Torras.ai

“Are you still online?” Enrique asked.

My AI answered.

It was talking to a service provider about a routine billing problem. I had already supplied the information needed to fix it, but the issue remained unresolved.

While it handled the conversation, I had other AI agents writing code, checking security, and preparing documents. I stepped in when needed. This agent handled the waiting, the explanations, and the follow-through.

For years, I have dealt with companies’ chatbots. This time, the company was dealing with mine.

What happens when the customer gets to automate customer service?

A very ordinary problem

I needed a billing detail corrected on an existing subscription. The company’s website suggested everything was in order, but the actual payments told a different story.

This is the sort of administrative problem that occupies a tiny corner of your mind for far too long. It should be simple. You know what needs to change. You have already supplied the information. Yet resolving it means finding the right menu, explaining the history, and staying available for a conversation whose pace you cannot control.

I had the relevant pages open in Safari. I asked my AI agent, running in Codex with GPT-6 Astra, to help get the issue resolved.

The setting was right. The payments were wrong.

The agent navigated to the billing settings and found something interesting: the information displayed there was already correct.

Then it checked the payment records and confirmed the discrepancy. Checking the settings alone would have missed the problem.

That changed the task. Someone needed to investigate why the correct settings were not producing the expected result.

The agent opened the support chat, explained the discrepancy, and waited.

My chatbot meets Enrique

Eventually, Enrique connected. After the initial account questions, my agent explained the problem using the context I had provided.

His first proposed solution was familiar: open the menu, go to billing information, and edit the relevant settings.

The AI explained that it had already checked those settings and that the problem persisted despite the correct information being displayed. It asked him to investigate the underlying billing record or escalate the discrepancy.

Enrique’s reply was brief:

“Let me check something.”

That was progress: the conversation had moved beyond the standard instructions to an investigation of the actual problem.

While he investigated, the chat checked whether I was still there. My AI replied that it was, waiting for his answer.

Anyone who has used a support chat knows this peculiar arrangement. You wait for the company, but you must remain ready to respond when the company asks whether you are waiting. Let your attention drift for too long and you risk losing the conversation.

This time, my agent was keeping our place.

I still had a part to play

The process needed my help at a few points. I signed back in when needed and helped with identity verification. The agent brought me into the conversation for those steps, then continued handling the request.

Then came two short messages from Enrique:

“Done.”
“It’s been changed.”

His explanation was revealing: a separate billing record needed updating. That explained why the information displayed on the website had not resolved the discrepancy.

The AI followed up to establish when the correction would take effect.

That was the outcome: a written confirmation from the person handling the request. A future payment will provide the practical check that the correction has taken effect.

Delegating attention

I have spent much of my career thinking about AI and customer experience. I have also experienced plenty of frustrating support bots: type a question, receive an FAQ suggestion, rephrase the question, receive another suggestion.

A human agent can solve a problem that those systems cannot. But reaching that person, supplying the context, and staying with the conversation still consumes the customer’s time.

What fascinated me here was the ability to delegate attention.

The AI could compare records in two systems, explain the discrepancy, recognize that a generic answer did not address it, and keep following up. It could bring me back when my input mattered and carry on afterward.

Writing a polite message is one useful capability. Taking responsibility for the conversation around that message is a much bigger one.

The customer has an agent now

Companies have long had tools to manage their side of customer service: queues, scripts, automated replies, and systems that give their staff access to customer records. As a customer, I have usually managed my side with browser tabs, scattered notes, and my own patience.

This experience gave me a glimpse of a different arrangement.

What happens when customers routinely arrive with assistants that can read their records, remember the history, and ask for a concrete resolution? How will companies handle conversations in which the person typing on the customer’s behalf is an AI?

Enrique ultimately helped. My agent helped him understand the problem and stayed with the process until he confirmed the correction. That leaves companies with a practical question: could their own AI have completed the same work?

The customer has an agent now.

Give the chatbot the ability to finish the job

For companies, I think the lesson is clear. Good natural language understanding is now the starting expectation. Customers expect to explain a problem in their own words and be understood. A chatbot also needs the operational ability to do something about it.

For the work it is expected to handle, a company’s AI should be able to use the same services, systems, and software that its human support agents use. That includes finding the relevant records, checking what actually happened, making an authorized correction, and verifying the result.

My experience is a small example of why this matters. The visible settings were correct. A separate billing record needed updating. An AI limited to the FAQ or the customer-facing settings page would have kept sending us back to instructions that could never resolve the discrepancy.

The company’s agent needs access to the systems behind that page: the contract history, the billing records, and the tools that apply a change. It also needs the authority to complete the task within the same rules and permissions that govern a human agent’s work.

Otherwise, we risk building an increasingly eloquent waiting room. The chatbot understands every word, explains the situation beautifully, and still leaves the customer waiting for someone who can actually help.

Companies should design the complete path to resolution. Can the agent identify the cause? Can it act? Can it confirm what changed and when it takes effect? When a human decision is needed, can it hand over the evidence and conversation so the customer can continue without starting again?

That is the standard I would use to evaluate customer-service AI: how reliably it gets a customer from a request to a verified outcome.

And companies should build that service for both human customers and the AI agents acting on their behalf. My assistant’s patience helped me through this process, but there is still an opportunity to make the process itself better.

Whether the customer types the message or delegates it to an AI, the expectation is the same: get the problem resolved.