Every agent platform demo ends with the same mental calculation: the cost of the tool versus the cost of doing it manually. The math usually favors the agent. What the demo doesn't show is the full cost of the agent — not just the subscription, but everything else that comes with it.
This note is a cost accounting exercise. It doesn't argue against AI agents — the case files on this site show plenty of situations where the ROI is real. But it argues for going in with the complete picture before you commit.
API costs: the line item that surprises everyone
Most agent platforms charge a flat subscription. Behind that subscription is a language model API — OpenAI, Anthropic, Google — that the platform calls every time your agent processes something. Some platforms include API costs in their subscription up to a certain volume; many pass them through or have usage tiers that escalate.
At low volumes — a few hundred agent actions per month — this is negligible. At higher volumes, it matters. An agent handling 500 email responses per day, each requiring a 1,000-token input and 400-token output using a frontier model, can generate $80–150/month in API costs alone — on top of the platform subscription.
The fix is easy once you know about it: ask the platform explicitly whether API costs are included, what the per-action cost structure is, and what the ceiling looks like at your expected volume. Run the numbers at 2x and 5x your current volume, because agents often get more use as you build confidence in them.
Setup and configuration: it's not an afternoon
The demo makes it look like you fill in a few fields and the agent runs. The reality is closer to the cases documented on this site: the yoga studio booking agent took 11 days. The inventory monitoring setup took 14 hours. These are not unusual numbers.
Setup time includes: understanding what data the agent needs to access, configuring integrations, writing and refining the instructions the agent operates from, testing edge cases, and supervising the first batch of real outputs before running it unsupervised. Don't budget setup time based on the demo. Budget it based on the complexity of the actual workflow you're automating.
Ongoing monitoring: the cost nobody budgets for
An agent running autonomously still requires human oversight. The failure patterns described in our note on how agents fail quietly are all things that a periodic check would catch — and all things that go undetected if nobody is looking.
A realistic monitoring budget for a production agent is 1–2 hours per week: reviewing a sample of outputs, checking that the agent is processing the expected volume of actions, and investigating anything that looks unusual. For most small businesses, this time comes from the same person who was doing the task manually — and that person often underestimates how much time monitoring will require when the agent produces an error that needs investigation and cleanup.
Error correction: occasional but not free
Every agent produces errors. A communication agent sends an incorrect confirmation. An inventory agent misses a reorder point. A support agent drafts a response that misreads the customer's question. The rate varies enormously by use case and how carefully the agent was configured — but zero is not a realistic expectation.
The cost of an error is the time to detect it, fix the downstream consequence (a customer who got a wrong confirmation needs a follow-up; an inventory miss might mean expedited shipping), and update the agent's configuration to prevent it recurring. Small errors at low frequency are a manageable cost. Errors on high-stakes outputs or at high frequency are a sign the agent isn't ready for autonomous operation.
A realistic total cost model
- Platform subscription: €25/month
- API costs (included in this tier): €0/month
- Monitoring time: 1.5 hrs/week × €40/hr equivalent = €240/month
- Error correction (average): 0.5 hrs/week × €40/hr = €80/month
- True monthly cost: ~€345 (not €25)
The agent saves 2.5 hours of manual work per day in this example. At the same €40/hr equivalent, that's €2,400/month in saved time. The ROI is still strongly positive — but the calculation is different from "€25/month versus doing it manually."
The point isn't that agents are expensive. It's that the cost structure is different from what the demos suggest — heavier upfront, with an ongoing time component that belongs in the model. An agent that saves you two hours a day and costs you one hour a week in monitoring is still a good deal. An agent that saves you twenty minutes a day and costs you two hours a week in monitoring is not.