How AI chatbots answer from your own data, and when they're worth it

Older website chatbots worked like a menu: press 1 for orders, press 2 for returns. Modern AI assistants are different. They understand questions in plain language and reply in full sentences. The catch is that a general AI model doesn't know anything about your business, your prices or your policies.

The fix is to connect the assistant to your own information, so it answers from that rather than from general knowledge.

How it works, in plain language

The most common approach has a technical name, retrieval-augmented generation, but the idea is simple:

1. Your documents, FAQs, policies and product details are collected and stored in a searchable form.

2. When a customer asks a question, the system first searches that information for the most relevant passages.

3. The AI model then writes an answer using only those passages, and can show which document each part came from.

4. If nothing relevant is found, the assistant says it doesn't know and hands over to a person.

Where it helps most

AI assistants are most useful when the same questions come up again and again:

• Order status, delivery times and return policies

• Fees, timings, locations and what's included in a service

• Booking appointments or demos

• Product questions, like sizes, compatibility or which plan to choose

• Internal questions from staff about processes and policies

If your team answers similar questions on WhatsApp or email every day, that's a strong sign an assistant could save real time.

Where it doesn't

AI isn't the right tool for everything. It's a poor fit when:

• Each conversation is unique and needs expert judgement

• Mistakes would be costly and there's no way for a person to review them

• You don't have written information for it to answer from

• You only get a handful of questions a week

In these cases, a good contact form or a well-organised FAQ page may serve you better.

Keeping answers accurate

A few practices make the difference between a helpful assistant and an embarrassing one:

• Ground it in your data. Answers should come from your own documents, not the model's general knowledge.

• Test with real questions. Before launch, try it on questions your customers have actually asked.

• Show sources. Linking each answer to its source makes mistakes easy to spot and fix.

• Hand over gracefully. Customers should always be able to reach a person, with the conversation history passed along.

• Review regularly. Looking at real conversations each month shows where the assistant needs better information.

What about privacy?

Before building, decide where your data will be stored and processed, and choose AI providers and settings that don't use your data to train their models. For sensitive information, there are options for private or region-specific hosting. These choices are worth making at the start, not after launch.

What it costs to run

There are usually two parts: a one-time build cost and a monthly running cost for the AI provider, which depends on how many conversations you have. Choosing the right model for the job and adding usage limits keeps the monthly bill predictable.

A sensible way to start

Pick one channel, like your website or WhatsApp, and one set of questions, like order and delivery queries. Launch, review the conversations for a few weeks and expand from there. Starting small lets you see real value before investing more.

If you'd like to know whether an assistant would actually save your team time, we're happy to look at the questions you get and give you an honest answer.