The useful distinction in business AI is not which model is behind it. It is whether the thing is allowed to change the state of your business. A system that retrieves text and paraphrases it can be judged on whether its answers read well. A system that can create a booking can be judged on whether the booking exists, and that is a question with a yes or no answer.
Most disappointment with AI in small business comes from buying the first and expecting the second. A widget that cannot act still absorbs the customer's attention, and an inquiry that would have been a phone call becomes nothing at all.
The other recurring finding is that the constraint is rarely the model. It is that there is no live source of availability to read from and no customer record to write to. The agent project turns out to be a systems project, and doing that part first is what makes the agent worth having.
What this covers
- What separates an agent from a chatbot in practice
- What an AI receptionist costs to build and to run
- Why AI support tools fail in production
- Owning the model layer instead of renting access to it
Read in this order
- AI Agent vs Chatbot Widget: What Each One Can DoBoth appear as a box in the corner of a website. One retrieves text, the other takes actions against your systems and can be held to an outcome. The distinction decides whether the thing reduces work or adds to it.
- AI Receptionist Cost: Build, Run, and the Failure ModesAnswering services quote per minute, AI vendors quote per month, and a built agent has a different shape entirely. The comparison only works once you know what each one is allowed to do and what happens when it is wrong.
- Why AI Wrappers Fail in Customer SupportThin LLM wrappers pose operational and legal risks. Citing the Air Canada court precedent, we explore why conversational AI must run in bounded JSON environments.
- Open-Weight AI Models vs Frontier Model AccessThe most capable proprietary models are getting gated, ID-walled, and repriced while open-weight models match them at a fraction of the cost. Here is the regulatory and economic case for building on models you own instead of access you rent.
Common questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions from a body of text and cannot change anything. An agent is given a defined set of tools and permission to use them, so it can read live availability, create a booking, take a payment, look up a customer, and escalate to a person with the conversation attached. The test is simple: ask what the system is allowed to change.
How much does an AI receptionist cost?
It depends on which of three products is meant. A human answering service prices per minute. A packaged AI voice product prices monthly with a usage component. A built agent carries a build cost and then a running cost of model usage, telephony, and hosting, which makes it the most expensive to start and the cheapest per interaction at volume. Only the third can be made to follow rules specific to your business.
Why do AI support tools fail once they are live?
Usually because they were given no grounding and no boundaries. A model with gaps in its knowledge will fill them, and a confidently wrong answer about price or availability creates a dispute with a customer holding a transcript. The controls that prevent this, narrow scope, real data, an explicit instruction to escalate, and reviewable logs, are the product and not an optional extra.
Do you need your own database before deploying an AI agent?
For anything that transacts, yes. An agent needs somewhere to read availability from and somewhere to write a booking to. Without a system of record the agent can only talk, which means paying agent prices for chatbot results. The honest sequence is to build the system first and add the agent to it.
What we build for this
One agent, one workflow, connected to your actual systems, with an escalation path and logs. Not a chat bubble. Read how ai agents works.