Use case assessment
Is an AI customer support chatbot worth building?
Genuinely AI. The question is not whether to build it but how much of the problem to give it.
Real AI with real value. The failure mode is almost never the model; it is a scope that covers everything and is trusted on nothing.
What it usually means
A conversational system answering customer questions, typically grounded in a help centre and sometimes able to act on an account.
Is it genuinely AI?
Yes. Understanding a question asked in the customer's own words, matching it to knowledge written in someone else's, and answering in context is not something rules do well. Decision-tree bots have been available for fifteen years and people route around them, which is the evidence.
The decision that actually matters
Not which model. How much of the problem it is given.
The instinct is to cover everything, because a bot that says it cannot help feels like a failure. It is not. An assistant that handles the top questions well and hands off cleanly keeps its credibility. One that attempts everything and is unreliable at the edges loses it permanently — and the second bad answer is the one customers remember.
On a change-management agent I scoped, cutting from the full process to the common path only was the unpopular decision and the one the numbers came from: request time from fifteen minutes to seven, policy compliance from 72% to 89%.
What the simpler version looks like
Better search over the help centre. Genuinely: a large share of support volume is people failing to find an answer that exists, and improving retrieval is cheaper, faster and carries none of the disclosure obligations.
Worth measuring before building: what proportion of tickets are answered by content that already exists? If it is high, you may have a findability problem wearing a chatbot's costume.
What it costs to run
Inference scales with volume and is the visible part. The rest: keeping the knowledge base current, because a confidently wrong answer from stale content is worse than no answer; evaluation against a labelled set of real questions; monitoring for drift; and a human path that is genuinely staffed.
The escalation path is the part that gets underfunded and the part customers judge you on.
Where it sits under the EU AI Act
Article 50 applies directly and has been enforceable since 2 August 2026. A customer talking to your bot has to know it is a bot, unless that is obvious from context.
Not normally Annex III high-risk. It can become so if it determines access to essential services. If it acts on accounts rather than only answering, the risk profile changes: the cost of being wrong stops being a bad answer and becomes a bad outcome.
When it is worth building
When volume is high, when a real share of it is repetitive, when your knowledge base is good enough to ground answers, and when you will resource the handoff.
Not when the underlying problem is that your documentation is poor. The bot will surface that faster and more publicly.
Common questions
Do we have to tell customers they are talking to a chatbot?
Yes. Article 50 of the EU AI Act requires disclosure when a person interacts with an AI system, unless it is obvious from context. That obligation became enforceable on 2 August 2026.
What makes support chatbots fail?
Scope, almost never the model. An assistant that covers the common questions well and hands off cleanly keeps its credibility; one that attempts everything and is unreliable at the edges loses it after the second bad answer.
What should we try before building a chatbot?
Better search over the existing help centre. A large share of support volume is people failing to find answers that already exist, and improving retrieval is cheaper, faster and carries no disclosure obligation.
Is a support chatbot high-risk under the EU AI Act?
Not normally. It can become an Annex III high-risk use if it determines access to essential public or private services. If it takes actions on accounts rather than only answering, the risk profile changes materially.
Cutting enterprise change requests 15 to 7 minutes — the scope cut described here, and what it moved.