AI automation ROI is often decided with back-of-the-envelope math. Someone takes the order volume, multiplies it by the time each one takes, estimates how much the machine will handle and that is it, project approved. The trouble usually shows up months later. The licenses arrive, then the integrations, then the tasks that bounce back to the team, along with that feeling that the promised savings got lost somewhere along the way.
A more sensible way to look at this comes from an unexpected place: a car factory. In an opinion piece published in Mexico Business News, Miguel Verduzco, COO of Cody, a company that builds AI agents to support salespeople, serve customers and coordinate workflows between systems, explains that he applies to artificial intelligence the same logic he used back in 2014. At the time, while the Audi plant in San José Chiapa, Mexico, was being built, he devoted his engineering thesis at RWTH Aachen to a single question: was it worth automating the distribution center for engines and transmissions, or should everything run manually?
On one side, there was an option with a high upfront investment in equipment. On the other, a cheaper choice at the start, but one that leaned much more heavily on labor. Technology has changed a lot since then. The right way to make an investment decision, not so much.
Here you will see how to calculate AI ROI in a simple and honest way. That means putting the full cost into the math, along with the return that actually turns into money and the time the tool needs before it pays off. That way, you can judge AI project feasibility before you spend, not after.
Why asking how much AI will save is the wrong question
Almost every project starts with a rosy estimate. The example shared by Mexico Business News will sound familiar to anyone who has put together a proposal: the company gets 10,000 requests a month, each one takes five minutes, AI could automate 70% of them, so the savings would be this much. Sounds logical. But, as the author himself puts it, that is a hypothesis, not an investment assessment.
The suggested fix is to compare two concrete scenarios. The first is the current process, projected over the next three years, with growth, new hires and every inefficiency that already exists today. The second is the process redesigned with AI, where agents carry out part of the work and people handle the rest. The word that matters there is redesigned.
That is because AI rarely replaces an entire process. It changes how work is split between people, systems and agents, a point that ties in closely with the conversation about using AI for collaboration, not just automation.
There is also a detail that often slips through the cracks. Both scenarios need to deliver a comparable result. If the AI version handles more volume or offers a different level of service, that difference has to show up in the math. Otherwise you are comparing technologies, not ways of running the business.
Picture, just as an illustration, a clinic that currently books appointments by phone during business hours. If the AI version also starts booking at night and on weekends, it is not doing the same job for less. It is delivering something else. That can be great, but it needs to go into the analysis as a separate gain, not get mixed in with savings.
Where the real return comes from
Not every operational improvement turns into cash in the bank. This may be the most misunderstood part of automation return on investment. According to the article, economic value usually comes from three different mechanisms, and each one calls for a different kind of measurement.
The first is operational efficiency. An agent might save 4,000 hours of work, as the author illustrates, but that does not automatically wipe out the matching payroll cost. Saved hours only become money when they have a clear destination: avoiding new hires as demand grows, cutting overtime, ending duplicated work or moving the team to tasks with measurable results. If the company cannot explain how it will capture that gain, the honest thing is to call it a productivity gain, not cash savings.
This matters even more when team size stays the same. Freed-up time can have value, sure. But that value depends on what people do with it and on whether the company measures the outcome.
The second mechanism is sales performance. When an agent helps salespeople prioritize opportunities, broaden product coverage or run promotions, the benefit is not time saved. It is the extra contribution margin that comes from better execution. More revenue, on its own, is not a return.
The third is quality and risk reduction. Fewer errors, less rework and fewer customer service failures also create value, as long as the frequency and financial impact of those problems can be estimated with historical data.
How this shows up in the daily life of a Brazilian company
Imagine an online store that puts an agent in place to answer delivery time questions on WhatsApp. If the support team stays the same size and nobody knows what to do with the free time, the gain is real, but it never shows up in the cash flow. Now, if the store was planning to hire extra help for Black Friday and no longer needs to, that is concrete savings, with a date and an amount you can check.
On the sales team, the logic is similar. Think of a distributor that uses an agent to organize its accounts and remind each salesperson to offer the right product to every customer. That tool only pays for itself if the margin on the extra sales covers its cost. Bringing in more revenue by selling low-margin items can even make the bottom line worse.
In marketing, the same reasoning applies to campaigns and content production, something we have already covered when talking about AI and automation in digital marketing.
The rule of thumb, which the author sums up well, is to measure what changes in the business, not just what the technology can do. A slick demo does not pay the bills.
The full cost almost nobody puts in the spreadsheet
In industrial automation, the heavy money goes into machines built to last for many years. With AI agents, the upfront investment can be much smaller, but recurring costs become decisive, according to the Mexico Business News article. That is where a lot of the math falls apart.
The upfront investment covers more than it seems: process design, setup, integrations with company systems, data preparation, testing, rollout, team training and change management, which simply means getting people to actually use the tool. Then come the costs that keep repeating:
- platform licenses, model usage and API consumption;
- maintenance, monitoring and quality control;
- human oversight and handling the cases the agent hands off;
- governance, meaning someone keeping an eye on the rules, limits and behavior of the agent.
