Choose a metric
The agent writes, sends, and makes recommendations. It isn’t accountable for a wrong price or a message sent to the wrong person. You respond.
To respond calmly, you need one number. Just one. And it has to come from the pain point you found in Track 2—not from what the agent does, but from what improves in the business.
🆕 New here? What is a “metric”?
Metric is a number you can count the same way every time that shows whether the pain has decreased. “Absences per week” is a metric. “Happier customers” isn’t—no one counts happiness on Friday.
How to read the diagram: the three numbers at the top come from the same pain point, but only the highlighted one at the bottom changes when the pain eases. "Messages sent" goes up even if no one shows up — it measures the agent's work, not your result.
✓ A useful metric
- ✓Move on from the pain point: "patients who showed up".
- ✓You can count it on the calendar, without relying on opinions.
- ✓It changes if the agent helps and doesn't change if it only does the work.
✗ Misleading metric
- ✗“Messages sent” — measures effort, not results.
- ✗“Better service” — nobody can count that.
- ✗Five numbers at once—one always goes up and makes excuses.
a countable number
not from the tool
no excuses
of counting, always
Write down today’s value
If you don’t write down today’s number, you won’t be able to prove anything two months from now. Not that things improved, and not that they got worse. You’ll be left with an impression—and an impression won’t hold up in front of a business partner.
O today's value is counted before of turning on the agent, looking back: the last month, the last four weeks. Once it’s on, the number is already contaminated.
🆕 New here? “Today’s value”
It's the metric's starting point, measured before any changes. Elsewhere you'll hear "baseline." Here, we call it today's value — because that’s what it is: how the business works today, without the agent.
Open the source
Renata opened last month’s schedule. She didn’t ask the front desk “about how many patients missed their appointments” — she counted.
Count the way you'll count later
38 appointments booked, 9 no-shows, 29 attended. Rescheduling with notice doesn't count as a no-show—and it will continue not to count.
Write it with a date
“Today’s value, recorded on September 1: 29 of 38 showed up.” On a sheet, in a document, somewhere you’ll find it later.
Only then turn on the agent
From here on, every comparison uses this line. If the way you count changes, the comparison no longer holds.
💡 Don’t have the number?
Marcos didn’t know how much time Júlia spent checking documents. Instead of guessing, he asked her to keep a log for a week: when she started and when she finished. A week of tracking is worth more than a year of guesses.
the starting point
of turning on the agent
written and saved
not estimated
Set a goal and deadline
With today's value in hand, two lines are missing: where you want to get to (goal) and until when (deadline). Without both, the agent stays on "test" forever.
The goal must be verifiable by counting. The deadline must be short enough for you to remember to check—30 to 60 days works for almost every small business.
How to read the diagram: the gray circle is today’s value; the highlighted circle at the top is the goal. The blue line isn’t a forecast—it’s the path you expect to take. The two dots mark when you stop and check: halfway through and at the end.
| Timeline | Renata (clinic) | Marcos (office) |
|---|---|---|
| Metric | patients who showed up | hours per day checking documents |
| Today's value | 29 of 38 | a whole morning |
| Goal | 34 out of 38 | one hour |
| Deadline | 60 days | 30 days |
What to look for in the table: the four lines in each column fit on a sticky note. If yours doesn't, there's probably more than one metric hidden in there.
💡 Bold goal or honest goal?
Zero missed appointments isn’t a goal; it’s a wish. Cutting missed appointments in half in 60 days is a goal that a well-made reminder can achieve—and if it doesn’t, you learn something about the pain point.
verifiable by counting
30 to 60 days
in the middle and at the end
four lines are enough
Name the owner
An agent can deliver a polished answer that's completely wrong. Someone needs to check it. If "everyone" checks, no one does.
O owner is a person with a name. It answers three questions: who reviews mistakes, how often, and when a human takes over the conversation from the agent.
