AI ROI: how to measure whether AI is paying off

The ROI of artificial intelligence shows up when AI use is tied to a process the company already measures, with a baseline recorded before the pilot, the full cost on the books and an indicator that finance recognizes in the results. Most companies measure something else: how many licenses are active and how many hours the vendor estimates each person saved. Those numbers show access and goodwill from the team, and none of them reaches the income statement. I organize measurement in three layers: adoption and use, process efficiency, and revenue or cost results. Each layer has its own data source and its own timeline for showing movement, and much of the frustration boards feel with AI comes from demanding in week four what only the third layer can answer.

At Iasbeck&Co we implement CRM and AI agents in B2B operations and train the teams that work with them. In the workshop I ran in October, one of the most frequent questions in the chat was how much each agent action costs and how to know whether the spend pays off. I explained that each agent consumes credits according to the complexity of the request, and that average consumption gets planned before the agent is switched on. I wrote this for the executive team and finance, who approve the budget, for the sales, marketing and CS leaders, who answer for the indicators, and for HR, which will need to show the effect of training.

Why don’t active licenses and hours saved measure AI ROI?

The McKinsey global survey on AI from 2026 shows the gap between the two measures. Eight in ten respondents said AI improved their individual productivity, and 37% attributed some EBIT impact at their company to AI use, nearly the same share as in 2025. The person feels the gain in the task, and the company can’t find that gain in the results. In the same survey, the organizations McKinsey classifies as AI high performers are twice as likely to have defined processes to measure the impact of their initiatives.

An active license measures access. A salesperson who opens the assistant once a day to rewrite an email counts as an active user, and the vendor’s dashboard can’t tell you whether that email changed anything in the negotiation. Hours saved have a different problem, because they turn into money only if someone decides what to do with them. If a CS analyst saves four hours a week preparing for meetings and leadership neither grows that analyst’s book nor postpones the next hire, those four hours dissolve into the routine and payroll stays the same size.

How do you measure AI results in three layers?

Finance wants to know which line of the results will change and how soon, and the area’s leader needs to know which team behavior moves that line. The layers connect the two answers and build on each other: without use in the right process, efficiency doesn’t change, and without efficiency, results have no way to move.

Layer 1: adoption and use

The indicator that matters is use within the chosen process, measured in the CRM or in the tool where the work happens. In sales, count how many meetings got an agent-generated summary and how many follow-up drafts the salesperson approved without editing. In support, measure how many tickets received automatic classification and how many of those classifications the analyst corrected. The edit and correction rates say more than the number of users, because they show whether the context the agent reads is good enough for the team to trust it.

Layer 2: process efficiency

The second layer compares the process before and after, on the same team and over a comparable period. The time between a meeting and the logging of the next step, and the number of proposals stalled for more than 15 days, are indicators the CRM already records, as long as the base has reliable dates. This layer tends to move between weeks four and eight, and it bridges the conversation with finance, because a shorter sales cycle brings revenue forward and frees salesperson hours for more opportunities.

Layer 3: revenue or cost results

The third layer gathers the numbers finance recognizes: proposal-to-win conversion, revenue per salesperson, renewal rate and cost per resolved ticket. It takes longer because it depends on the sales cycle and the renewal cycle, and because pricing, campaigns and seasonality move the same numbers. That is why it calls for a comparison group or, at the very least, a baseline covering several months. In a B2B sale with a 90-day cycle, expect an effect on conversion from the third month and on renewal only in the following cycle.

Layer	Area	Indicator	Source	When to expect movement
1. Adoption and use	Sales	Meetings with agent summary; drafts approved without edits	CRM, activity log	2 to 4 weeks
1. Adoption and use	Marketing	Share of sends personalized by the agent	Automation tool	2 to 4 weeks
1. Adoption and use	Support and CS	Tickets classified by the agent and analyst corrections	Help desk integrated with the CRM	2 to 4 weeks
2. Efficiency	Sales	Time from meeting to next step; proposals stalled over 15 days	CRM, activity and stage dates	4 to 8 weeks
2. Efficiency	Marketing	Campaign production time; leads accepted by sales	Automation and CRM	4 to 8 weeks
2. Efficiency	Support and CS	First response on urgent tickets; renewal preparation	Help desk and calendar	4 to 8 weeks
3. Results	Sales	Proposal-to-win conversion; revenue per salesperson	CRM and finance	1 to 2 sales cycles
3. Results	Marketing	Personalized group versus standard; opportunities generated	Automation and CRM	1 sales cycle
3. Results	Support and CS	Cost per resolved ticket; renewal rate	Help desk and finance	1 renewal cycle

Layers 2 and 3 depend on a base that records date, stage and source with discipline. If the team moves deals between stages to make the weekly report look better, the sales cycle the CRM calculates won’t work as a baseline, and the ROI that comes out of it becomes an estimate dressed up as a number. That is the subject of how to make data-driven decisions, and the AI uses that already deliver in each area are in AI in sales, marketing and CS.

