What an AI System Actually Costs

The fear I hear most

Every CEO I talk to has the same fear: “What if I spend €100K and get nothing?”

It’s a reasonable fear. Most of them have already spent money and gotten nothing. Or they’ve heard stories from peers who did.

I’m going to de-risk this for you. Not by promising it’ll be cheap - but by showing you exactly where the money goes, what you should pay, and what you shouldn’t.

No “it depends.” No “let’s schedule a call to discuss.” Real numbers from real projects.

The cost breakdown nobody shows you

Here’s where your money goes in a typical AI project:

• Data preparation: 40% - Cleaning, organizing, piping, validating, labeling
• Infrastructure & integration: 25% - Servers, APIs, security, connecting to your systems
• Model development: 15% - The AI itself
• UI & workflow: 10% - The interface your team uses
• Monitoring & maintenance: 10% - Keeping it running after launch

The model - the thing everyone talks about - is 15% of the cost. The data work - the thing nobody talks about - is 40%.

This is why projects go over budget. You budgeted for the model. You forgot the data.

Tier 1: API integration - €15K-€50K

What it is: You call an existing AI model (OpenAI, Anthropic, Google) via API, wrap it in a workflow, add prompt engineering and guardrails.

What you get: Chatbots, summarization, classification, content generation, translation. The AI does the thinking; you provide the context and workflow.

Timeline: 4-8 weeks.

Hidden costs: API usage fees (per-token pricing, scales with volume), latency optimization, rate limiting.

When to choose this tier: Your problem maps to something existing models already do well. Your data doesn’t need heavy processing. You want to test the waters before committing more.

This is the lowest-risk entry point. If you’re scared of wasting money, start here. Prove the value. Then decide whether to invest more.

Tier 2: RAG system - €50K-€150K

What it is: Your data + existing model. The AI retrieves information from your documents before answering. Grounded answers, no hallucination. The model lives on someone else’s servers, but it reads your data before responding.

What you get: Internal knowledge base, document Q&A, customer support that knows your products, compliance document search, technical documentation assistant.

Timeline: 8-16 weeks.

Hidden costs: Data cleaning (always more than expected), vector database hosting, embedding computation costs, ongoing data pipeline maintenance.

When to choose this tier: Your AI needs to reason over your specific knowledge. An API alone doesn’t know your business. RAG gives the AI your company’s brain without training a custom model.

You’re not training a model from scratch. You’re connecting existing intelligence to your data. Lower risk than Tier 3, higher value than Tier 1. This is where most mid-market companies land.

Tier 3: Custom model - €150K-€500K+

What it is: Fine-tune or train a model on your data for a specific task. The model is yours. Optimized for your use case. No per-query costs.

What you get: High-accuracy specialized systems - fraud detection, predictive maintenance, custom classification, domain-specific generation.

Timeline: 4-9 months.

Hidden costs: Compute (€10K-€50K per training run), MLOps infrastructure, evaluation pipelines, retraining cadence, AI talent to maintain it.

When to choose this tier: APIs don’t meet your accuracy needs, you have enough data to train on, and the use case justifies the investment.

When NOT to choose this tier: If an API gets you 85% accuracy, the 15% improvement rarely justifies 10x the cost. Don’t jump here. Earn it by exhausting Tiers 1 and 2 first.

If a vendor recommends Tier 3 without proving Tiers 1 and 2 don’t work, get a second opinion. They’re either upselling or experimenting on your budget.

The costs nobody tells you about

Data cleaning: always 2-3x what you budget. Every company thinks their data is clean. It never is. Duplicates, missing fields, inconsistent formats, stale records, orphaned data in systems nobody remembers. Budget for it.

Integration: where projects stall. Connecting the AI to your CRM, ERP, ticketing system, or document store. Each integration has its own API, rate limits, authentication, and quirks. This is where the timeline slips.

Ongoing maintenance: 15-20% of build cost per year. Models drift. Data changes. APIs update and break things. Someone has to monitor, retrain, and fix. If you don’t budget for this, you’ll have a system that works for 6 months then silently degrades.

Talent: someone must own it after the consultant leaves. This isn’t a full-time AI researcher. It’s an existing engineer who understands the system and can debug issues. But they need training, documentation, and ownership - which means the consultant needs to provide those things.

Compliance: 10-15% of project budget if done right. EU AI Act documentation, GDPR assessments, security reviews, audit trails. Skip this and you’re saving 10% now to risk €35M in fines later.

How to avoid overpaying

Start with an audit. Know your tier before spending. An audit costs a fraction of a failed project and tells you exactly what you’re buying.

Don’t let a vendor push a custom model when an API works. This is the most common upsell. Custom models are more profitable for vendors. They’re not always better for you.

Don’t skip data work. It’s where projects fail, not where you save. Cutting data prep budget doesn’t save money - it moves the cost to the deployment phase, where it’s 5x more expensive to fix.

Fixed-scope audit, phased implementation. No blank checks. Audit first. If the audit makes sense, build Phase 1. If Phase 1 delivers, build Phase 2. You control the spend. You see results at each stage. You can stop.

The real question

Not “what does AI cost?” but “what does the problem cost, and does AI solve it for less?”

Example 1: The math works. Problem costs €200K/year in manual labor. AI solves for €80K build + €15K/year running. Payback in 12 months. Year 2+ saves €185K/year. Build it.

Example 2: The math doesn’t work. Problem costs €20K/year. AI costs €150K to build + €15K/year running. You’re losing €145K in Year 1 and €5K every year after. Don’t build it. Fix the process manually.

The math either works or it doesn’t. An audit tells you before you spend.

FAQ

Can we start small and scale? Yes. API first. Prove value. Then invest more. This is the safest path - you spend €15K to validate, not €150K to gamble.

What about open-source models? They save API costs but increase infrastructure and talent costs. The total is often similar. The real advantage of open-source is control and data privacy - you’re not sending your data to someone else’s servers.

What’s the ongoing budget? 15-20% of build cost per year for maintenance, monitoring, and retraining. If you build for €100K, budget €15-20K/year to keep it running. Skip this and your system degrades silently.