Playbook · ai chatbot vs ai agent for business
A chatbot answers. An agent acts. Most businesses are quoted the wrong one.
Chatbots answer known questions from your documents. Agents take actions in your systems. The price gap is $4,000 to $15,000 and the wrong pick shows.
Lasse Pettersen
A chatbot answers. An agent acts. Getting quoted for one when you needed the other is the most common expensive mistake in this category, and the gap between them is roughly $4,000 to $15,000.
Statistics Canada put virtual agents and chatbots at 28.2% of Canadian business AI use in the second quarter of 2026, third behind data analytics at 36.6% and text analytics at 34.5%. It is a mature use, which means the failures are well documented rather than novel.
The four tiers, with what each actually is
| You want | What it is | Rough cost |
|---|---|---|
| Answers to known questions | A retrieval bot grounded in your own documents. It quotes your text and does not invent. | $4,000 to $7,000 |
| Qualify and book | An agent that asks, scores against your rules, and writes into your calendar or CRM. | $7,000 to $12,000 |
| Handle a process end to end | An agent with tools, permissions and a human approval step where money or commitments are involved. | $12,000 to $15,000+ |
| A box that says hello | Nothing. Spend the money on an internal automation instead. | $0 |
The jump between rows one and two is where the price moves, and it is not about conversation quality. It is about write access. A system that only reads is a system whose worst failure is an unhelpful answer. A system that writes into your calendar can double-book a technician, and everything about how it is built has to account for that.
Grounding is the whole job
An assistant answering from a general model will answer confidently about your return policy, your service area and your prices, and it will be wrong at a rate you cannot predict. That is not a defect in a particular product. It is what a language model does when nobody gave it your facts.
A grounded system works differently. The question is used to retrieve passages from documents you supplied, and the answer is composed from those passages. When the documents do not cover the question, the correct behaviour is to say so and hand off to a person with the conversation attached.
Building that grounding is most of the cost, and most of it is not technical. It is finding out what your policies actually are. In a large share of the businesses where I have done this, it is the first time anybody has written the answers down in one place, and several of them turn out to disagree with each other. That discovery is worth something on its own.
Handing off to a human is a feature, not a failure. A system that escalates cleanly beats a system that answers everything, because the second one is answering some things wrong.
Where a chatbot genuinely earns its cost
Three situations, all of which share the same shape: high question volume, stable answers, and the answers already exist in writing.
A Niagara Falls hospitality business in season. The same questions arrive thousands of times: parking, accessibility, check-in times, what is included, whether it runs in the rain. Every answer is already on the website.
A Burlington food processor or precision manufacturer. Customers and auditors ask about specifications, allergen statements, certificates of analysis and inspection records. Somebody spends 40 minutes finding an answer that definitely exists in a shared drive organised by a convention that lives in one person’s memory.
A Collingwood short-term rental operator. Guest questions, constantly, answerable from a document that mostly already exists.
In all three the test passes: the questions repeat, the answers are authoritative, and the failure mode of getting one wrong is embarrassment rather than harm.
Where an agent earns its cost, and where it does not
An agent is worth building when a person is currently doing lookup and data entry inside a conversation. Answering “where is my load” from your own dispatch data. Qualifying an enquiry against rules you can state. Booking a standard job into a calendar with the right duration and the right technician.
It is not worth building where judgement carries a consequence. Two rules I hold to on every build:
No model produces a number that has to be correct. A model that composes an invoice total will eventually compose the wrong one. Automation moves data and applies rules; where judgement is required it drafts and a person approves.
No model issues a decision that affects a person. Approvals, refusals, prices, assessments. It can assemble everything up to the decision, and the decision is signed by somebody who can be asked about it.
Those two lines are not caution for its own sake. They are the difference between a system you can leave alone and one somebody has to check, and a system somebody has to check has not saved you the labour it was sold on.
Before features, the privacy question
An intake agent collects personal information by design. That is easier to handle properly than the accidental cases, because you get to design it, but it does have to be designed.
The Office of the Privacy Commissioner of Canada found in May 2026 that OpenAI’s initial training of ChatGPT did not comply with Canadian privacy law, and the guidance around it treats prompts, retrieval corpora and outputs as personal information handling like any other. In practice a public-facing agent that collects a name, a phone number and a description of a problem needs a stated purpose, a retention rule for the transcripts, and a processor that will sign an agreement. That is achievable, it is not free, and it belongs in the quote. The reasoning is set out in what PIPEDA means for your AI tools.
The retrieval case carries a quieter risk. Point a system at a shared drive and everything in that drive becomes reachable through a new interface, including the folder of scanned identity documents somebody saved years ago. The corpus is the exposure, not the model.
How to size it before spending anything
For a chatbot: count the questions. Pull 200 recent enquiries, sort them into buckets, and see how concentrated the top ten are. If those ten cover more than half, the volume is there. If the enquiries are all different, you have a sales conversation rather than a support load, and a bot will make it worse.
For an agent: count the calls arriving outside working hours and how many go unanswered. Your phone provider can tell you in one email. The loss is invisible because you never hear about the job that called somebody else, and it is frequently larger than any internal saving available in the same business.
Which of those two numbers is bigger decides the build. That comparison, done properly with hours and dollars on both sides, is what the $999 assessment produces, and the fee comes off the build. If the internal number wins, the answer is workflow automation instead, and the ranking method is set out in which jobs to automate first.
What handover should include
Whichever you build, ask for these before you agree:
- The system running on accounts in your name, with the model provider billed to you.
- The document set it is grounded in, editable by your staff without calling the supplier.
- A transcript log you own, with a retention period you chose.
- An escalation path to a human, and a written rule for when it fires.
- The monthly running cost, seen before you agreed to build, not after.
If any of those is awkward to ask for, you have learned more than the answer would have told you.