An AI agent doesn't cost what the model's price tag says, and it's the most common budgeting mistake we see when a small or midsize business compares vendors. Someone asks "how much does it cost to bring AI into my company?" and gets a number per token, per message, or per user, and decides based on that. The problem is that number describes one of four line items, and the other three are what actually determines whether the project turns a profit or becomes a hole nobody saw coming.
We won't give you figures (ours aren't published, and a generic vendor's would be made up), but we will give you the decision rule we apply with every client before they sign anything: compare vendors on all four line items, not one.
The "price per token" trap
Price per token is easy to advertise, easy to put in a comparison table, and easy to sell as "AI is cheap." It's real, but it only covers the cost of the model processing text. It says nothing about what it actually takes for that model to become something that works inside your business: reading your ERP, writing to your CRM, doing something people trust without checking every output by hand.
Comparing two vendors only by token price is like comparing two cars only by the price of a tank of gas. The cheaper one at the pump might be the one that costs more at the shop.
The four real line items
These are the four cost items of an AI agent in production, in the order they usually show up (and surprise people):
| Item | What it covers | When it shows up |
|---|---|---|
| 1. Model | Model usage: input and output tokens, the context size it needs to reason well | From the first message; the only one visible in the initial quote |
| 2. Integration | Connecting the agent to your ERP, CRM, or real data system, and maintaining that connection when those tools change their API | Before the agent goes live, and repeatedly after |
| 3. Human oversight | Someone reviewing what the agent does, fixing its mistakes, and adjusting the rules, especially in the first weeks | Ongoing, heaviest at the start, lower once the agent is well built |
| 4. Governance and testing | Testing the agent before it touches real data, guardrails, role based permissions, what happens if the vendor changes the model or its pricing | Before launch, and every time something in the environment changes |
Only the first item shows up in most quotes you'll see out there. The other three are real work, and if nobody accounts for them, they don't disappear: someone pays for them anyway, usually as your own team's hours cleaning up what the agent broke, or as a project that stalls because nobody budgeted for the integration part.
Why the other three items aren't optional
There's a temptation to treat integration, oversight, and governance as "extras" you can cut when the budget is tight. In our experience building agents that actually write to real systems, it works the other way around: cutting any of the three doesn't lower the cost, it just moves it somewhere else.
- Cut integration, and the agent ends up disconnected from the real system, turning into an expensive chatbot that answers questions but does nothing; we cover this in more detail in why an agent without write access is an expensive chatbot.
- Cut oversight, and the mistake the agent makes in week three goes unnoticed until it already hit a real customer or a real invoice.
- Cut governance and testing, and you're letting a new system touch production data without ever checking what it does when something goes wrong, which is exactly what how an AI agent gets tested before touching real data is meant to prevent.
The AI sales agent we built for an industrial distributor with over 50,000 products is a good example of why the four items travel together: it quotes over WhatsApp inside Odoo, validates the customer, applies the right rate, checks stock, and creates the real order. The model's price is the smallest part of that work; the Odoo integration, the rules for what it can and can't touch, and the testing before it was allowed to talk to real customers are what make it run today without constant supervision. You can see the full case in AI sales agent inside the ERP.
The question to ask your vendor
Before comparing two AI offers on price, ask this and compare the answers, not just the token number:
- What part of the price covers only model usage?
- Who does the integration with my ERP or CRM, and what happens if that tool changes its API next year?
- How many hours of human oversight does this agent need per month, and who does them, you or my team?
- How is the agent tested before touching real data, and what guardrails does it have if it gets something wrong?
If a vendor can't answer all four with something concrete, you don't have a price: you have one line item, and the other three are coming your way regardless, just without warning.
Checklist before you sign
- I have the model's price, and I know what unit it measures (token, message, user, minute).
- I know who does the integration with my real systems, and whether maintenance is included or separate.
- I know how many hours of human review the agent needs per month, and who owns that.
- I know how the agent gets tested before it touches real data, and what happens if it fails.
- I compared vendors on all four line items, not just the first one.
The rule, in one line
A price per token isn't a budget, it's one line item. If your vendor only talks about that one, the conversation hasn't really started yet.
At AutoBoost we build custom software and AI integrated into the client's real system, not generic templates, so all four line items are on the table from day one, not as a surprise in month three. If you want to see what this would look like for your business, check out our services or get in touch and we'll go through your specific case together.
