculation is not the core of the measurement; it is the annual conclusion. If done too early, it relies on assumptions that are too fragile. If done once a year, in conjunction with management control and using figures already published during the year, it closes the debate instead of opening it.

On the cost side, the overlooked item is always the same: internal time. A lead working 40% of a full-time position for a year costs roughly the same as a hundred licenses, and omitting it undermines the calculation from the very first question. The toolkit details the five items to consider: licenses, training and support (including participants' time), internal management time, technical and integration time, and compliance and legal time.

In terms of earnings, three conventions allow for an accurate presentation of the figure: reducing the declared amount by half, valuing only regular users and not the total workforce, and applying a realization rate of 30% to 50% because not all hours saved translate into value for the organization. The toolkit's standard formula is: weekly active users, multiplied by the hours recorded per week, multiplied by 44 weeks, multiplied by the loaded hourly rate, and multiplied by the realization rate. The result is presented in three scenarios, with the conservative scenario serving as the benchmark, rather than a single figure that invites dispute.

5. The five mistakes that ruin a balance sheet

Apply the gain to the entire workforce. This is the first correction your CFO will make. Make it before him: the benefit only applies to regular users, which is why adoption is the primary driver of profitability.

Forget about internal time in the costs. Referents, ambassadors, pilot project leaders, training participants: all this time is valued at the full hourly rate, like any other project.

Extrapolate a successful pilot program to the entire organization. A pilot program is conducted with supervised volunteers. Applying its benefits to untrained teams massively overestimates the result. Extrapolation is limited to the actual scope of the training, with a corresponding reduction in value.

Announce job cuts. Unless management makes an explicit and deliberate decision, the hours saved do not represent job losses. Stating this will trigger reactions that will cost more than the announced gain.

Show only gains. A dashboard without indicators that could deteriorate is not credible. The recovery rate of generated deliverables and the incident log lend credibility to the entire system and protect the project when an incident occurs. A log with zero incidents after a year is also suspicious: it most often indicates that no one is reporting them.

6. What the kit contains

The kit is a 27-page, editable Word document with a decision area at the end of each step. It contains, in order: the user guide and the three measurement principles; the "Deciding What to Measure" section, with the four categories, the useful indicator test, what to forgo measuring, and the data collection schedule; the "Measuring Usage" section, with data sources, definitions of the three rates, and an interpretation of adoption by department; the "Measuring Value" section, with the three methods and the detailed before/after protocol; the "Calculating ROI" section, with the five cost items, conversion conventions, three scenarios, and six calculation pitfalls; and the "Reporting and Deciding" section, with the five-page summary, the three trade-offs, and answers to the five common objections. and eight annexes: complete dashboard, before/after measurement sheet with its record table, barometer questionnaire in fifteen questions, return on investment calculation sheet, template for the report to the management committee, frequently asked questions, self-assessment of the system rated out of 40, and resources to go further.

The kit, part by part: what each part contains and what it is used for. The templates to fill in are in the Word document.

Part 1. Deciding what to measure

1.1 The four questions and the four families

Four questions from the management committee yield four categories of indicators: adoption ("Are people using it?"), usage ("For what purpose?"), value ("What does it bring us?"), and costs and risks. These four categories are sufficient for year 1, with one main figure for each.

1.2 The principle of the useful indicator

An indicator deserves its place if it meets four conditions: it changes a decision, it is unambiguously defined, its source is stable, and it is comparable over time. The cost of collecting an indicator is paid monthly, not once: the toolkit recommends a maximum of four to six indicators, validated with management control.

1.3 What we must stop measuring

Three deliberate sacrifices: individual productivity, because measuring usage per person breaks trust (we measure by team or by department); the intrinsic quality of AI productions, replaced by the recovery rate; directly attributable revenue, replaced by a process indicator that the business already maintains.

1.4 The measurement timetable

A starting point for the four families before deployment, a monthly report on adoption, usage, and incidents, a quarterly survey of employees, a before/after measurement for each pilot project, and an annual review with management control. This pace requires half a day per month for the lead contact and one full day per quarter.

Part 2. Measuring usage

2.1 Where do the figures come from?

Five sources, each with its limitations: the tool's administration console, whose definition of "asset" varies from one publisher to another; the directory and the HRIS, to calculate rates by department; the training register; the internal library of use cases, which only measures what is shared; the quarterly barometer, which is declarative and therefore biased.

2.2 Define "active user" once and for all

Recommended definition: an individual within the scope who has interacted with the reference tool at least three times during a calendar week, relative to the number of employees with a license. Three rates should be distinguished: the equipment rate (licenses issued per employee), the activation rate (individuals who have used the tool at least once per license issued), and the regular usage rate, the only one that reflects established practice and the figure to be presented to the management committee.

