An AI bot looks brilliant in a demo: it answers in plain language, understands typos, has no button tree. Then a real customer asks whether you deliver to their city on a Saturday, and the bot confidently invents an answer. The difference between those two bots is not the model — it is where the bot gets its facts and what it is forbidden to do. Below: where an AI assistant genuinely beats a scripted bot, how to wire in a knowledge base, what it costs per month in tokens, and the failure modes that lose you a customer.

How an AI assistant differs from a scripted bot

These are two different tools, not an old and a new version of one. A scripted bot is deterministic: you know exactly what it will say. An AI assistant is probabilistic: it composes every answer from scratch, which is precisely why it needs constraints a scripted bot never requires.

The rule we apply on every project: AI answers, buttons act. Booking a slot, creating an order or taking a payment must never be handed to a language model — those steps belong in ordinary deterministic code. If you have not settled your base scenario yet, start with our guide to bot scenarios that pay back.

Where AI is genuinely cheaper than a person

What these tasks share is that a correct answer already exists somewhere in your data, and the cost of being wrong is low. Where no correct answer exists, or a mistake costs money, AI should assist a person rather than stand in for one.

The knowledge base is 80% of the result

Quality is determined far more by what you let the model read than by which model you picked. Without your data it answers in generic phrases scraped from the open web. With a clean knowledge base and a rule against going beyond it, the same model answers accurately and boringly — exactly what support needs.

1. 1

   Collect the sources

   Price list, delivery and returns terms, service descriptions, FAQs, and above all your archive of real customer conversations. That archive gives you the wording customers actually use, rather than the wording your marketing team prefers.

2. 2

   Remove contradictions

   The most common cause of a lying bot is two documents quoting different prices. The model picks one at random. One evening of reconciliation saves a month of complaints.

3. 3

   Chunk and index

   Documents are cut into 300–800 word pieces and turned into vectors. On each question the system retrieves the three to five closest chunks and hands them to the model as context. That retrieval step is what stops the bot inventing.

4. 4

   Write the behaviour rules

   A ban on answering outside the retrieved context, a mandatory "I do not know, let me fetch a colleague", response format, tone, length. This is the system prompt, and it gets rewritten about ten times in the first month.

5. 5

   Replay 100 real questions

   Take the archive and compare bot answers with what your team actually replied. Below 80% agreement, the knowledge base is incomplete — the model is rarely the problem.

# Minimum constraints before a bot is allowed near customers
system:
  grounding: strict          # answer only from retrieved chunks
  max_context_chunks: 5
  on_missing_context: handoff  # no data -> pass to a human
  never_answer:
    - individual_pricing
    - refunds_and_disputes
    - legal_or_medical
  always_include_source: true  # cite the document each answer came from
  max_reply_chars: 600
  temperature: 0.2

What it costs every month

Token pricing worries clients more than it should. A support conversation is typically three to eight exchanges, and in our projects that lands around one or two cents on mid-tier models. The real money goes not on answers but on bloated context: stuffing the entire price list into every request instead of five relevant chunks multiplies the bill many times over.

1. 1Do the volume maths. A thousand conversations a month at $0.01 each is $10 of tokens. Even if that estimate is ten times off, it is not the line item that should decide whether you build.
2. 2Keep context small. Five chunks of 500 words instead of a full catalogue is roughly a twentyfold cost difference at identical answer quality.
3. 3Cache the common questions. Half of all enquiries are the same twenty questions. A cached answer costs nothing and returns instantly.
4. 4Use a smaller model for classification. Deciding what a message is about needs the cheapest tier; the expensive one is only needed to compose a difficult answer.
5. 5Set a per-user limit. Thirty messages a day per account protects you from both abuse and a surprise invoice.

Three ways AI loses you a customer

• Confident invention. The bot quotes a price that does not exist or promises delivery you do not offer. Fixed by strict grounding and a forbidden-topics list — not by "a better prompt".
• The trap with no exit. A customer types "give me a human" three times and the bot keeps offering articles. The handover button must be visible from the first message, and any mention of a person should trigger it.
• Silent decay. Prices changed in February; the document in the knowledge base is still the November one. The bot confidently quotes the old figure for months. Tie knowledge base updates to the same process that updates your website.

All three failures are organisational, not technical. That is why an AI assistant is worth launching only once your processes are defined: who updates the data, who reads the logs, who answers escalations. If none of that exists yet, read up on business process automation before you read about models. And decide where your line between bot and human sits — our comparison of chatbots and live chat works that through.

Frequently asked questions about AI chatbots

In short

• AI answers, buttons act: keep booking, payments and order changes in deterministic code.
• The knowledge base decides 80% of quality, not the model. Contradictory documents are the top cause of invented answers.
• Tokens are rarely the problem: keep context small and cache the repeat questions.
• Forbidden topics, a mandatory "I do not know", and a visible exit to a human are minimum safety, not options.
• With nobody responsible for keeping documents current, an AI bot becomes a source of stale prices.