The €5k and €15k version of the same project look identical in the demo
Two vendors quote you "automate quote-to-invoice with GPT" and land at €5,200 and €14,900. Both will show you the same demo: a Zapier or n8n workflow, a prompt, a webhook firing against your ERP. Both work. The difference is almost entirely in what happens when the API times out, the customer data format shifts, or the agent hallucinates that it posted an invoice when it didn't.
The visible build—prompt engineering, webhook integration, the n8n node wired to your CRM—is about 20% of the cost. The other 80% is idempotency keys so you don't double-charge customers, external-state assertions that actually verify the invoice landed in Holded or Odoo instead of trusting the agent's confidence score, retry semantics with exponential backoff, an eval harness that catches regressions before production, and observability that lets you debug at 2am without guessing. The €5k vendor skips this. The €15k vendor doesn't.
Three weeks in, when your CRM has duplicate entries and your finance team is repairing data in bulk, the cheaper quote retroactively becomes the expensive one. We've gotten this wrong before on both sides—pushing clients to cheaper builds and paying the debt in escalations, and over-engineering for clients who genuinely didn't need the robustness.
The €1,500 assessment: what you're actually paying for
Before we quote a build, we do a 1.5–2 week assessment. This is where the real price variance lives.
The assessment is process mapping (how is the quote-to-invoice actually happening now?), a data access audit (who can write to which system, and under what conditions?), and an integration surface audit (HubSpot, Factorial, Holded, Odoo, custom ERP—each has different retry semantics and rate limits). The deliverable is an architecture doc, a token-cost projection per run at your actual volume, and a build estimate with a confidence interval.
We refuse to quote a build without one. The "simple RPA replacement" that sounded like 20 lines of prompt engineering turns out to need OCR plus human-in-the-loop checkpoints plus audit logging because it touches payroll. The €4,000 flat quote a vendor throws back from a 30-minute call should be a red flag—it means they're either guessing or they've pre-decided their scope and will fight you when reality doesn't fit.
Sometimes the honest assessment conclusion is "don't build this, buy Make.com and a €200/month seat." We've said that. Clients respect it more than a upsell.
What pushes a build from €5k to €15k (and what doesn't)
The drivers that actually move the needle: the number of external systems with write access (each one doubles your retry surface and your state-assertion complexity). SLA on tail latency—if you need a 6-tool chain to stay under 8 seconds p99, that's engineering work. Whether you need an eval harness (you do if money or customer comms are involved). Regulated data (GDPR plus sectorial compliance means audit logging and data retention policies that cost real time).
Human-in-the-loop checkpoints aren't just a Slack message. They're a state machine, a UI, a way to route back to agents if the human rejects the suggestion. That's 4–6 days of work, not 4 hours.
The non-drivers people overpay for: model choice (GPT-5 vs Claude vs Qwen rarely changes build cost, only token throughput at scale). "Custom RAG" when keyword search over a 20-document knowledge base would ship just fine. Streaming endpoints when polling every 500ms is good enough.
We price multimodal and voice separately. Streaming budgets are their own engineering problem—you're managing concurrent connections, frame buffering, tail latency in the audio loop. It's not a checkbox on the same project.
Fixed price vs hourly: pick the one that penalizes the right party
Fixed-price works when the assessment is done and the integration surface is frozen. We quote €5k–€15k against a signed scope: the systems you'll connect, the workflows, the SLAs. If scope moves, we renegotiate.
Hourly (€80–€140/hour in Spain in 2026) is honest when the external API is undocumented or the client's data model is still moving. You pay for the unknown, and so do we—neither party can pretend the scope is certain.
Our default hybrid: fixed-price build plus hourly retainer for the first 60 days of production, because that's when the real failure modes surface. A client on €10k fixed + €120/hour retainer for two months pays for confidence, not gambling.
Watch for "fixed price, unlimited revisions." It's priced for the worst-case client, and you're subsidizing someone else's chaos. Revisions that stay within the signed scope are free. Scope creep costs.
Our milestone structure: 30% on assessment sign-off, 40% on staging with eval harness green, 30% after two weeks in production. The vendor is incentivized to actually ship something that works, not just hand over code.
The line items clients always underestimate
Token costs at scale. A "cheap" agent at €0.03 per run becomes €900 per month at 1,000 runs per day. This needs to be in the quote, projected against your actual expected volume. If the vendor hand-waves it, they're pricing for a pilot and you're paying for production.
Agents can't self-verify external state—you have to assert it in the loop. That's a separate line item: 2–5 days of work to build the assertions, write the tests, and integrate them into your workflow. Non-negotiable for anything touching money or customer comms.
Observability stack. Langfuse, Helicone, Datadog spans aren't free. Without them, debugging a 2am production regression is guesswork. Budget 10–20% of your build cost for observability infrastructure and dashboards.
Spend caps and graceful degradation. What happens when OpenAI 429s at 11pm on Black Friday? Does your agent queue the work and retry, or does it fail silently? Does it fall back to a cheaper model? Does it page you? This logic has to be architected and tested. It's not free.
Handover docs and runbooks. If the vendor won't quote this separately, they're not planning to write them. You'll inherit an undocumented n8n instance that nobody but the original engineer can touch. Budget 3–5 days.
How to read a Spanish AI automation quote in 2026
Does the quote name the external systems and the specific write operations? "Automate order processing" is vague. "Write to Holded via REST API using idempotency key X with retry on 429" is a contract. If the vendor hasn't named the systems, scope will move.
Is there a line for state assertion and verification, or just "agent execution"? If it's just execution, you don't have failure recovery. The risk is on you.
Token cost projection with assumed volume. Present or hand-waved? Anything hand-waved becomes your problem at month two when your monthly bill is €4,000 higher than expected.
Who owns the API keys and the observability dashboard on day 61? If it's the vendor, can you export your data? If it's you, do you have the credentials? Ownership matters.
Exit clause: can you take the prompts, evals, and infrastructure-as-code with you, or is it locked in the vendor's n8n instance forever? If you can't leave, you can't negotiate.
Send us your quote
If you've got a quote on your desk right now, send it to us at /contact. We'll tell you in 30 minutes which line items are missing and whether €5k, €10k, or €15k is the honest number for what you're actually asking for.