How much does AI really cost?

The vendor’s proposal shows the license. The company’s results also feel the credits consumed per agent action, the implementation, data cleanup, training and the time people spend reviewing what the agent produces. Gartner predicted in 2025 that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. The first two reasons show up when a company buys without a business case and discovers the real cost in month four.

Cost item	What goes in	Nature	Who usually forgets it
License or seat	AI-enabled plan, extra seats, edition upgrade	Recurring	Nobody, it is the line in the proposal
Credits per agent action	Consumption per summary, personalized email, ticket classification	Recurring and variable	Whoever estimated from pilot volume
Implementation	Agent setup, company context, approval rules, integrations	One-time	Whoever thought switching it on was enough
Data cleanup	Deduplication, standardized fields, reliable dates and stages	One-time, with upkeep	Whoever didn’t look at the base first
Training	Training for leaders, operators and the CRM administrator	One-time, with refreshers	Whoever counted only the vendor tutorial
Human review	Hours spent approving drafts, correcting classifications and auditing samples	Recurring	Whoever calculated hours saved without subtracting review

On platforms that charge per agent action, estimate consumption from the process volume. If support receives 400 tickets a month, triage runs 400 times, and if nurturing sends 20,000 emails, the coverage percentage the manager sets in the agent decides how much of that volume consumes credits. Human review also changes over time: in the first weeks the team checks everything, and once the correction rate drops, leadership can move to sample audits. Put that reduction in the plan with a date and a criterion, because it is what brings second-quarter cost below the first.

How do you build the AI business case before buying?

The business case is a two-page document that the executive team, finance and the area leader sign before the purchase. It fixes the process, the baseline, the hypothesis and the stop criterion, and it takes the project’s evaluation out of the hands of whoever likes or dislikes the tool.

Choose a process. Start with a high-volume process that works on text the operation already produces and has a clear owner, such as proposal follow-up or ticket triage. The reasoning for choosing the first uses is in AI for companies, and the design of each agent, with what it reads, what it executes and who approves, is in AI agents.

Record the baseline. Pull the three layers’ indicators from the CRM for the last three to six months, for the same team that will join the pilot. If the data doesn’t exist or isn’t reliable, the first stage of the project becomes recording it for four weeks before switching the agent on.

Write the hypothesis and the break-even point. Define how much use you expect in layer 1, how much the process should gain in layer 2 and what movement layer 3 needs to show to cover the full 12-month cost, with one-time items amortized over the period.

Convert to money through margin. The gain that goes into the ROI is the contribution margin on new revenue or the cost the company no longer incurs, and finance validates that conversion before the pilot.

Define the stop criterion. Write down what makes the project expand, adjust or end on day 90. Shutting down an agent that didn’t pay for itself is also good management, and with the criterion in writing nobody needs to defend the tool out of pride.

Illustrative example: proposal follow-up agent

The numbers below are hypothetical, built to show the math, and don’t represent a market average or a client case. Picture an operation with ten salespeople sending 100 proposals a month, with 22% proposal-to-win conversion averaged over the last six months, an average deal size of R$ 15,000 and a 40% contribution margin. The executive team wants an agent that lists stalled proposals and writes a follow-up for each one for the salesperson to approve.

Business case line	Hypothetical monthly value
Licenses and extra seats	R$ 2,500
Agent credits	R$ 1,500
Implementation (R$ 24,000 amortized over 12 months)	R$ 2,000
Data cleanup (R$ 12,000 over 12 months)	R$ 1,000
Training (R$ 12,000 over 12 months)	R$ 1,000
Human review (10 salespeople, 2 hours a month, R$ 80 an hour)	R$ 1,600
Total cost	R$ 9,600
Conversion rises from 22% to 24%: 2 more deals	R$ 30,000 in revenue
Contribution margin on those 2 deals	R$ 12,000
Monthly result (margin minus cost)	R$ 2,400

Calculated on revenue, the project seems to return more than three times its cost, and that is the number that usually goes into the presentation. Calculated on margin, the return is 25% over total cost for the month. If conversion rises only one point, to 23%, the additional margin falls to R$ 6,000, below the R$ 9,600 cost, and the project loses money. Break-even sits at 1.6 additional deals a month, or 23.6% conversion, and that is the number the executive team tracks at the 90-day review.

The hours-saved math would give a different number. If the vendor estimates four hours per salesperson per week, that is 160 hours a month, or R$ 12,800 at R$ 80 an hour, a return higher than the margin-based one that puts no money in the bank. Those hours enter the case when leadership turns them into more proposals per salesperson or a postponed hire, and the effect shows up in layer 3.

How do you run the 30, 60 and 90-day reviews?

The review is a 45-minute meeting with the area leader, finance and the CRM administrator, over the same dashboard at each milestone. The dashboard lives inside the CRM, with the three layers’ indicators and credit consumption, so nobody brings numbers from a private spreadsheet.

