Everyone with a system prompt calls themselves an AI agent development company now. Here's how to tell engineering from theater before you've paid for the difference.
A real one delivers four layers, not one: the agent behavior (prompts, flows, guardrails — designed, not improvised), the machinery (retrieval over your data, tool integrations, orchestration), the quality system (eval suites that measure behavior before and after every change), and operations (monitoring, cost control, fallbacks). Shops that only do the first layer sell demos. Production agents live or die on layers two through four — the split we describe across our agent designer and Engenheiro de IA roles.
On quality: How do you measure agent behavior? (The only good answer contains the word "evals".) What's your regression process when prompts change? Can we see an eval suite from a past project? On production: What happens when the model provider has an outage? How do you control cost per conversation? Who gets paged when the agent misbehaves at 2 a.m.? On data: Where does our data live and what touches it? How is retrieval kept current when our content changes? On the relationship: Who owns the IP? (You should.) What exactly transfers at handover? What does the engagement cost — fixed entry or open-ended discovery? And the closer: what have you refused to automate, and why? Companies with real experience have refusals; sellers don't.
| Model | What it tells you |
|---|---|
| Fixed-price audit → scoped delivery | Provider owns discovery risk; incentives aligned. The mid-market standard worth demanding. |
| Open-ended T&M discovery | You're funding their learning. Acceptable only with tight sprint gates. |
| Per-resolution / outcome pricing | Mature for support agents; verify the counting methodology. |
| Big-consultancy program pricing | What enterprises pay. If you're mid-market, you're subsidizing overhead built for banks. |
Red: no evals anywhere in the proposal; "proprietary platform" that means lock-in; case studies without named workflows or numbers; a team that's all prompt-writers and no production engineers. Green: fixed-price entry, IP transfer as default, human-in-the-loop designed where errors are expensive, capability transfer contracted — the pattern enterprise buyers now demand, and the way our own implementation pods are built (first agent live in six weeks, your team trained to run it).
Agent development is iteration-heavy — daily contact between your ops people and the builders. That's why time-zone overlap matters more here than in classic outsourcing, and why Benelux buyers increasingly source this work nearshore (CET ±1) instead of far-shore: same-day iteration at rates well under local consultancies. If you'd rather own the capability than buy the project, the same vetting applies to hiring AI developers directly.
Mid-market entry: fixed-price audits and first-agent deliveries in the low tens of thousands of euros; enterprise consultancy programs run an order of magnitude higher. Insist on a fixed-price entry before any open-ended engagement.
With a provider that has patterns to reuse: an audit in two weeks and a first production agent by week six is a realistic, demandable timeline for a well-chosen workflow.
Sequence it: buy the first outcome from a team with evals discipline (fixed price, capability transfer), then decide between a partner pod and hiring your own AI engineers based on roadmap depth.
Fixed-price audit, evals in every delivery, IP transfer as default — and refusals when automation isn't worth it.
Marque uma chamada de apresentação