The question is placement, not belief
Arguments about AI in support tend to collapse into whether you are for it or against it, which is not a useful frame for someone configuring a widget on Tuesday. AI is a tool with a shape: excellent at retrieving and rephrasing information that exists in your documentation, weak wherever the right answer depends on judgment, authority, or emotional read. Deciding where it sits is an engineering decision, and getting it wrong in the cautious direction costs you a little efficiency, while getting it wrong in the confident direction costs you customers.
Emotional conversations
When someone is angry, frightened, or grieving, the content of the reply matters less than the sense of being heard by a person who can be affected by what they said. AI can produce sympathetic sentences; what it cannot do is bear responsibility, and that is what an upset customer is actually looking for. An automated apology to someone who has just described a serious problem tends to read as dismissal, and it converts a recoverable complaint into a public one. Route detectable frustration to a human immediately and accept the occasional false positive.
Anything irreversible or expensive
Refunds, cancellations, account deletions, contractual commitments, and pricing exceptions should involve a person. Not because AI cannot process the request, but because the cost of a confident error is asymmetric: a wrong answer about your opening hours is a minor annoyance, while a wrongly promised refund is a real financial commitment a customer will reasonably hold you to. The rule of thumb is simple — if getting it wrong costs money or cannot be undone, a human decides.
Questions your content does not answer
An AI grounded in your help centre is only as good as the help centre. Ask it something the documentation does not cover and the failure mode is not silence but plausible invention, which is far more damaging than an honest gap because it is indistinguishable from an answer. Before widening what AI handles, look at what your content actually covers, and prefer a system that says it does not know and offers a human over one that always produces something.
High-stakes and regulated topics
Medical, legal, financial and safety questions deserve a human, and often a specifically qualified one. The reputational and legal exposure of an automated answer in these areas is out of all proportion to the handling time saved. The same caution applies to anything touching identity or account security: a public chat widget is not an authentication channel, and it is not the place to automate decisions about who someone is.
Where AI genuinely earns its place
None of this argues for switching it off. Repetitive factual questions — hours, policies, order status, how a feature works — are answered faster and more consistently by AI than by a tired human at 2am, and every one it handles is a conversation your team does not have to. The strongest setup is layered: AI takes the volume that is genuinely repetitive, and hands over the moment the conversation stops being routine.
How MyLiveChat fits
MyLiveChat runs AI and human agents on the same queue, so a handover happens inside the conversation the visitor is already in rather than restarting it somewhere else. The AI answers from content you supply, which keeps it inside what you have actually documented, and a visitor can ask for a person at any point. That layering is the practical version of everything above: automate the repetitive volume, and make the exit to a human immediate and obvious.
The placement decision, page by page
Because the answer is placement rather than belief, it is worth making the decision explicitly per surface instead of switching AI on globally and hoping. Walk your main templates and write the choice down.
- Documentation and help pages. Strong fit. The questions are factual, the content exists, and a wrong answer is cheap to correct.
- Pricing. Mixed. Plan mechanics answer well; anything touching a negotiation, a custom quote or a renewal should reach a person.
- Checkout and payment. Usually not. The cost of a confident wrong answer is a lost order or a refund, and the visitor is at their least patient.
- Account, billing and cancellation. Not without care. These are irreversible or contested by nature, and they are the pages where identity questions arise.
- Complaint or outage pages. No. Someone arriving angry or worried wants acknowledgement from a person, and an automated reply reads as an evasion.
Writing this down as a short table has a second benefit: when someone later asks why AI is not answering on checkout, there is a reason on file rather than an argument.
Reversing the decision without drama
The scope you set at launch will be wrong somewhere, and the teams that do well are the ones that can pull it back in an afternoon without treating it as a failure of the whole project.
Decide the triggers in advance. A cluster of complaints about a specific topic, a review finding the same wrong answer twice, or a policy change that has not reached your source content are all reasons to narrow scope immediately and widen it again once the content is fixed. Make sure one named person can make that change without a meeting, because a scope problem that waits a week for approval does a week of damage.
Tell the team when you change it, and say why. Agents notice when the mix of chats reaching them shifts, and an unexplained change gets read as the AI breaking rather than as scope being managed.
What to measure before widening scope
Before you extend AI to a new area, look at three things on the areas it already covers. Handover rate should be stable or falling; rising handovers mean the current scope is already past the content. Accuracy on a sample of closed conversations, because handover rate alone tells you nothing about the answers nobody escalated. And repeat contacts after an AI-only conversation, which is the closest thing to an honest resolution signal — a chat that ends and then comes back the next day did not resolve, whatever it looked like at the time.
If all three look healthy, widen by one topic and watch the same numbers for a fortnight. Incremental scope changes with a measurement window between them are slower than a big switch-on and very much cheaper than the alternative, which is discovering the limits through customer complaints.
Ruling AI out of a conversation does not rule it out of the work around one. On tickets where you would never let a model answer unattended, a drafted reply you edit before sending keeps the judgement with the agent.