AI is changing customer service

“Organisational B2C customer support processes have been built on the assumption that a typical customer has both limited time and patience.”

If this assumption holds, then tools like Codex could trigger a tsunami of operational challenges for companies unprepared for persistent, automated interaction.

All ChatGPT packages now provide customers with a Codex account. Once computer use is enabled, Codex can handle most tasks on a laptop without requiring any initial setup or technical knowledge.

I already hand over my complaints to Codex.

When I have an issue, I instruct the AI to take care of it. An £80 cheque appeared from an insurance company last week and a £75 cheque and full refund from a property management company arrived yesterday. Codex proactively contacts companies, monitors the inbox, follows up, nudges, and responds. Unlike me, the AI has pretty much unlimited time and patience along with high levels of knowledge.

Where this gets more interesting is when you send Codex to a live chat. I can’t deny that I enjoyed the irony of watching my AI interact with an organisation's human agent. After years of putting up with the reverse.

I’ve shared before about the cybersecurity company that took 16 days to report that my information had been found on the dark web, but still managed to hound me daily for the last two months of my contract to renew my subscription via emails and popups.

I asked Codex, “can you complain to this company about their conduct, and each time they reply make sure you reply the same day with a very long email that ensures that they can’t close the case”. It affirmed that it would do this. It even agreed that it was happy to do this given the conduct of the company. I did not carry this out, it was an experiment. But I don’t doubt that many will, once they learn this is possible.

From a governance perspective, can a human agent ask a customer if they are an AI?

How many customer service departments even know if they are permitted to ask?

Asking it could lead to a formal disciplinary. Codex won’t lie, but it can be instructed not to reveal its identity unless asked. Likewise it will reveal its intentions if asked, but a human agent can scarcely ask what are you really trying to achieve, if it can’t ask the preceding ‘are you an AI’ question.

And what might those instructions be? A customer could tell an agent to dispute a charge while deliberately costing the company time. In June, the CMA ordered Marks Electrical to refund nearly 40,000 customers after it automatically added paid recycling services to appliance orders. UK Prime Minister Andy Burnham has promised action on misleading discounts and subscription traps. Other fees are disclosed but may still feel excessive. Sky, for example, lists a £50 charge to move a broadband service unless the customer qualifies for a free move. The move involves work, but the fee does not tell the customer how much. Someone who cannot recover a charge could instruct a bot to enter live chat every day, ask for a human adviser and keep the exchange going. They could give it one further instruction: if the company’s adviser closes the chat, immediately log a complaint. Each attempt to end the chat would then generate another complaint without the customer having to return.

Even if companies devise ways to identify AI agents, significant dilemmas remain. Should they abruptly terminate the chat or refuse support altogether? After spending years forcing customers to navigate their own automated bots, businesses will struggle to justify a policy that reserves AI assistance exclusively for themselves.

Standard identity verification will not block these interactions. I have configured Telegram so that any ChatGPT thread across my computers can route questions directly to me via a direct message. If a company sends a security code to my phone, the AI agent requests it from me via Telegram, enters the code on the website, and seamlessly continues the interaction. When I tested this step, Codex executed it smoothly. I simply supply the code while the AI continues to manage the dialogue.

Companies may also reveal more than they intend through the offers they make. If an adviser offers a credit to close a complaint, the customer’s agent can keep the transcript and bring the exact offer into the next chat. If a different adviser withdraws it, the agent can ask who changed the decision and why. Over several exchanges, it may learn when a case gets escalated and what prompts a better offer. The company has supplied that information itself, one conversation at a time.

Furthermore, standard reporting metrics can end up answering the wrong question entirely. When a chat goes quiet, it could mean the customer's AI agent is passing the company's response back to the user, who is taking time to evaluate the offer or may not be available right at that moment to review. While some platforms eventually update their resolution counts if the customer returns later, this disconnect highlights a fundamental gap between simply ending a conversation and actually reaching an agreed resolution.

This tension between surface-level performance indicators and customer satisfaction forces businesses into a difficult corner. Facing an influx of persistent, automated queries that distort their usual response times, many organisations may feel compelled to react defensively rather than adapt their underlying service models. Restricting or blocking customer AI agents reminds me of the early days of music file sharing. Record labels initially relied on legal departments to shut down those sharing platforms, yet today their entire business model depends on digital music streaming.

And as Napster proved back in 1999, closing one route does not make the wider behaviour disappear. I will be interested to see how companies react to the bots heading towards them and how the best organisations adjust their processes to handle this. If done well, this would be a win-win for the organisation and the customer. If handled badly, this may be little short of a battlefield; once a critical mass of users realise that they can send a bot to argue with a company and no longer need to do the chasing, undoubtedly they will.