"Look how it's climbing! The agent is paying for itself." Carlos said this while looking at a dashboard that started counting the same day Hora came on board. Nobody measured the before. In this lesson you build the monthly budget with a cap and the estimated ROI in three scenarios, and you practice the conversation where the owner asks for a savings number and you answer without making promises.
1The month's budget with a cap, an alert, and a limit per task
2Estimating the benefit when the baseline was never measured
3Estimated ROI in three scenarios, with the investment kept separate
4Why freed capacity is not savings
Goal: build the monthly budget and the estimated ROI of the operation in three scenarios, with uncertainties written down and no promised savings.
01 · From budget to estimated ROI
Five steps. The last one is a decision, not a number. Click each one.
Next month, what was measured replaces what was estimated
Click each block in the diagram to see what it does.
Formulas from the Gestoria method (INEMA): potential benefit, estimated net result, and estimated ROI, with the initial investment kept separate from the recurring one. Horizonte: fictional study company and values.
02 · How to build and defend the numbers
Four steps. At the end, the whole Horizonte calculation, so you can check it line by line.
1. Budget with a cap
Model cost is variable: it rises with volume, with long conversations and with the top-tier model. That is why it has a monthly cap and an alert before the cap, agreed with IT and, if possible, set up with the provider.
Also set a limit per task: number of attempts and the maximum cost of one conversation. If it goes over, the agent stops and calls a person. Human review is also a governance cost, and it grows: Gartner forecasts worldwide spending on AI governance rising from US$ 492 million in 2026 to more than US$ 1 billion in 2030 (forecast, 02/17/2026).
At Horizonte (fictional)Model: average R$ 1,100, cap R$ 1,400, alert at R$ 1,120 (80%). Per conversation: up to 2 attempts and R$ 1.50; above that, the conversation goes to Juliana.
2. Baseline as a range
Without the "before," any savings number is a guess. Rebuild the baseline with the people who did the work: how long did a WhatsApp order take before Hora? Ask for the shortest and longest reasonable time, not an average.
Write down where the range came from and start measuring now, over a comparable period. A month from now, part of the uncertainty becomes a measurement.
At Horizonte (fictional)Juliana and the salespeople: 8 to 12 minutes per order, most likely 10. With 680 orders and R$ 45 an hour, the human cost before is between R$ 4,080 and R$ 6,120.
3. Three scenarios
Human cost after: Patrícia R$ 720 + Juliana R$ 1,350 + rework R$ 432 = R$ 2,502. Potential benefit: R$ 1,578, R$ 2,598 or R$ 3,618. Minus the month's cost (R$ 2,040): −R$ 462, +R$ 558 or +R$ 1,578.
Estimated ROI: −23%, +27% and +77%. The R$ 3,600 setup cost pays back in about 6.5 months in the likely scenario and never in the pessimistic one. This gets written down, along with the numbers.
ReadingThe most likely result is positive and small; the pessimistic one is negative. The honest sentence is "it probably pays for itself, and we don't know yet," not "it is paying for itself."
4. Capacity is not savings
In the likely scenario, Hora frees up about 63 hours of people's time per month. That is capacity. It only becomes savings if the company cuts an expense or uses the hours on something with a result you can measure.
That is why the selection report separates "potential benefit" from "realized savings" and leaves the latter blank until there is evidence. A simulation is not presented as revenue or as savings.
At HorizonteJuliana's hours went to registering new customers and after-sales. Realized savings: not yet observed. Patrícia tracks for a month where the hours went before any other decision.
03 · Template to copy: budget and estimated ROI
A page that finance and the owner read. The tab next to it shows Horizonte's.
Monthly budget + estimated ROI
BUDGET AND ESTIMATED ROI · <process> · <month/year> · owner <name>
Values: [ ] measured [ ] estimated (mark line by line)
COST OF THE MONTH (recurring)
| Line | Average | Cap | Alert | Measured or estimated |
|---|---|---|---|---|
| AI model | R$ <...> | R$ <...> | R$ <...> (80%) | <...> |
| Tools | R$ <...> | - | - | <...> |
| Maintenance | R$ <...> | - | - | <...> |
| TOTAL | R$ <...> | | | |
LIMIT PER TASK: <n> attempts · R$ <...> per <conversation/case> · over the limit -> <name>
INITIAL INVESTMENT (separate): R$ <...> (<what it was>)
POTENTIAL BENEFIT = human cost before - human cost after
Baseline: <min> to <max> min per task · source: <who/how> · measured? [ ] yes [ ] no
Human cost after: review R$ <...> + exceptions R$ <...> + rework R$ <...>
PESSIMISTIC LIKELY OPTIMISTIC
Human cost before R$ <...> R$ <...> R$ <...>
Potential benefit R$ <...> R$ <...> R$ <...>
Net result R$ <...> R$ <...> R$ <...> (= benefit - cost of the month)
ESTIMATED ROI <...>% <...>% <...>% (= result ÷ cost of the month)
Investment payback: <months> in the likely case · <...> in the pessimistic case
FREED CAPACITY: <h>/month (likely) · destination: <...>
REALIZED SAVINGS: <evidence or "not yet observed">
UNCERTAINTIES: 1. <...> 2. <...> 3. <...>
DECISION: <keep / expand / reduce / pause> · REDO THE CALCULATION ON: <date>
Prompt · where the ROI is too optimistic
You are the most skeptical chief financial officer in the company. Below is
the budget and estimated ROI of an operation run by an AI agent. Do not redo
the calculation: point out where it may be too optimistic.
