Website chat is one of the clearest conversion tools available: a visitor with a question who gets an answer is far more likely to buy than one who leaves to find it. The question is who answers.
What each is good at
Chatbot strengths
Instant response, always. No queue, no waiting, no business hours. A visitor at 11pm gets the same answer as one at 11am.
Unlimited concurrency. Fifty simultaneous conversations cost the same as one.
Consistency. The same question gets the same correct answer every time, which is genuinely hard for a human team.
Near-zero marginal cost. Each additional conversation costs very little.
Captures intent at the moment of hesitation, which is where sales are won or lost.
Chatbot weaknesses
No judgement. It cannot decide whether to make an exception, offer a discount, or recognise that this customer is unusually important.
No empathy. A frustrated or upset customer needs a person, and a bot attempting warmth frequently makes it worse.
Fails on the unexpected. Anything outside its scope should escalate rather than improvise.
Can invent. Mitigated substantially by grounding in real documentation, not eliminated.
Live chat strengths
Handles anything. Including situations nobody anticipated.
Reads tone. A person can tell that a customer is annoyed and respond to that rather than to the literal question.
Commercial judgement. Discounts, exceptions, reading buying signals and pushing appropriately.
Builds relationship, particularly in high-value sales.
Live chat weaknesses
Limited hours. Most visitors arrive outside them.
Limited capacity. A busy period means queues, and queues mean abandonment.
Inconsistent. Different agents give different answers; quality varies with workload and experience.
Expensive at scale, because it scales linearly with volume.
The answer is usually both
Framing this as a choice is the mistake. The real question is where the handoff sits, and that depends on your conversation mix.
A sensible default:
- Bot first, answering instantly from real documentation and live data
- Human escalation available at any point, obvious and one click away
- Automatic escalation on defined triggers: complaint language, high-value accounts, repeated failure, explicit request
- Bot-only outside staffed hours, with a clear promise about when a person will respond
Make the human route obvious
Counter-intuitively, making escalation easy increases satisfaction with the bot.
When visitors know they can reach a person, they are willing to try the automated route. When they suspect they are trapped, they fight it from the first message — typing "agent" repeatedly instead of describing their problem, which produces a bad experience for everyone and no useful data for you.
Hiding the handoff to improve deflection metrics is the single most counterproductive decision available in this area.
Where the handoff should sit, by business type
High-volume, low-value transactions — ecommerce, subscriptions. Bot handles most, humans take exceptions. The economics do not support staffing for volume.
Low-volume, high-value sales — professional services, B2B, property. Bot qualifies and gathers context, a person has the conversation. The deal is worth a human's time.
Technical support — bot handles documented issues with live system data, humans take anything requiring diagnosis.
Healthcare and regulated services — bot handles logistics and general information only, with anything substantive escalated by design.
What makes a bot worth having
Not all chatbots are equal, and a bad one is worse than none.
Grounded in your real content. Not keyword matching against a fixed list of answers. It should read your actual documentation and answer questions nobody scripted.
Connected to live data. A bot answering "where is my order" by explaining how to check your order status is useless. One that looks up the actual order has resolved the ticket.
Able to say it does not know. Confidently wrong is far worse than honestly uncertain.
Capturing leads naturally. Mid-conversation, when the visitor is engaged, not as an upfront gate.
Carrying context on escalation. The customer should never repeat themselves. Having to re-explain after a failed bot conversation is the specific experience that makes people hate chat automation.
Measure resolution, not deflection
Deflection rate — conversations that did not reach a human — is the metric vendors report and it is misleading, because a customer who gave up in frustration counts as successfully deflected.
Better:
Resolution rate. Did the customer get what they needed? Harder to measure and worth the effort.
Escalation rate by category. Which question types the bot handles, which it does not. This tells you where to invest.
Repeat contact rate. Someone returning within 24 hours about the same issue was not resolved.
Conversion rate for chat users versus non-users, which is the number that justifies the whole thing commercially.
Satisfaction split by path — bot-resolved versus human-resolved. If the gap is large, deflection is costing goodwill.
A reasonable starting point
- Read your last two hundred chat or email conversations and categorise what people actually wanted. This is your requirements document.
- Build the bot for the top three categories only, connected to live data.
- Route everything else to a human untouched, with context attached.
- Make escalation one click, visible at all times.
- Read transcripts weekly and expand categories based on what the bot could not answer.
Resist launching a bot that attempts everything. One handling three categories completely, with clean handoff for the rest, is received far better than one half-handling ten.
If you want your conversation data analysed to see where the handoff should sit, send us a sample.
