AI agent for SMEs: setup budget, run cost and traps

SME AI agent: what you actually pay for

An AI agent for an SME is not priced like a single SaaS subscription. The bill splits across three layers:

• Setup: scoping, build, connectors and guardrails
• Run: models, hosting and tools
• Human ops: review (HITL), incidents and improvements

If a vendor only shows you the setup price, look at the run. If they only show the run, look at the setup and ask who the business owner will be. Until all three layers are on the table, the quote is a teaser.

For rough orders of magnitude, here are the kengdev ranges (ex-VAT, France, 2026):

The detail then depends on the tools to connect and on how risky the flow is.

Setup: where the money goes

Setup breaks down into three phases, each with its own cost line:

1. Scoping fixes the target friction, the KPI, the action allowlist and the allowed data. Without it, the build drifts.
2. Build covers the prompts and skills, the tools (CRM, mail, Drive), the tests, and a validation UI if needed.
3. Go-live adds the secrets, the logs, the token caps and the runbook.

The closer the flow gets to customers or money, the more human review and tests weigh in the quote. And an agent that "works in demo" without monitoring is not in production.

Run: tokens, hosting, support

Token cost varies a lot from one flow to the next. A simple qualification flow is cheap, while multi-agent setups, long context and tool loops multiply the bill. Measure from the first week.

For hosting, you choose between a pure cloud API and a self-hosted EU VPS. Hermes-style self-hosting adds infrastructure and maintenance; a cloud API needs less ops but creates vendor dependency.

Support covers fixes, small additions and monitoring. Plan a light monthly retainer or a ticket budget: zero support is silent debt.

HITL: the cost people forget

Human-in-the-loop is not free: someone reads, validates and edits the agent's outputs. Early on, that is normal. But if 100% of outputs still go through manual review at month 3, the autonomy design is wrong.

So calibrate autonomy by risk level: green cases go out automatically, amber cases wait in a review queue, and red cases stay forbidden for the agent. Then measure the median review time and the edit rate.

Without a business owner, you just have a bot that creates noise. The budget must include that role, even at 2 hours a week.

Quote traps

Some signals should slow you down before you sign:

• A price with no KPI attached
• A "smart agent" with no bounded flow
• No token ceiling
• No escalation plan to a human
• A delivery with no handover
• Vendors compared on setup alone

A good quote, on the other hand, lists the target process, the tools, the HITL, the estimated run and the exclusions.

Finally, be wary of leaderboard screenshots, promises to replace a full role in week one, and offers that never mention human review.

How to decide in 20-40 minutes

A clean decision fits in six points, in this order:

1. One measurable friction
2. Reachable tools
3. An identified business owner
4. A risk threshold: what must never go out automatically
5. A setup plus 90-day run budget
6. A go / no-go on a bounded POC

If the POC misses the KPI, you stop or shrink the scope. That is always better than a six-month tunnel.

For build detail, see the guide on how to create an AI agent and the human-in-the-loop article. For operations, the LLMOps guide takes over.

If you want to scope an agent on a real process, we can do it in 20-40 minutes.