The chatbot projects that fail rarely fail on technology. They fail on scoping: fuzzy goals, missing data, no owner, success never defined. All of that is knowable before a line of code is written.
These are the ten questions we walk through in discovery calls. Answer them honestly and you'll either have a well-scoped project — or the valuable discovery that you shouldn't build one yet. Both outcomes are wins.
Questions 1–3: the job to be done
1. What specific job is the bot doing? "Answer support questions about billing and how-tos" is scopeable; "improve customer experience" is not. If you can't name the job, you can't measure whether it's done.
2. Who exactly is talking to it — anonymous visitors, logged-in customers, or your own staff? Each implies different knowledge, tone, and security.
3. What happens today without the bot? The current process (and its cost) is your baseline. Chatbots that replace an expensive, slow, or absent process show value fast; chatbots bolted onto a process that already works show nothing.
Questions 4–6: the knowledge and data
4. Where does the knowledge live, and is it current? A bot grounded in stale docs automates the distribution of wrong answers. Auditing the top fifty questions against your existing content takes an afternoon and predicts half your project outcome.
5. What systems must it read from or write to — CRM, helpdesk, billing, internal APIs? Integration count is the biggest cost driver, so name them up front.
6. What data can the bot absolutely not touch or reveal? Defining the forbidden zone early shapes architecture; defining it late causes rebuilds.
Questions 7–8: the failure plan
7. What should the bot do when it doesn't know? The right answer is some version of "say so and escalate to a human with context". If a vendor doesn't raise this question themselves, that's your red flag.
8. Who reviews the failed conversations? Unanswered questions are the improvement backlog. A named person spending an hour or two a month on them is the difference between a bot that climbs and one that plateaus.
Questions 9–10: ownership and success
9. Who owns the bot after launch — the artifact (code, prompts, data) and the operation (monitoring, updates)? "The vendor, forever" is an acceptable answer only if chosen deliberately with eyes open.
10. What number defines success, measured when? "50% honest deflection at stable CSAT within 90 days" is a goal. "See how it goes" is how projects drift into quiet abandonment.
Scoring yourself honestly
Crisp answers to eight or more: you have a well-scoped project, and any competent builder can give you a realistic fixed-scope quote from them.
Fewer than five: the cheapest next step isn't a build — it's a scoping exercise (ours is the AI Readiness Audit) that turns the unknowns into answers before real money moves. Discovering "not yet" for the price of an audit beats discovering it for the price of a build.
Frequently asked questions
Conclusion
Ten questions, one afternoon of honest answers, and you'll know whether you're looking at a high-value build or an expensive experiment — before spending a pound or dollar either way.
If you want to work through them with people who build these systems, that's exactly what our free discovery call is.