The bot that became a wall

import json import os from hurozo import Agent os.environ["HUROZO_API_TOKEN"] = "YOUR_API_TOKEN" agent = Agent("sentiment-analyser") inputs = { "prompt": "Analyze sentiment and return JSON. Text:", "input": "I am very happy!", } result = agent.run(inputs) print(json.dumps(result, indent=2))

Last week a member of TrainMore, a large European gym chain, tried to get a VAT invoice. What they got was a chatbot with a friendly human name, an error about a “temporary limit,” a one-time code the bot then refused to accept, and the same question put back to them four times. Every route to a person led back to the bot. It ended up on the city’s subreddit, upvoted into the hundreds, commenters working out which no-code tool and which model the company had wired together.

I read that thread twice. The first time I thought what everyone thought: AI support is bad. The second time I saw the worse thing underneath it.

I spent years running a chemical-materials marketplace that did about 1.5 billion euro across three continents, and I still could not always get a customer to the right person fast enough. So I know the pull of the cheap fix. When the queue grows faster than the team can, a bot that answers everything over a weekend looks like mercy.

But almost none of the anger in that thread was about a wrong answer. It was about the bot standing between a paying customer and the human who could have fixed the whole thing in ten seconds. TrainMore shipped more than a weak tool. It put a machine in the doorway and took the door off its hinges.

And here is the part that should worry anyone who cares about their brand. It was self-inflicted, it was public, and it does not scroll away. A bad review fades in a week. This is a case study in how not to treat customers, written by the customers themselves, sitting on the front page of a forum its future members read before they join.

The false choice

You get handed a choice, and both answers are bad.

Great human service does not scale. It is expensive, the good people are hard to keep, and the knowledge that makes it good lives in one person’s head. You feel it every time the queue outruns the team.

Cheap automated service scales, and it costs you the brand. A no-code flow plus a model is a weekend of work. It runs fine, until the Tuesday it doesn’t, and then it loops, dead-ends, makes up a policy, and cannot tell when it is out of its depth. A reputation bomb with a delay timer.

Most companies pick one and lose something they cared about. They cap their growth on the price of good service, or they trade the reputation they spent years building for a bot that will, sooner or later, end up in a screenshot. And it lands hardest on the operators who should care most, the ones who chose service as the reason a customer picks them over someone cheaper. For them a viral failure does more than dent a quarter. It burns the exact thing the business is built on.

The third option

The real problem sits one level up from cost. It is this: deliver the service that builds the brand, at a size you can afford, without it ever going off-script.

That points at a different kind of tool. An agent that answers from your own canon: your real policies, your prices, the exceptions your best people already carry in their heads. It acts behind a gate you hold, so nothing reaches a customer that you would not put your own name on. It knows when it is out of its depth and hands over to a person on the same thread, at once, with no wall in between. And because it runs on infrastructure you own, with its actions recorded, you can show afterwards exactly what it did, to a customer, to your team, or to a regulator.

Here is where that gets too clean, so let me break it myself. A gate does not make the model correct. It will still draft things you would never send. What the gate changes is who meets the mistake first: your team, early, with a record, while the customer only ever sees what you would stand behind. The promise is small and exact. Your customer is never the one who finds the error, and you can prove what happened when someone asks. If you were hoping for a model that never errs, no honest person can sell you one.

Two reasons, one direction

There are two reasons to build it this way, and they pull in the same direction. The first is growth. You say yes to more customers without watering down the service you are proud of, and your best people stop answering the same question for the hundredth time. The second is protection. You do not wake up as the punchline on a forum, because the agent cannot wall a customer off and cannot fold a mistake into silence. Most tools sell you the first and quietly hand you the opposite of the second.

The choice

The companies that win the next decade on service will be the ones who scaled the service they were already proud of. TrainMore ran the other race. It had that choice on an ordinary Tuesday, and answered it with a chatbot in the doorway.

The thread is still up.

This is the problem we are working on.

Common questions

What actually went wrong with TrainMore's support chatbot?

The failure was not a single wrong answer. A member trying to get a VAT invoice was looped through error messages and a one-time code the bot would not accept, and every attempt to reach a person led back to the bot. The chatbot became a wall between a paying customer and the human who could have solved the problem in seconds, and the exchange went viral on the city's subreddit.

Does this mean AI customer support is a bad idea?

No. It means an ungoverned bot with no path to a human is a bad idea. The problem is not automation, it is automation that can trap a customer, invent policy, and fail in silence. A governed agent that answers from the company's own rules and hands to a person when it is out of its depth does not carry that risk.

What is the alternative to a rented support chatbot?

An agent that answers from your own canon, your real policies, prices, and product truth, acts only behind a gate you hold, and hands the customer to a person on the same thread when it is out of its depth. Run it with its actions recorded, so you can prove what it did to a customer, your team, and a regulator. Owned, not rented.

Set up your first agency

Manage worker

Connect services for the current agency

New message

Hurozo CLI

Model routing

Approvals

Second opinions

Connected tools

First run

Hurozo SDK

Agent lookup

Structured inputs

Hosted runtime

Token setup