1. What cost is missing (maintenance, review, rework, switching
provider, higher volume)?
2. Where does the baseline come from? Was it measured or remembered?
3. Is any freed capacity being treated as savings?
4. Was the initial investment kept separate?
5. What question would I ask before approving this?
Document:
"""<paste>"""
Simulation: illustrative example
Hora · WhatsApp order · Oct/2026 · Horizonte, fictional study company · owner: Patrícia
Cost of the month: model R$ 1,100 (cap R$ 1,400, alert R$ 1,120) + tools R$ 400 + maintenance R$ 540 = R$ 2,040. Limit: 2 attempts and R$ 1.50 per conversation.
Initial investment: R$ 3,600 (40 h of Rafael).
Baseline: 8 to 12 min per order, estimated with Juliana and the salespeople · not measured.
Investment payback: ~6.5 months in the likely case; it does not pay back in the pessimistic case. Freed capacity: ~63 h/month (likely). Realized savings: not yet observed. Decision: keep and measure the baseline for 2 weeks; redo on 11/30.
A single number. "27% ROI" without the range hides that the pessimistic case is negative.
Capacity added up as savings. Hours freed are not money in the bank.
Investment mixed in. Adding the setup cost to the month drags the ROI down; forgetting it inflates the return.
Baseline from memory, unstated. You can use an estimate, but written as an estimate, with who gave it.
Check: before − 2,502 = benefit; benefit − 2,040 = net; net ÷ 2,040 = ROI. All three Horizonte scenarios have to add up.
04 · Simulation: Carlos wants a number
A simulated conversation for practice. The owner is going to the quarterly results meeting and wants to know how much the agent saves. You answer with the calculation from this lesson.
Carlos · owner of Horizonte (simulation)Horizonte Distribuidora, fictional study company
Choose your answer. The comment appears right after, and the conversation continues based on your choice.
05 · Practice
Horizonte's budget to train on, yours for real, and the conversation with the owner to rehearse.
1 · Horizonte's budget
25 min
Fill in the template with Horizonte's numbers, line by line, marking what is measured and what is estimated. Check the three columns against the formula. Run the skeptical CFO prompt and write three uncertainties and the decision.
Done when: the document has the monthly cost with a cap and an alert, the investment separate, the baseline as a range with its source, the three scenarios adding up (−R$ 462, +R$ 558, +R$ 1,578), "realized savings: not yet observed", three written uncertainties, and the date to redo it.
2 · Your process's budget
45 min
Use the math from lesson 1 for your process. Build the budget with a cap, estimate the baseline as a range with the person who did the work, calculate the three scenarios, and write the decision. Show it to someone in finance before taking it to the owner.
Done when: your document has a cap and an alert for the model, a limit per task, a separate investment, the three scenarios with the formula checked, at least three uncertainties, the freed capacity with a destination, and the date when the estimate becomes measured.
3 · The conversation with Carlos
10 min
Run the simulation above to the end. If you don't reach the top score, start over and see what changes with each answer. Then write, in three sentences, how you would answer the same question about your own process.
Done when: you reached 8 out of 8 points and have the three sentences written, with a range of values, the word "estimated", and what will be measured.
5 questions. The explanation appears after your attempt.
Question 1 of 5
What is the estimated ROI formula used in the course?
Correct
The benefit comes from the human cost before minus after; the net result subtracts the monthly cost. The initial investment goes into the payback time, not mixed into the month.
Not that one
The benefit comes from the human cost before minus after; the net result subtracts the monthly cost. The initial investment goes into the payback time, not mixed into the month.Review: From budget to estimated ROI
Question 2 of 5
The baseline was never measured. What should you do?
Correct
Waiting blocks the decision; making it up misleads. The written range, with its source, lets you decide now and correct later, when the measurement arrives.
Not that one
Waiting blocks the decision; making it up misleads. The written range, with its source, lets you decide now and correct later, when the measurement arrives.Review: How to build and defend the numbers
Question 3 of 5
In the likely scenario, Hora frees up about 63 hours per month. Is that savings?
Correct
Freed hours that nobody uses don't show up in the cash flow. The selection report leaves "realized savings" blank until there is evidence.
Not that one
Freed hours that nobody uses don't show up in the cash flow. The selection report leaves "realized savings" blank until there is evidence.Review: How to build and defend the numbers
Question 4 of 5
Why present three scenarios instead of one number?
Correct
With a single number (+27%), no one sees that the pessimistic case gives −23%. The range turns the question "how much does it save?" into "what do we need to measure to find out?".
Not that one
With a single number (+27%), no one sees that the pessimistic case gives −23%. The range turns the question "how much does it save?" into "what do we need to measure to find out?".Review: From budget to estimated ROI
Question 5 of 5
The model cost blew past the month's cap because of long conversations. Which control was missing?
Correct
A monthly cap with no per-task limit only warns you after the damage is done. With a per-conversation limit, the expensive case stops and goes to a person; with an alert at 80%, there's time to act.
Not that one
A monthly cap with no per-task limit only warns you after the damage is done. With a per-conversation limit, the expensive case stops and goes to a person; with an alert at 80%, there's time to act.Review: How to build and defend the numbers
Quiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.
Pick one answer. The explanation shows up after you try.
07 · Summary
Budget with a capVariable-cost model with a cap and an alert; per-task limit; investment kept separate.
Baseline as a rangeEstimated with the people who did the work, written down as an estimate, measured from now on.
Three scenariosPessimistic, likely and optimistic estimated ROI, with the uncertainties written down.
No promisesCapacity is not savings. Realized savings only with evidence.
