What Should an AI Chatbot Actually Be Able to Do?

If a chatbot can only paraphrase a help centre, calling it intelligent does not make it useful. The bar should be tied to the job visitors are trying to complete.

For most websites, the following capabilities matter more than personality settings or an animated avatar.

Answer from business knowledge

The chatbot should retrieve from approved website pages, product data and policies. It should distinguish “the model knows a general fact” from “the business says this is its return policy.” Those are not interchangeable.

Useful answers are concise and traceable. When a source link would help the visitor verify details, show it. When the source is absent, say so.

Understand follow-up questions

Customers do not repeat the full subject in every message:

• “Does the Trail 2 come in black?”
• “What about size 10?”
• “Can it arrive by Friday?”

The chatbot should understand that “it” still refers to the same product. Losing that context turns chat into a slower search box.

Search by meaning, not only words

A visitor may describe a problem without using the terms on the website. “Something for a small team that needs weekly reporting” should find the relevant plan or service even if that exact sentence appears nowhere.

This requires retrieval that handles meaning, filters and business rules. It also requires a fallback when two pages conflict.

Recommend with a reason

Recommendations should connect a stated need to a real attribute. “This shoe is lightweight and available under your $120 budget” is useful. “Customers love this amazing choice” is filler unless the business has evidence for that claim.

The bot should ask one focused follow-up when a missing preference would materially change the answer. It should not conduct a twelve-question interview to recommend a T-shirt.

Take a limited action

Depending on the business, a capable agent may collect contact details, create a lead, check appointment availability or add a selected item to a cart. Each action needs a real tool behind it and a clear confirmation step.

Do not let the model imply success before the system confirms it. “I can request that appointment” and “Your appointment is booked” are different statements.

Keep useful context

Short-term conversation context is essential. Longer-term customer memory can reduce repetition across visits, but it introduces identity, consent and deletion requirements.

Remember decisions that improve service, not every sentence by default. A customer’s preferred product category may be useful. An unrelated personal remark probably is not.

Know when not to answer

This is one of the strongest signs of a well-designed chatbot. It should stop when evidence is weak, the request is outside scope, the action is sensitive or the visitor wants a person.

An honest boundary protects the customer and gives the business a useful signal about missing content. Confident improvisation hides both problems.

Escalate with context

“Contact support” is not enough. A proper handoff includes the transcript, the customer’s goal, details already collected and the reason for escalation. Otherwise automation has simply added another queue.

A practical evaluation

Before choosing a chatbot, give every candidate the same twenty real questions and five failure cases. Include follow-ups, misspellings and requests that should be refused. Then compare:

• Was the answer grounded in the correct source?
• Did it ask a useful clarification?
• Did it complete the intended action accurately?
• Did it preserve context?
• Did it stop safely?
• Could a person continue the conversation without making the customer repeat it?

FAQ answering is still valuable. It is simply the beginning of the product, not the definition of a modern AI chat agent.