An AI agent is not something you set up once and forget. Its behavior, cost and effectiveness need to be tracked and tweaked all the time.
There is also a hidden cost the author makes a point of highlighting. When the agent tries to handle a task, fails and passes it to a person, the company may end up paying twice: once for the AI processing and once for the human work. Ignoring those failed attempts skews the entire calculation.
That is why the metric that matters is cost per successful outcome, not cost per message. Imagine a support agent that replies fast but needs three or four rounds of back-and-forth with the customer and, in the end, still transfers the case to a human rep. Each message was cheap. The full interaction, not so much.
Adoption, coverage and resolution: the math that changes everything
An agent almost never delivers the promised performance on day one. Results depend on how many people use it, on the quality of the information, on the integrations and on whether the team works the tool into its routine. To avoid fooling yourself, the article suggests separating three variables.
Adoption shows how much of the target activity actually runs through the AI workflow. Coverage tells you how many of those cases fall within what the agent was built to handle. And resolution shows how many of the eligible cases end correctly, with no one needing to step in.
Multiply the three and you get effective automation. The example in the article makes this very clear: 60% adoption, 80% coverage and 90% resolution add up to 43.2% effective automation of the target activity. A far cry from the 90% that usually lands on the slide for the executive team.
The author also points out that an agent can resolve 90% of the cases it handles and still have limited impact if only 30% of the organization uses it. In practice, this happens a lot when the team is wary of anything new. Picture a company where half the sales reps would rather stick to the old way. The agent might be excellent, but it will be running at half its potential.
That is why it pays to include a realistic learning curve. As an example, Mexico Business News suggests assuming something like 40% of expected performance in the first quarter, 70% in the second and a steady level after that, always adjusting based on evidence from your own process. The goal is not to nail the forecast. It is to avoid building a business case that assumes perfection from launch day.
Simple payback vs. discounted cash flow
For the author, this may be the clearest lesson he brought over from his industrial experience. He cites a type of proposal that still shows up a lot: implementation costs US$20,000, monthly savings are US$5,000, so the investment pays for itself in four months. It is easy to explain in a meeting. When it comes to deciding where to put your money, though, it falls short.
The main reason is that money has a time value. A dollar today is worth more than a dollar a year from now, because you could invest it, pay off a debt or use it somewhere else in the business. Math this lean can also hide recurring costs, the learning curve and the risk that the savings take a while to show up.
When you lay out inflows and outflows month by month, or quarter by quarter, and discount the cost of money, the picture gets more accurate. Sometimes less exciting, too. And that is fine: better to find out in the spreadsheet than on the bank statement.
When it is time to take the numbers to the executive team or your business partner, a good practice is to present a base case, a pessimistic case and an optimistic case. Show the cash flow over the three years, spell out your assumptions on adoption and resolution and explain where each gain comes from. Transparent math tends to earn more trust than a percentage that looks a little too neat. If capital allocation is on your mind, it is also worth seeing how AI and automation have been driving investments.
Step by step to calculate the ROI of your AI project
The article by Verduzco brings together eight lessons in total. Putting together the ones we have covered so far, you can build a short roadmap that works for most AI automation projects for businesses, from customer service to finance. You do not need a complex spreadsheet, just the discipline not to skip steps.
- Define the two scenarios: the current process projected over three years and the process redesigned with AI, delivering a comparable result.
- Map the value: state where each gain comes from (efficiency, sales margin or fewer errors) and how it turns into money.
- Add up every cost: implementation, recurring costs and the cost of tasks that go back to people.
- Apply adoption, coverage, resolution and the learning curve: use effective automation, not the success rate of the agent.
- Project the cash flow: spread inflows and outflows over time and factor in the cost of money.
Some mistakes inflate the result without anyone noticing. Counting saved hours as saved money is the classic one. Treating the resolution rate as if it were total automation is another. Forgetting human oversight, monitoring and the cost of handoffs happens too, and so does assuming the tool will deliver everything it promises right out of the gate.
In smaller businesses, with little cash to absorb a mistake, each of these slip-ups hits harder. If that sounds like you, check out this piece on AI and business automation in startups, which helps you figure out where to start.
The logic that connects factories and AI agents
The story Verduzco tells carries a lesson that goes beyond the numbers. In 2014, the job was not to prove that automation was technologically superior. It was to figure out when it made economic sense, taking into account volumes, operating costs, depreciation, cost of capital and uncertain demand. With AI, the question is the same. Technology moves fast, but the logic behind a good investment decision still holds up.
In practice, a prudent path is to start with a small process that is easy to measure. A specific customer service flow, one stage of the sales process, a repetitive task in finance. That way you gather real data on adoption and resolution, adjust the math based on evidence and only then scale up.
An ambitious project is not the problem. The risk lies in not knowing exactly what you are measuring, how much it really costs and how long it takes to pay off.
Before you sign your next AI contract, pick one process in your company and run it through the five steps above. If the math works even in the pessimistic case, you have a solid candidate on your hands. And if you want to dig deeper, here at Método Viral you will find more content on use cases and best practices for AI in business.
