There are two versions of the enterprise chatbot. The demo version: it answers everything perfectly, the client is satisfied, the support team is freed up. The production version six months in: 40% of conversations end in "I don’t understand your question", the team is still taking the same calls and nobody knows whether the system is generating value or not.
This guide is about the real version.
What the real data says
The figures circulating about chatbots almost always come from chatbot providers. It is worth reading them with scepticism. What does exist is independent research with more honest ranges.
Autonomous resolution rate: between 40% and 80% depending on the sector and the complexity of queries. 40% corresponds to basic implementations with little personalisation. 80% to advanced implementations with a complete knowledge base and real-time access to systems. The actual average in SMEs is around 55-60%.
User satisfaction: comparable to human support for informational queries. Significantly lower for complex or emotional queries. The net promoter score of a well-implemented chatbot for FAQ queries is typically between 30 and 50, comparable to that of a human agent for the same type of query.
Reduced team workload: between 30% and 60% of support tickets. The figure depends directly on the proportion of repetitive queries in the total.
The sectors where they work best
Hospitality and catering: bookings, opening hours, availability, menu information. 70-80% of queries are always the same. The chatbot handles them perfectly and the ROI is immediate.
E-commerce: order status, returns policy, product availability, sizes. With access to the management system API, the chatbot can respond with real-time data. The autonomous resolution rate exceeds 75%.
B2B professional services: initial lead qualification, information about services, quote requests. The chatbot collects prospect data and qualifies them before passing them to the sales team. It reduces the time salespeople spend on discovery calls by 40%.
Clinics and health: appointments, schedules, information about specialists, visit preparation. High sensitivity in queries — the design of the handoff to a human is critical.
The sectors where they perform worst
Highly personalised services: consultancies, advisory firms, creative services. Queries have no standard answer. The chatbot frustrates the user because it cannot give the specific answer they are looking for.
After-sales support for complex products: industrial machinery, specialist software, technical installations. Queries require deep technical knowledge and specific customer context.
Complaints and claims management: although the chatbot can collect information, resolving a complaint requires human judgement, empathy and negotiating ability. A chatbot that attempts to resolve a serious complaint without immediate escalation can make the situation worse.
The metrics worth measuring
Containment rate: the percentage of conversations resolved without escalating to a human. It is the primary efficiency metric.
Correct escalation rate: when the chatbot escalates, does it do so at the right moment with the full context? Incorrect escalation — too soon or too late — has the same negative impact on the experience as not escalating at all.
Post-conversation CSAT: user satisfaction after interacting with the chatbot. It should be measured separately for conversations resolved automatically and for conversations that have been escalated.
Time to resolution: compare the average resolution time with chatbot vs without chatbot for the same type of query.
Cost per ticket: the calculation that justifies the investment. Cost of the human team per ticket without chatbot vs operating cost of the chatbot per conversation with the percentage of autonomous resolution.
The most costly implementation mistake
The most costly mistake is not technical. It is launching the chatbot without a clear process of continuous improvement. A chatbot that is configured once and left to run degrades in quality over time because company information changes, prices change, products change, and the knowledge base becomes out of date.
Successful implementations have a weekly review process: someone from the team reviews the week’s conversations, identifies incorrect responses and updates the knowledge base. Without that process, the chatbot becomes a generator of frustration rather than a problem solver.