Robot or human: the triggers that define when to transfer the conversation

What should never be automated

Price negotiation and discount. Any conversation about value requires context reading that automation does not do.

Objection. "I found it expensive," "I’ll think about it," "the competitor does it for less." A canned objection response sounds exactly like what it is.

Complaint. An unhappy customer who receives an automatic menu becomes more dissatisfied, and the cost is not the conversation: it’s the public review that follows.

Sensitive situation. Health, mourning, financial difficulty, legal issues. No savings justify the risk.

Exception. Any request that deviates from the standard. Automation only knows how to answer what was foreseen, and the customer immediately notices when it wasn’t.

Closing. The last mile of the sale is human. A robot trying to close produces abandoned carts with extra steps.

The six immediate transfer triggers

These are the signals that should take the conversation out of automation immediately, without completing the flow.

1. Explicit purchase intent. "I want to buy," "how do I contract," "send the payment link." Holding someone who has already decided is the most costly mistake of all.

2. Mention of price, discount, or condition. Any conversation about money goes to a person.

3. Explicit request for an attendant. "I want to talk to a person" must always work, at any point in the flow. Ignoring this request is the main cause of irritation in automated service.

4. Negative tone. Complaint, urgency, words of dissatisfaction. Transfer immediately and preferably to someone prepared.

5. Two unexpected responses. If the customer wrote twice something the flow didn’t understand, they won’t understand it the third time. Transfer.

6. High potential value. Known company, large volume, corporate order. Worth a priority path, even if the question seems simple.

How to transfer without making the customer repeat everything

Poorly done transfer destroys the gain from automation. Four precautions solve this.

Pass the context along. The attendant needs to see, before writing, everything the customer has already informed. If the salesperson asks again for the name and city, the customer rightly concludes the robot only served to waste their time.

Announce the transfer with name and timeframe. "I’m going to transfer you to Ana, who handles this. She will respond within 10 minutes." Silent transfer feels like abandonment.

Start the human conversation with what has already been said. "Hi, this is Ana. I saw you need X for next week, in Campinas. Let me confirm two things..." This proves there was reading and recovers any accumulated friction.

Have a plan for when no one is available. Queue with an estimate, not silence. "All attendants are busy. You’re next, and I should call you in about 12 minutes."

Automation outside business hours

This is where automation has the best return because it competes with no one: the alternative is not a human, it’s silence.

Outside business hours, the flow should do three things: acknowledge immediately, collect essentials for the next day, and inform a concrete return time. Nothing more — it’s not time to try to sell or navigate long menus, because the person is writing at night, probably to more than one company.

A technical detail that makes a difference in the official API: when the customer responds to an automation question, they keep the service window active, which facilitates resuming the next morning without depending on an approved template.

What changes with artificial intelligence in the middle

Menu flow and AI assistant are different things, and the distinction matters because it changes where the risk lies.

The menu flow errs by rigidity: it only knows what was foreseen and freezes in front of any question outside the script. The AI assistant errs by the opposite — it always answers, even when it doesn’t know. In commercial service, this means it can give a deadline that doesn’t exist, confirm a condition the company doesn’t practice, or invent a product feature. And the customer will hold exactly that against you later, with the conversation history in hand.

Three safeguards solve most of the risk. Closed knowledge base: the assistant answers only with information you provided and says it will check when the question goes beyond that. Declared prohibited topics: price outside the table, exceptional deadline, differentiated payment condition, and any promise of results automatically go to a human. Weekly sample reading: twenty conversations per week read by someone from the sales team, specifically looking for incorrect information.

Well configured, the AI assistant solves something the menu never did: understanding the customer’s written question as they wrote it, without forcing them to choose an option. It’s a real experience gain — as long as the design assumes it will err and that the error needs somewhere to go.

What to measure

Without these four numbers, it’s impossible to know if automation helps or hinders.

Abandonment rate in the flow. How many customers stop responding during automation. Above 25%, the flow is too long or confusing.

Time until human, when transfer happens. If the sum of automation and queue exceeds 15 minutes, automation is creating wait instead of eliminating it.

Comparative conversion. Compare the closing rate of conversations that went through automation with those that went straight to human service. If automation converts significantly less, it’s holding ready customers.

Requests for an attendant. If many customers ask to speak to someone right at the first interaction, the flow isn’t solving — it’s delaying.

Where to start without breaking what works

Those who have never automated anything tend to want to design the entire flow at once. The reverse path works better, in three two-week stages.

Stage 1 — just acknowledgment. A single instant automatic message, requesting two pieces of information. No menu, no tree. This alone solves the first response problem and creates no risk of holding customers.

Stage 2 — simple routing. Two or three clear options directing to the right sector, always with an exit to talk to someone. Measure abandonment before adding any step.

Stage 3 — factual answers and reminders. The most repeated questions, answered immediately, and scheduling and delivery confirmations sent automatically.

After these three stages, most companies discover they have captured almost all available gain — and that the elaborate flows they intended to build would solve little and cost conversations.

Mistakes that make automation cost money

Long menu in the first message. Five numbered options before any conversation scare away precisely those in a hurry.

No exit offered. A flow without the option to talk to someone is the design that generates the most public complaints.

Automating to save instead of to respond quickly. The justification changes the entire design. Automation that exists to cut costs tends to hold the customer; automation that exists to respond in seconds tends to release them early.

Never reading conversations. Reading twenty automated conversations per week reveals more about the flow than any report — and that’s where you find the points where the customer gave up.

Well-calibrated automation doesn’t replace the salesperson: it delivers to the salesperson an already identified customer, with the need described, in less time. When it starts holding conversations, it stops being efficiency and becomes an obstacle — and the difference between one and the other is entirely in the timing of release.