The hardest part of building a good chatbot isn’t the model — it’s the scope. A tight scope makes evaluations meaningful, guardrails testable, and launch dates real. A loose scope produces an assistant that either hallucinates or refuses everything.
Why scoping matters
Every bad chatbot we’ve seen had the same origin story: everyone agreed it should "help customers" but no one agreed on which customers, which questions, or which answers were off-limits. So the bot tried to answer everything, and did most of it badly.
A one-week scope isn’t a compressed research project — it’s a forcing function. It gets the right people in the room, extracts the answerable questions, and produces a written contract before code is written.
Day 1: intents on a whiteboard
We start with the questions people actually ask. Not what marketing wishes they asked — what shows up in the support inbox last month. From those, we cluster into 8–15 intents that will pay for the build.
- Pull the last 30 days of support tickets.
- Cluster into intents; discard anything under 1%.
- Mark refusal-list topics next to the intent list.
Day 3: the knowledge base
By mid-week, we know which documents ground each intent. Almost always we discover the docs need work — product pages contradict the help center, or key answers live only in Slack. Fixing that is scope, not a delay.
If the answer isn’t written down anywhere, the bot won’t answer it. Retrieval only works over what exists.
Day 5: guardrails
Now the hard question: what should the bot refuse? Legal, medical, financial advice, opinions about competitors, promises about pricing — the specific refusals depend on the business. We write them into the prompt and the eval set on the same day.
- Draft explicit refusals.
- Draft PII handling rules.
- Draft the escalation flow to a human.
Day 7: the written scope
We land with a document that fits on a page: the intents, the knowledge sources, the refusals, the escalation path, and the metrics. Everyone signs. Building starts Monday with almost no ambiguity left.
Key takeaways
- Scope is a forcing function — not a research project.
- Intents come from real ticket data, not wishful thinking.
- Refusals and escalation matter as much as answers.
- A written, one-page scope beats a Slack thread.
If this sounds like the process your project needs, we’d be happy to run it. Get a quote — a senior lead will reply within a business day.