Milestone	Review question	What to look at	Possible decisions
30 days	Is the team using AI in the chosen process, and is the agent getting it right?	Layer 1: use, edit and correction rates, credits versus cap	Adjust context and instructions, fix fields, revisit coverage
60 days	Did the process get faster or gain volume?	Layer 2 versus baseline; actual cost versus forecast	Expand coverage, switch full review to sampling, revisit the credit cap
90 days	Did results move enough to cover the full cost?	First signs in layer 3; distance to break-even	Scale to other teams, adjust and run another cycle, or shut it down

Demanding revenue on day 30 leads leadership to cancel projects that haven’t yet had time to change the process. On day 60, if the process hasn’t changed against the baseline, the cause usually lies in the context the agent reads or in a team that approves without checking. On day 90, layer 3 starts to speak in short-cycle operations, and in long-cycle ones the decision rests on layer 2 and the break-even projection, with a new review scheduled for the end of the first sales cycle.

What are the common mistakes in measuring AI?

Measuring the pilot with the best team

Starting with the most engaged team makes adoption easier and skews the comparison. If the pilot runs with the three best salespeople and the baseline is the company average, the measured gain mixes AI with the talent of the people involved. Compare the pilot team with itself before the pilot, or set up a comparison group with a similar profile.

Comparing periods with different seasonality

Comparing December with October, or a big campaign month with an ordinary one, produces gains and losses that AI didn’t cause. Use the same period of the previous year when the business has strong seasonality, or a group that doesn’t use the agent during the same interval.

Switching the indicator midway

The project starts out promising conversion, conversion doesn’t rise, and in the day-90 presentation the indicator becomes salesperson satisfaction. The switch may have a legitimate reason, as long as leadership records it as a new hypothesis, with its own baseline, separate from the evaluation of the original project.

Crediting AI with gains that came from the cleanup

For the agent to work, the company standardizes fields, defines stage criteria and writes down its ICP, and part of the gain comes from that cleanup. That gain is a return on the project and belongs in the case, but the executive team needs to know its origin to decide on the next investment. In some projects the biggest return was in the process, and AI gave the company the reason to organize it.

How do HR and L&D measure the effect of AI training?

HR recognizes the need for training in layer 1 signals. Paid licenses with use concentrated in a few people, a high edit rate on drafts and leaders who can’t say which process AI should change indicate that the team received the tool without receiving the method. The satisfaction survey at the end of class captures none of these signals.

The training that sustains ROI involves different audiences. The executive team and finance learn to build and review the business case, and the sales, marketing and CS leaders define the indicators for each layer and the approval rules for the agents. Professionals in each area practice on the processes they will use, and the CRM administrator or the RevOps team learns to configure agents, control credits and build the review dashboard. The content and format of that track are in AI training for companies.

To measure the effect, HR uses the project’s own indicators. Compare layer 1 in the four weeks before and the four weeks after training, with the share of the team using AI in the chosen process and the edit and correction rates.

AI’s return comes from the company’s ability to choose the process, record the baseline and run the reviews without changing the yardstick halfway. That is the work of the in-company applied AI training, built on the company’s process and data, with cohorts for executives, leaders and operations. It delivers the prompt library for each area and the agent architecture, with what each agent reads, what it executes on its own and who approves the rest, along with the indicators leadership will use to measure the return.

Frequently asked questions

How do you calculate the ROI of artificial intelligence?

Choose a process, record the baseline in the CRM and add up the full 12-month cost, including license, credits, implementation, data cleanup, training and human review. The gain comes in as contribution margin on new revenue or as avoided cost, and ROI is that gain minus the cost, divided by the cost.

How long does it take for AI to pay off?

Use and quality show up within two to four weeks, and process efficiency between weeks four and eight. Revenue or cost results depend on the business cycle, and in a B2B sale with a 90-day cycle the effect on conversion starts to appear from the third month.

Which AI metrics should the executive team track?

One metric from each layer for each process: use in the process with the edit rate, an efficiency indicator against the baseline and the business case’s financial indicator with the break-even point. Credit consumption against the cap goes on the same dashboard.

Are hours saved a good way to measure AI results?

They work as a layer 2 efficiency indicator. An hour saved turns into money if leadership uses it for more volume per person or to postpone a hire, and that effect needs to show up in revenue or cost to count toward the return.

How much does an AI agent in the CRM cost?

It depends on the platform, the process volume and the complexity of each action, because many platforms charge variable credits per agent run. On top of consumption, add implementation, data cleanup, training and human review hours, which in the first quarter tend to weigh more than the license.

How does HR measure the results of AI training?

Compare in-process usage indicators before and after the training, on the same team: the share of people using AI in the chosen task, the edit rate on drafts and the correction rate on classifications. If those numbers move, the training changed behavior, and the results layer shows over the following months whether that reached revenue.