2.3 Read the adoption by management

The organization's average always masks discrepancies of one to five between departments: it is this dispersion that can be addressed. The toolkit combines usage and training: low usage despite extensive training points to access to tools or management; high usage without training indicates personal accounts that need to be equipped and managed quickly.

2.4 Count uses, not just users

The number of documented and reused use cases is the best leading indicator of value. Three figures suffice: the number of published cases, the proportion reused by a team other than the one that created them, and the number of departments represented. A catalog of sixty cases all produced by the same department is worth less than twenty cases spread across eight departments.

Part 3. Measuring the value

3.1 Method 1: Framed declaration

A single question in the quarterly barometer, asked identically in each wave, with time slots ranging from less than 30 minutes to more than 8 hours per week. The median of regular users is used, not the average. Self-reported usage overestimates: the toolkit applies a 50% discount for any economic conversion and presents both the self-reported and the actual usage.

3.2 Method 2: Before/after measurement on a pilot

The convincing method requires two to three days of work and produces the only truly defensible figure of the year. The seven-step protocol involves a repetitive task of at least ten occurrences per month, a review of at least fifteen deliverables before and after, two weeks of familiarization, and a blind quality review.

3.3 Method 3: The Process Indicator

For general use, we track an indicator that the business already uses: response time to a call for tenders, time to process a complaint, time to produce the monthly report. We never discuss causality: we present the curve, indicate the deployment date, and note other changes during the period. The toolkit lists these indicators for six functions.

3.4 Also measure quality and risk

Two indicators that can deteriorate lend credibility to the dashboard: the rework rate, the proportion of generated deliverables requiring significant correction, tracked across fifteen deliverables per month, and the incident log. A log with zero incidents after a year is suspicious: it most often indicates that no one is reporting them.

Part 4. Calculating the return on investment

4.1 Count all costs

Five key areas: licenses, often 40 to 60% of the total; training and support, 20 to 40%, including participant time; internal management time, 10 to 30% and almost always overlooked; technical support and integration, minimal in year 1; and compliance and legal matters, concentrated in the first quarter. A dedicated contact person working 40% of a full-time position for a year costs roughly the same as one hundred licenses.

4.2 Convert the hours earned, with caution

An hour gained is not an hour saved. The kit reduces the declared value by half, only values ​​weekly active users, and applies a completion rate of 30% to 50% in year 1 for support functions. Formula: weekly active users × hours retained per week × 44 weeks × hourly rate × completion rate. Conservative example: 300 × 1 hour × 44 × €45 × 30% = €237,600 in annual valued savings.

4.3 Present three scenarios rather than one number

The conservative scenario (declared discount of 50%, realization rate of 30%) is the figure that we defend and that we record in the balance sheet; the central scenario (discount of 30%, realization of 40%) is the working target; the high scenario (gain measured on the pilot, 60% if a redeployment is demonstrated) gives the order of magnitude of the potential.

4.4 The six calculation pitfalls

Apply the savings to the entire workforce, disregard internal time in the costs, add up overlapping gains, extrapolate a successful pilot to the whole organization, announce job cuts and present a ROI without intervals or assumptions. For each of these, the toolkit indicates what to do instead.

4.5 Beyond hours: other forms of value

Five forms of value are named rather than converted into euros, each with concrete proof: the quality and homogeneity of deliverables, responsiveness, decongestion of experts, attractiveness and competence (with Article 4 of the AI Act), and risk reduction when regulated uses replace personal accounts.

Part 5. Returning and deciding

5.1 The five-page annual report

Beyond five pages, the management committee stops reading. Recommended order: where we are (three figures), what it brings us, what it costs us, with the cost per regular user, the return on investment in three scenarios, then what we propose: value before costs, assumptions always visible, the decision last.

5.2 The three arbitrations

Discontinue activities that haven't been used regularly for three months and reallocate licenses; maintain services that provide a service at a constant cost (library, rituals, ambassadors, training log); invest in processes whose pilot has demonstrated a measurable benefit. Making these trade-offs prevents the automatic renewal of the budget.

5.3 Responding to the five classic objections

The kit provides a standard response to five objections: unaccounted-for hours saved, statements rather than measurements, the causality of AI, licenses paid for inactive users, and the discrepancy with press promises. For example: "We don't prove causality, we document it."

5.4 Putting the measure into practice

Three safeguards: include the data collection on the management committee's agenda starting in January, keep definitions and results in a single document, and entrust the monthly report to the tool's administrator. Publishing the results internally, even modest ones, makes the measurement a lever for adoption.

