Skip to content
PRYNODE LABS, home
Notes

Before you build an AI chatbot, answer these five questions

Sadashiv Kumar, . 6 min read

We get a lot of messages that start with "We want an AI chatbot." Sometimes it's exactly the right call. Sometimes a better FAQ page would do the job for a tenth of the cost. Before we quote anything, we walk through these five questions with you. You can do it yourself in about fifteen minutes.

1. What questions will it answer?

Open your inbox, your WhatsApp, or your support tool and pull the last 100 customer messages. Sort them into groups. If most of them are variations of a dozen questions, and the answers are written down somewhere, an agent can probably handle them well.

If every message is different, or the answers live only in someone's head, start by writing the answers down. That alone often fixes half the problem.

2. Where do the answers come from?

An AI agent is only as good as the information you give it. It needs a source: your policies, product catalogue, past tickets, or a live system like an order database. If that source is out of date or contradicts itself, the agent will confidently repeat the mistakes.

The most useful work in an AI project usually happens before any AI is involved: cleaning up the information it will rely on.

3. What should it never do?

Write this list down before anything else. Common items:

  • Promise refunds, discounts or delivery dates
  • Give medical, legal or financial advice
  • Discuss a specific order without verifying the customer
  • Make up an answer when it doesn't know

A well-built agent follows these limits and hands the conversation to a person when a question goes past them. A badly built one tries to be helpful and gets you into trouble.

4. Who takes over, and how fast?

Every agent will hit questions it shouldn't answer. Decide in advance where those go: an email inbox, a WhatsApp number, a support tool. Decide who checks it and how quickly. An agent that hands over to nobody is worse than no agent at all.

5. How will you know it's working?

Agree on one or two numbers before you start. For example:

  • The share of conversations resolved without a person
  • How often a person has to correct the agent
  • Average time to first reply

Then build a test set: 50 to 300 real questions with the answers you'd expect. Run the agent against it before launch and after every change. This is how you catch problems before your customers do.

What it usually costs to run

For most small businesses, the AI model itself costs a few hundred to a few thousand rupees a month, depending on volume. The bigger costs are building it properly and keeping the information it uses up to date. Anyone who tells you otherwise is probably skipping the testing.


If you've answered these and still think an agent makes sense, tell us about it. If you're not sure, send us your 100 messages (with personal details removed) and we'll tell you honestly whether an agent is worth it.

Tell us what’s slowing you down.

A few lines is enough. You’ll hear back within one working day with honest thoughts, a few questions, or a straight “this isn’t for us”.