A chatbot is the right choice when users arrive with questions in their own words and the answers exist in your documentation or data, such as support, onboarding and internal knowledge. It is the wrong choice when the task has a known structure that a form or a button handles faster, or when the AI's job is to process data rather than talk. Most products that need AI need custom features embedded in the workflow, not a conversation, and many need both: a chatbot for questions and embedded AI for work.
When founders say they want AI in their product, they usually picture a chat window. Sometimes that is exactly right. Often it is the wrong shape for the job, and the feature that would actually help is invisible: a field filled in automatically, a list sorted intelligently, a document read for you. This guide helps you tell which you need.
What a chatbot is good at
- Answering questions from your content. Support, product documentation, policies, internal knowledge bases. The user asks in their words, the assistant answers from your sources and shows them.
- Onboarding and guidance. "How do I..." questions during first use, answered in context.
- Triage. Understanding what the user needs and routing to the right place or person.
- Tasks with unpredictable input. Where the user's request could take many forms and a form would need fifty fields.
What a chatbot is bad at
- Structured tasks. Booking a slot, entering an order, changing a setting. A form with three fields beats a conversation with three exchanges every time.
- Precision. Anything where the exact value matters and a misunderstanding costs money.
- Discovery. Users do not know what to ask a blank box. A menu shows them what is possible.
- Processing data at volume. Classifying ten thousand tickets is not a conversation.
- Being the only interface. Products that hide their functionality behind chat frustrate users who know what they want.
The four kinds of custom AI feature that are not chatbots
| Kind | What it does | Where it appears in the product | Example |
|---|---|---|---|
| Extraction | Reads documents, images or messages and fills in structured data | A form that is already filled in when the user opens it | Invoice fields captured from a photo |
| Classification and ranking | Sorts, routes, prioritises, flags | Lists that are in the right order, queues that route themselves | Support tickets routed by topic and urgency |
| Generation in context | Drafts text or content where the user is working | A suggested reply, a draft description, a summary at the top of a thread | A reply drafted inside the support tool |
| Prediction | Estimates a number or a likelihood from history | A forecast on the dashboard, a risk score on a record | Expected demand next week per product |
None of these involve talking to the product. They make the product do more of the work. For most business applications, they deliver more value than a chat window, and users often do not notice them as AI at all. Our guide to AI features that actually work lists which ones reliably succeed.
How to decide
| Situation | Choose |
|---|---|
| Users have questions and the answers are in your content | Chatbot, bounded to your sources |
| Users arrive with requests in unpredictable forms | Chatbot for intake, custom features for the work |
| The task is structured and repeated | Custom feature embedded in the flow, no chat |
| The value is in processing data the user never sees | Custom feature, invisible |
| You want to reduce support load | Chatbot on documentation plus classification for routing |
| You want to reduce data entry | Extraction, no chat |
If you build a chatbot, build it bounded
- Give it sources. It answers from your documentation and data, and it shows where the answer came from.
- Give it a boundary. It says what it cannot help with and hands off to a person, cleanly.
- Give it an escape. A visible way to reach a human at any point.
- Measure it. Resolution rate, hand-off rate, and user ratings per conversation.
- Keep the rest of the product usable without it. Chat is an addition, not a replacement for navigation.
Cost and effort
A bounded chatbot on your documentation is a few weeks with a hosted model. Custom features range from a week for classification to a couple of months for extraction with a review workflow. Both are covered in what adding AI costs. The more expensive mistake is building the wrong shape: a chatbot for a structured task, or a custom feature for a question-answering need.
How 7L approaches it
We ask what the user is trying to do and where they spend time, and we choose the shape from that. Many products end up with both: a bounded assistant for questions and embedded features for work. What we avoid is the chat window as a default, because it is the most visible way to add AI and frequently the least useful. Describe the job your users are doing and we will suggest the shape.
Frequently asked questions
Can a chatbot replace my support team?
It can answer the common questions that your documentation covers, which is often a large share of tickets, and route the rest. It cannot handle the unusual, the angry or the ambiguous well. Plan for it to reduce support load, not remove it.
Is a chatbot cheaper than custom AI features?
A bounded chatbot on existing documentation is one of the cheaper AI features to build. Whether it is the right one depends on whether users have questions or tasks. The wrong shape is expensive at any price.
Should the chatbot have a personality?
A consistent, plain and honest tone that matches your brand. Avoid pretending it is a person. Users prefer a capable assistant that admits limits over a chatty one that does not.
Can I use a general AI assistant instead of building one?
For internal use, often. For your product, a general assistant does not know your content, your data or your users, and you cannot bound or measure it. Building on a hosted model with your sources gives you both.
What about voice?
Voice is a chatbot with a different input, and the same rules apply. It suits hands-busy situations and accessibility, and it is worse than a screen for anything with options to compare.