How to calculate the ROI of an AI agent — a guide for CEOs and CTOs
Calculating the ROI of an AI agent isn't subtracting costs from revenue. It's identifying the person-hour saved, the lead not lost and the error not made.
If you're considering adopting AI in your company, the executive committee's first question is always the same: how much will it cost me and how much will I get back? This guide gives you the framework to answer with concrete numbers, not promises.
The most common mistake when justifying an AI project is comparing it to traditional software. AI has different operating costs (tokens, compute, drift monitoring) and creates value in ways a spreadsheet doesn't capture: service hours saved, errors avoided, leads that didn't go cold, faster decisions.
The 3 components of AI ROI
The return of an AI agent is made up of three families of value. Ignoring any of them understates the business case.
1. Direct savings in operating cost
These are the person-hours a task stops consuming. If a person took 15 minutes to draft a quote and the agent does it in 20 seconds without intervention, you're freeing up ~14.5 minutes per quote. Multiply by monthly volume and by the employee's hourly cost (salary + benefits + overhead).
2. Enabled revenue
These are sales that didn't exist before. An agent that responds in 3 minutes where you used to take 24 hours captures leads that would have gone cold. Quantify it with your historical conversion rate and your average ticket.
3. Avoided risks
Human errors cost money: a miskeyed invoice, an out-of-range quote, an inconsistent answer to a customer. Estimate frequency × average cost of each error. In regulated processes (finance, healthcare, legal) this component can be the largest part of the ROI.
The real costs, no fine print
For the ROI to be honest, the other side of the equation must include everything — not just the initial development.
- Initial implementation: discovery + development + deployment (4-12 weeks depending on complexity)
- LLM tokens: a variable, usage-based cost. A typical customer service agent consumes $300-2,000 USD/month in tokens
- Infrastructure: hosting, observability, storage. Usually $200-1,500 USD/month
- Maintenance: improvements, drift monitoring, retraining — budget 15-25% of the initial cost per year
- Opportunity cost of the internal team: your PM and your technical people will dedicate time to the project
A practical payback formula
The formula is simple:
A well-applied AI project in mid-sized Mexican companies usually pays for itself in 3-8 months. Projects with ROI beyond 12 months are a red flag — the use case probably wasn't chosen well.
A real example: the quoting agent at ABSA
In the ABSA case (see our portfolio) the AI agent handles quote requests for industrial automation projects. Before: 24-hour response time, the sales team drafting manually, leads going cold.
| Variable | Before | After | Impact |
|---|---|---|---|
| Response time | 24 h | 3 min | −99% |
| Qualified leads/month | 120 | 450 | +275% |
| Conversion rate | 8% | 22% | +175% |
| Sales rep hours/month on quotes | 80 h | 12 h | −68 h freed up |
With an average ticket of $180k MXN and 30 additional closed deals per month, the agent enables revenue of ~$5.4M MXN per month. Payback was 6 weeks after launch.
What NOT to include in your ROI
Avoid inflating the business case with promises you can't measure. Don't include:
- "Improved customer experience" without a before/after NPS
- "Better decision-making" without specific KPIs
- "Digital transformation" — it's not a benefit, it's a consequence
- Future value from use cases you haven't prioritized yet
A credible ROI is always conservative on benefits and generous on costs. If the project still makes sense with that calculation, it's a good project.
When ROI isn't the right metric
For some projects — regulatory compliance, security, differentiating capability — ROI doesn't capture all the value. If your competitors already operate with AI and you don't, your business case is staying competitively viable, not a direct EBITDA increase. Make it explicit with leadership: sometimes the right number is the cost of NOT doing it.
What you may also be wondering
What is a typical ROI for an AI agent in Mexican companies?
In well-chosen projects, payback is between 3 and 8 months. The 12-month ROI is usually between 200% and 500%. Projects with payback >12 months should be reconsidered.
Do I need outside consultants to calculate this?
Not necessarily — the formula is simple if you know your operating costs and volumes. But an outside consultancy can give you a second pair of eyes that challenges optimistic assumptions.
How do I justify an AI project without historical data?
Run a 4-6 week pilot with a narrow scope and clear KPIs. The pilot generates the data that justifies the full project. Avoid asking for full budget approval without a prior pilot.