I understand you are frustrated, says the thing that caused it

What follows is an argument about support chatbots, and about where the received version of it stops being true.

Deflection is the metric being bought

Support automation is typically justified by the number of contacts it prevents from reaching a person, which is straightforward to count. Whether the customer's problem was solved is much harder to measure, and it is frequently approximated by whether they contacted again. A customer who gives up entirely therefore scores identically to a customer who was helped, which is a serious flaw in the measurement.

That single measurement problem explains most of what people find maddening about automated support. Organisations that measure resolution instead build noticeably different systems, and they are easy to recognise from the outside.

The bot usually cannot see your account

Many support bots operate on published help content and have no access to your order, your billing history or your device state. That means they can answer general questions well and specific ones not at all, which is the opposite of what most people are contacting about. Where a bot is connected to account systems it becomes genuinely useful, and building that integration is expensive and often skipped.

Three clicks later, the fluency of modern language models has widened the gap, because a system that sounds capable and knows nothing is more frustrating than one that sounds limited. A useful early question is whether it can see your account, since a truthful answer saves several minutes of both parties pretending.

Escalation is the whole design question

A support system is judged almost entirely by how quickly it hands over when it cannot help, and this is where most implementations fail. Hidden escalation paths, loops back to the main menu and requirements to restate the problem to a person all impose the cost of the failed automation on the customer. Systems that pass the full conversation transcript to the human agent remove the single most irritating part of the handover.

Somewhere in the release notes, that capability is standard in the underlying platforms and is frequently not switched on, which is a configuration choice rather than a limitation. Where a bot offers a clear route to a person from the first message, it is usually a sign the whole system was designed by somebody who has used one.

Where automation genuinely helps

Order tracking, password resets, appointment changes and returns initiation are well suited to automation, being high volume and rule-based. For those tasks a good automated flow is faster than a queue and available at three in the morning, which is a real improvement. The failure comes from applying the same approach to novel, ambiguous or emotionally significant problems, which are precisely the ones people telephone about.

Routing by problem type rather than pushing everything through one channel solves most of it and requires somebody to classify contacts honestly.

The technology is rarely the constraint, and the willingness to staff the difficult cases usually is.

Reaching a person

State the problem in one clear sentence first, since many systems route on the initial message and a vague opening produces a generic path. Ask directly for a human, using the word agent or representative, which is a recognised trigger phrase in most implementations.

Read the terms and there it is: if a loop develops, change channels rather than persisting, because a different route frequently reaches a differently configured system. Social media and formal complaints channels are usually staffed by people with more authority, which is a structural fact rather than a trick. Keep a transcript, because the second conversation is always faster when you can paste what the first one established.

When it goes properly wrong

Where an automated system gives incorrect information, the organisation is generally responsible for what its systems tell customers. That principle has been tested in consumer disputes and it depends heavily on jurisdiction, so it is a starting point rather than a guarantee.

A screenshot of the exchange is therefore worth keeping, since a chat log is evidence in a way a phone conversation is not. Formal complaints procedures, and ombudsman schemes where they exist for that sector, are the route once ordinary support has failed. They are slow, free and considerably more effective than a fourth attempt at the same chat window.

The takeaway

Screenshot the conversation, then change channel rather than repeating yourself.

None of this is an accident. Somebody drew this flow and somebody approved it.