🆕 New here? “Owner” isn’t whoever configured it
The person who turned on the agent may have been your nephew, the provider, or you yourself in the middle of the night. The owner is the person who checks its work every week and responds when something goes wrong. In a small business, that’s almost always you—but write down the name.
✗ No owner
Patient: "I got the reminder. How much does a facial cost?"
Agent: "A skin treatment starts at R$ 90."
No one gave the agent that price. It made it up — and the patient will complain at the front desk.
✓ With an owner and handoff rule
Patient: "I got the reminder. How much does a facial cost?"
Agent: "I’ll pass your question along to the front desk, who will get back to you later today."
The agent recognizes the limit and hands the conversation to a person.
| Question | Renata | Marcos |
|---|---|---|
| Who reviews? | the receptionist | Marcos himself |
| How often? | every morning, on the day’s schedule | per sample, on Fridays |
| When does a human take over? | price, complaint, or health question | question about taxes or a legal deadline |
⚠️ Attention
Don’t know everything that could go wrong? Start with three required checkpoints: price, complaint and anything involving health, money, or a legal deadline. Add the rest as it comes up in the log.
one person, by name
on a set day
when a human takes over
where the owner checks
Answer both questions
When the deadline arrives, you answer two questions. Did it work? — the number changed, counted the same way. Was it under control? — you can show what the agent did, who approved it, and what happened when it didn’t know.
Neither one adds up on its own. Results without control are luck. Control without results is a cost.
How to read the diagram: move right if the number changed; move up if you can show the record and approvals. Only the highlighted box in the upper right authorizes keeping the agent as is. The yellow box is the most dangerous: it looks like success, but no one knows why.
📈 Did it work?
The number that changed, counted the same way as at the start. Marcos: “Document checks went from taking all morning to taking one hour.”
🧾 Was it under control?
A record of what the agent did, who approved it, and what happened when it didn’t know. Marcos: "Two messages were held for approval. One question was passed on to me."
Quick test (optional): at the clinic, missed appointments were cut in half, but nobody knows what the agent said to the patients. What’s missing?
the number changed
can show
uncontrolled outcome
control without results
Build the one-line dashboard
Tracking doesn’t require a system. It needs one line per week, always with the same columns, filled out by the owner in five minutes.
Eight rows later, you have both answers ready: the weekly number answers "did it work?"; the error and handoff columns answer "was it under control?"
Paste into the AI chat you already use (ChatGPT, Claude, Gemini). Replace what's inside < >.
Quero acompanhar, uma vez por semana, o resultado de um agente de IA no meu negócio. Meu negócio: <ex.: clínica de estética com 2 funcionárias>. O que o agente faz: <ex.: envia lembrete e pede confirmação de horário pelo WhatsApp>. Métrica: <ex.: pacientes que compareceram>. Valor de hoje: <ex.: 29 de 38, contado em 01/09>. Meta e prazo: <ex.: 34 de 38 em 60 dias>. Dono: <nome de quem revisa>, revisa <ex.: toda sexta>. Um humano assume quando: <ex.: preço, reclamação, saúde>. Monte um painel de UMA linha por semana, em forma de tabela, com as colunas: semana | valor da métrica | erros achados na revisão | vezes que um humano assumiu | ações que esperaram aprovação | nota do dono Preencha só a primeira linha com o valor de hoje e deixe mais 8 linhas vazias. Depois da tabela, escreva em 3 frases como eu respondo, no fim do prazo: 1) funcionou? 2) estava sob controle? Não invente números: onde eu não informei, escreva "preencher".
💡 Renata’s line in week 3
Week 3 · 32 of 38 · 1 error (reminder with the wrong time) · 4 handoffs to reception · 0 sends without approval · “fix the calendar time zone.” One line. Five minutes. And the error became a task, not a surprise.
one line per week
on the owner’s calendar
become tasks
ready by the deadline
🎓 Module summary
Next module:
5.2 — The spreadsheet in 4 blocks