You automated, and now nobody answers

Every automation project starts with the same request: we want more customers writing in.

Almost nobody asks what happens the day they show up.

This is what we measured at an auto shop in Puebla where we nearly quadrupled inbound messages in five weeks. The technical side worked. What did not hold up was something else.

The numbers

Marcas de Puebla runs two locations and sells tires and auto service. It has a WhatsApp bot that quotes against real inventory, logs leads and follows up. In August we turned on a Google Ads campaign for them, and in September we added social media ads.

This is what happened to the messages coming into the shop's WhatsApp:

• August: 17.7 inbound messages a day.
• September 1 to 8: 67.6 a day. Almost four times as many.
• New conversations: from 4.3 to 10.8 a day.
• Leads logged: from 3.3 to 7.3 a day.

Seen that way, it is a success story. The shop went from three prospects a day to seven, and they are not weak prospects: of the August leads, 16.7% ended in a paid service order.

The problem shows up when you look at who was answering.

The number that set off the alarm

Throughout August, of every message that went out from the shop's WhatsApp, 1.3% was written by a person. The bot answered everything else. Nobody on the team was touching the inbox, and nobody needed to.

On September 7 and 8, half of the outgoing messages were already being written by a human.

And there is a second, more uncomfortable signal: the team started switching the bot off. In the entire history of the system there are 37 conversations where the bot was turned off by hand. Thirty-six of those 37 are from September.

Switching the bot off is not bad in itself: that is exactly what the option is for, so a service advisor can take over when the conversation calls for it. What matters is the change in pace. It went from being an exception to an everyday thing, in the same two weeks that volume spiked.

That is not a software problem. It is a team hitting its limit.

What volume does to response time

We measured how long the first reply written by a person took, counted from the customer's first message.

• Week of August 31: median of 15 minutes.
• September 7 and 8: median of 148 minutes.

To be honest about it: that last figure comes from just eight conversations. It is a signal, not a conclusion, and it needs another week of data to stand on. We include it because it points in the same direction as everything else, not because it closes the case.

But the business question does not need more precision than that. If someone asks the price of a set of tires and gets an answer two and a half hours later, they have already bought somewhere else or gone cold.

Automating demand does not automate the response

Here is the design mistake we see over and over, and it is not the client's: it belongs to whoever sells the automation.

A well-built bot handles the first contact. It answers at eleven at night, quotes against inventory, logs the prospect, leaves no message unread. That scales with practically no limit.

What does not scale is everything that comes after and still needs a person: settling the price on a bigger job, scheduling against the shop's real workload, deciding whether to repair or replace. That capacity does not grow because you turn on a campaign. It grows by hiring, training or reorganizing.

When you multiply the input by four and leave the output the same, you did not generate four times the sales. You generated a line.

What we did

The first thing was to lower the ad budget. The owner asked for it and he was right: not because of cost, but because his team was not handling the messages well. It is the least common conversation at an agency, and it is the right one.

Spending less right when the campaign is working sounds backwards. It is not. A lead nobody attends to is not a lead; it is a customer you taught that writing to you is pointless.

The second was to review the Google Ads campaign through that lens. We found that an automated recommendation, applied from a phone, had set a cost-per-conversion target at double the real one, and that was making every click 54% more expensive without bringing in a single extra prospect. It was reverted.

The third, and what comes next: align the hours the ads run with the hours when someone is able to answer. If the budget is already maxed out, scheduling does not reduce prospects. It only moves them to the time when someone is actually there to receive them.

How many conversations one person can handle

"Hire someone else" is an expensive answer and usually a premature one. It pays to put a number on it first.

The math we use has no mystery to it. A sales conversation that has already gone through the bot —the prospect arrived with their need identified and a price on the table— takes between 8 and 15 minutes of real attention spread across the day: reading, checking something with the shop, replying, and coming back a couple of times.

With that, one person who also has other responsibilities —because at a small business nobody handles WhatsApp full time— can sustain between 12 and 20 sales conversations a day without their response time degrading.

Apply it to your case: divide your new conversations per day by that range and you get how many person-shifts you need. If the result is higher than the number of people actually watching the inbox, you already know why your prospects are going cold, and you know it without arguing about it.

The exact number varies by industry. A shop with simple quotes can handle more; a consultative sale that runs several weeks, far fewer. What does not vary is that the number exists and almost nobody calculated it before turning on the campaign.

What the bot should absorb and what it should not

When capacity saturates, the reflex is to ask more of the bot. Sometimes that is right, and sometimes it is exactly what breaks the customer's trust.

The line that works for us is this: the bot keeps everything that has a verifiable answer, and the person keeps everything that involves a commitment.

• For the bot: prices against real inventory, availability of a given tire size, locations and hours, the status of an existing order, and logging the prospect along with what they asked for.
• For the person: committing to a date, authorizing a discount, deciding between repair and replacement, and any case where getting it wrong costs money or credibility.

The most valuable rule is often the one that prevents an answer. A bot that promises a date the shop cannot meet does more damage than one that says "an advisor will confirm that in the morning."

Before hiring anyone, there is almost always room to move things to the first list that sit on the second one today out of habit rather than judgment.

How to tell if this is happening to you

You do not need a dashboard to spot it. Three questions to run against your own data:

• What percentage of replies are written by a person, and which way is that curve heading? If it rises month over month, your automation is coming apart on its own.
• How many times was the bot switched off this month, compared with last month? It is the most honest thermometer there is: people switch it off when they feel it is in the way or when they cannot keep up.
• How long does the first human reply take, measured as a median and not an average? A single conversation answered three days later ruins the average. The median tells you how the typical case feels.

If you cannot answer those three questions with data, that is the first finding.

What this case reminds us

Response capacity is part of the system, even though nobody programs it. An automation project that does not account for it is not finished: it is just delivering its bottleneck one step further along.

That is why, when someone asks us for more prospects, the first question is no longer which channel. It is how many they can serve well today, with the people they already have.

The answer is usually a smaller number than they expected. Knowing it before turning on the campaign is far cheaper than finding out with a line out the door.

The data in this article comes from Marcas de Puebla's operational database and is published with their permission.