How much does an AI agent cost? Real numbers from real builds

The headline numbers

What we charge, by engagement type:

These are real ranges from real engagements, not pitch-deck numbers. Where you land in each range depends on integration complexity, document/data variety, and how clean your existing systems are.

What drives cost

Five factors, ranked by impact:

1. Integration count and complexity

The single biggest cost driver. An agent that talks to one system (your own database) is cheap. An agent that talks to NetSuite + Salesforce + Microsoft 365 + a custom data warehouse + your CRM is expensive — because each integration needs auth, schema validation, retry, idempotency, observability.

Budget rule of thumb: each meaningful external integration adds €5,000–€15,000 to the build, depending on whether it's a well-documented modern API or a legacy system.

2. Data variety

A document agent processing invoices from 3 vendors with similar layouts is cheap. The same agent processing invoices from 200 vendors with wildly varied layouts costs more — because the eval suite, the schema, and the reviewer queue all scale with variety.

Budget rule: add ~20% to the document agent build if vendor / format variety is high.

3. Compliance and auditability

Regulated industries (healthcare, finance, legal, EU public sector) add cost — data residency requirements, audit trail granularity, approval gates, encrypted-at-rest, role-based access. None of these are technically hard; they're each an engineering tax.

Budget rule: add 15–30% for heavily regulated environments.

4. Channel count

Text-only agents are cheaper than agents that also do voice. Add a voice surface and you add Twilio integration, realtime model usage, voice-specific evals, recording compliance. Add a separate mobile surface and you add another platform's build effort.

Budget rule: each new channel adds €15,000–€30,000.

5. Production-grade rigor

The difference between "it works in the demo" and "it's been running for 6 months without a Sunday-evening outage." Production rigor includes: real evals, real observability, real on-call runbooks, real load testing, real failover paths.

We always build production-grade. Some shops sell "MVPs" that skip this. Their MVPs cost less; their MVPs also break.

Per-call infrastructure cost

The other half of the cost story is what each agent invocation costs in production. Honest numbers from our deployments:

Plus the operational overhead — observability tooling (€100–€500/month at typical volumes), hosting (€50–€500/month), and minor fixed costs.

The retainer question

After the build, you have three options:

1. In-house: take the code and the docs, run it yourselves. Many clients prefer this. We hand over runbooks, eval suites, and CI/CD configured to your team.
2. Light retainer: we handle evals, prompt tuning, model upgrades, occasional firefighting. €1,500–€2,500/month.
3. Full retainer: we operate it. Includes on-call, ongoing improvements, monthly business review. €3,500–€5,000/month and up.

Roughly 60% of our clients take option 2 or 3 because operating production AI agents is its own job. The rest take it in-house.

What's NOT included

A few things we explicitly don't bundle into the build price:

• LLM API costs (Anthropic, OpenAI, Google). These pass through directly. We help you set spend caps and surface costs on the dashboard.
• Twilio / Telnyx voice charges. Per-minute pass-through.
• Vector store hosting (if you use a managed one like Pinecone). Usually trivial; sometimes €50–€500/month at scale.
• Custom model training. We rarely recommend it; if you need it, separate engagement.
• Hardware / on-prem infrastructure. We design for cloud; if you need on-prem, scope changes.

Common pricing traps

"Cheap" agencies that don't include evals

Some shops sell €15,000 "AI agent MVPs" that ship without an eval suite, observability, or proper guardrails. The agent works in the demo and quietly degrades in production. Three months later you pay someone else 2× the price to do it properly.

Fixed-price quotes before discovery

Anyone quoting a fixed price for a production AI agent from a one-hour intro call is either over-quoting (to cover the unknown) or under-quoting (and will discover scope mid-build). Insist on a paid discovery sprint as the first engagement.

"Per-agent" SaaS pricing for unique workflows

Some platforms charge per agent or per AI workflow on a SaaS basis. Fine for generic use cases. Expensive for unique ones. Math out the 24-month cost vs custom-build before committing to a multi-year SaaS contract.

Hidden vendor lock-in

The agent works great with Vendor X's tools. Migrating off costs €40,000. This shows up most often with low-code platforms (Power Automate at extreme scale, Microsoft AI Studio with proprietary connectors). Worth verifying portability up front.

The buyer checklist

Before you sign anything, confirm:

• Scope was defined in a paid discovery sprint, not a free pitch call.
• Eval suite is part of the build, not an add-on.
• Observability is part of the build, not an add-on.
• Code is yours, in your git repo, on day one.
• Provider abstraction exists — you can swap LLM providers without a rewrite.
• Approval gates exist on every irreversible action.
• Cost ceilings are configurable per agent.
• Handover is built into the engagement — runbooks, eval suites, CI/CD configured for your team.

If any of these are missing, ask why.

Where to go next

For the broader hiring-an-agency checklist, see 12 questions to ask when hiring an AI agency. For our internal playbook on how we ship efficiently, see The AI Development playbook.

If you have a project in mind and want a fast cost take, drop us a note. One paragraph is enough.