Annexes

Eight annexes: the complete dashboard, the before/after measurement sheet, the quarterly barometer questionnaire, the return on investment calculation sheet, the framework for the management committee report, frequently asked questions, the self-assessment of the system and resources to go further.

Appendix C. Quarterly Barometer Questionnaire

Fifteen questions, five minutes to complete, sent to the entire scope on a fixed date, divided into three sections: usage, perceived value, obstacles and needs. The wording remains the same from one wave to the next: it is their consistency that makes comparison possible.

Appendix G. Self-assessment of the measurement device

Ten statements, each scored from 0 to 4, to be repeated at six and twelve months. From 0 to 10, no action taken; from 11 to 20, only usage measures implemented; from 21 to 30, measure installed without an economic component; from 31 to 40, complete system. Each reading refers to the section of the kit to be addressed as a priority.

7. Frequently Asked Questions

Where do we begin if we haven't measured anything so far?

Three things to do, in this order: export the regular usage rate for the last three months from the administration console, launch the first wave of the barometer, and choose a pilot program to measure before and after the next deployment. In six weeks, you have a starting point and evidence in progress. Don't try to reconstruct the history.

Does the declarative statement have any value?

Yes, provided it is treated for what it is: a perception measure, biased upwards, whose value lies in its evolution over time and comparisons between directions. Discounted by half, it provides an acceptable basis for conversion; not discounted and presented as a measure, it destroys the credibility of the rest.

Should it be measured per person?

No. Individual measurement transforms the tool into an instrument of control, undermines the integrity of the statements, and raises legitimate questions from employee representatives. The appropriate level of measurement is management or the team, provided it comprises enough people to avoid being de facto a system of individual targeting.

What should be done if the ROI is negative in the first year?

This is common and normal: implementation, training, and management costs are concentrated in year 1, while gains increase with adoption. Present the quarterly trajectory rather than the annual balance, highlighting the decreasing cost per regular user and the projection for year 2 based on the scope of the system. An honest assessment with an acknowledged negative balance is more likely to secure renewal than a contested, optimistic one.

How can we know that the gains come from AI and not something else?

We don't prove it, we document it: scope, deployment date, other changes during that period. A before-and-after measurement of a pilot program, on the other hand, compares the same people performing the same task a few weeks apart, making it the strongest level of evidence accessible without an experimental protocol.

How do you measure the value of an agent or a delegated file?

Like a pilot, but with two distinct timeframes: the remaining human time (briefing, review, validation) and the total elapsed time. The gain is measured in human time, the delay in total time. Add a reliability indicator: the proportion of executions producing a usable deliverable without rework.

8. Going further

This kit expands on the fifth part of our AI Adoption Roadmap Kit in 12 Months If you're starting from scratch, begin with this one. NextStart.AI also assists large organizations with the implementation of the measurement system: choosing indicators, measuring a pilot project before and after implementation, and preparing the report for the management committee, as part of our services. adoption support.

Read also : AI Adoption Roadmap in 12 Months: The 2026 Toolkit, AI Charter in Business: The 2026 Model (Downloadable), AI Ambassador Network: The 2026 Toolkit et AI systems governance: the 2026 toolkit.

The entire series of free kits. To drive adoption: AI charter template, AI Ambassadors Kit, Manager's toolkit for dealing with AI, Adoption Roadmap Kit, AI maturity grid et AI Governance KitFor the AI ​​project sponsor: AI Act kit, GDPR kit, Security and data kit, AI and Social Dialogue Kit, frugal AI kit, AI Sovereignty Kit et responsible AI kitBy profession: AI HR kit, AI marketing kit, AI finance kit, AI kit for lawyers, AI kit for salespeople, AI toolkit for project managers, AI innovator kit et communicator's AI kit.

9. sources

The method described on this page is based on our field experience. The measurement principles it employs (discounting of self-reported data, before/after measurements on the same task, quality control, usage indicators from administration consoles) are based on the following public references.

• METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 10, 2025 (difference between perceived gain and measured gain).
• Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work, The Quarterly Journal of Economics, vol. 140, No. 2, 2025.
• Fabrizio Dell’Acqua et al. (Harvard Business School, Boston Consulting Group), Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, September 2023.
• Microsoft, Microsoft Copilot Usage Report, Microsoft 365 admin, Microsoft Learn documentation accessed in September 2026.
• Microsoft, Connect to the Microsoft Copilot Dashboard for Microsoft 365 customers, Viva Insights documentation accessed in September 2026.