Replacing Support with ChatGPT Cut Sales from 20% to 3%: What AI Can't Handle

Every support manager who launches an AI assistant sees the same pattern. At first the team is afraid: "I'll be replaced." Then it calms down: "It's dumb." And when the bot starts genuinely helping, the anxiety returns. Let's look at where AI is truly stronger than people, where it fails, and what to do with the team.

What AI Has Already Taken Over

A significant part of first-line work isn't communicating with the customer but routing and finding information. Here the machine is objectively faster. At one large IT company, an assistant merged scattered knowledge sources into a single search. Engineers used to spend hours digging through documentation and ticket archives; now they get the relevant sections and descriptions of similar problems within minutes. Response time to requests from the bank client fell by 29%.

An important detail: the company also rethought employee motivation. People were encouraged not only to use the tool but also to add to the base of standard solutions. Without that, search quickly runs into outdated data.

The same applies to classifying requests by topic and priority, drafting replies to standard questions (with human review), and extracting facts from ticket history. An agent can prompt the operator right in the conversation: "A similar case came up in March, here is the solution." If a specialist spends 60% of their time on such tasks, AI doesn't take away their job but frees up time.

Where AI Fails

A telling experiment: a product team replaced its first line with ChatGPT. About 20% of sales went through support, with managers turning conversations into orders. After automation the share fell to 3%. The bot answered questions but didn't pick up on doubts, offer alternatives or overcome objections. It informed rather than sold.

The second problem is familiar to anyone who has connected an LLM to support: the model confidently invents facts and almost never says "I don't know." A practical fix from discussions: explicitly write in the prompt "if you can't find the answer in the context, say you don't know," and add a mandatory check of the answer before sending. But a human still devises and monitors those rules.

There are other weak spots:

• Emotions. The phrase "I understand your frustration" doesn't change the tone of a conversation when the customer is at the breaking point.
• Judgment under uncertainty. A bug can't be reproduced, two departments say different things, and you must decide whether to escalate or wait. AI will suggest options; a human decides.
• Systemic problems. AI searches within what it's given. A human notices that this is the fifth such request this week and launches a root-cause analysis.
• Trust. Customers forgive a human's mistake, but a bot that made an error and didn't admit it loses trust for a long time.

What Managers Should Do

The main risk isn't AI itself but management that decided "AI will do everything" and cut staff. That is exactly how sales through support fell sixfold.

Practical steps:

• Change your KPIs. If people are evaluated by the number of closed tickets, they will resist a bot that closes them faster. Add the metrics "knowledge base quality" and "share of cases solved with the AI assistant," and reduce the weight of response speed for complex categories.
• Set aside about 10% of time for improving the knowledge base and reviewing AI errors.
• Explain to senior leadership: AI speeds up search but doesn't replace judgment.

The author describes three growth paths for those whose tasks AI is taking over. A knowledge engineer writes instructions, checks model answers and updates the base (3–6 months of training). An L2+ complex-case specialist investigates causes, reads logs and handles negotiations (6–12 months). A support quality analyst looks at what AI closes and what it fails at, and serves as a bridge to the product team (up to a year).

Limitations and Takeaway

The article relies on individual cases without disclosing the companies or methodology, and the figures (29%, 20% → 3%) should be taken as illustrations, not benchmarks. Also, the text ends with an ad for a course and webinar, so the emphasis on "new professions" is partly promotional.

But the practical conclusion is sound: don't measure support by response speed per ticket, because you can't win that race against a machine. Give AI the search, sorting and drafts, and leave humans with sales, escalations, root-cause analysis and responsibility for what the bot knows and says. The winner is whoever builds a system where answering is needed less often, because the problem was fixed in the product.