The quote says 80 thousand euros. After a year, you've spent 300 thousand. Nobody lied — they just didn't tell you the whole truth. This is the most common pattern in enterprise AI projects, and it happens because the visible cost is a fraction of the real one.
GRAL believes cost transparency is a prerequisite for any serious project. This article is the document we wish we could hand every company before the first meeting.
The AI Cost Iceberg
The cost of an AI project is like an iceberg. The visible part — the model, initial development, software license — is 20-30% of the total. The rest is submerged.
The Visible Part
Initial development and integration. This is the number on the quote. It includes: requirements analysis, model development, integration with existing systems, testing, initial deployment. For a medium-complexity enterprise project, we're talking 50-200 thousand euros. This is the part everyone understands and everyone quotes.
Licenses and infrastructure. If you use cloud services for AI (GPUs, model APIs, storage), the initial cost is often low — a few thousand euros per month. Seems manageable. The problem is it scales with usage, and production usage is always higher than testing.
The Submerged Part
Data preparation. At GRAL, we estimate that 40-60% of an AI project's time goes into data preparation. Cleaning, normalizing, labeling, validating. If your data is in good shape, this phase is quick. If it's fragmented across ten different systems with inconsistent formats — as in most companies — it's the biggest cost item.
Concrete numbers: manually labeling a dataset for computer vision can cost 10-50 euros per hour of specialized work. If you need 10,000 labeled images, do the math.
Production infrastructure costs. The model that ran fine in testing on 100 requests per day now handles 10,000. Cloud costs scale. GPUs aren't cheap. A single medium-complexity model in production can cost 2,000-15,000 euros per month in infrastructure, depending on volume and required latency.
Monitoring and maintenance. A production AI model isn't traditional software that "works until it doesn't." Models degrade over time — a phenomenon called model drift. Data changes, context evolves, performance drops. You need continuous monitoring and periodic retraining.
Typical cost: 15-25% of initial development cost, every year. Forever. Or at least as long as the system is in production.
Internal personnel costs. Someone in your company needs to manage the vendor relationship, supervise the system, handle anomalous cases, train users. Even if you don't hire a data scientist, existing people's time has a cost.
Opportunity costs. While the team is engaged in the AI project, they're not doing other things. Alignment meetings, testing sessions, feedback collection — all subtract time from other activities. This cost is invisible in budgets but real in operations.
The Most Common Cost Models
Fixed-Price Project
How it works: you pay a defined amount for a defined deliverable. "We'll build you a visual quality control system for 120 thousand euros."
Pros: predictable budget, vendor's risk. Cons: the vendor has an incentive to minimize scope. Every additional request becomes a paid change request. The result is often the minimum viable product, not the optimal system.
When it makes sense: for well-defined problems, with clean data and stable requirements. Rare in the real world.
Time & Materials
How it works: you pay for the vendor team's actual hours. Typical rates: 800-2,000 euros per person per day, depending on seniority.
Pros: maximum flexibility, the project adapts to reality. Cons: unpredictable budget. Without tight governance, costs balloon. The vendor has no incentive to be efficient.
When it makes sense: for exploratory projects or with evolving requirements. But you need a competent internal project manager controlling spending.
Platform with Subscription
How it works: you pay a monthly or annual fee to use an AI platform. Includes infrastructure, updates, basic support.
Pros: predictable costs, maintenance included, faster time-to-value. Cons: less customization, vendor dependency, cumulative costs long-term.
When it makes sense: when your problem fits what the platform solves well. Don't force a custom problem into a standard platform.
How to Calculate Real TCO
The Total Cost of Ownership of an AI project over 3 years includes:
Year 1:
- Assessment and data preparation: 20-40% of total budget
- Development and integration: 30-40% of total budget
- Infrastructure and deployment: 10-15% of total budget
- Training and change management: 5-10% of total budget
Year 2-3 (per year):
- Operational infrastructure: 2,000-15,000 euros/month
- Monitoring and maintenance: 15-25% of initial development cost
- Periodic retraining: 5-15% of initial development cost
- Internal support and management: 0.5-1 FTE equivalent
Concrete example: a document intelligence system for a mid-size company.
| Item | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Development and integration | €120,000 | — | — |
| Data preparation | €40,000 | — | — |
| Infrastructure | €36,000 | €48,000 | €48,000 |
| Maintenance and retraining | — | €25,000 | €25,000 |
| Internal personnel (partial) | €20,000 | €20,000 | €20,000 |
| Total | €216,000 | €93,000 | €93,000 |
| Cumulative TCO | €216,000 | €309,000 | €402,000 |
The initial quote said 120,000 euros. The 3-year TCO is 402,000 euros. Three and a half times more. It's not a scam — it's the reality of any enterprise software system. But if you don't know it upfront, you can't calculate ROI.
Questions You Must Ask Your Vendor
Before signing any contract, GRAL recommends getting written answers to these questions:
What is the estimated monthly infrastructure cost in production? Not in testing — in production, with real volumes.
What does maintenance include and what doesn't it? Security updates, retraining, performance optimization — what's included in the fee and what's extra?
How do costs scale with volume? If you process 1,000 documents per month today and 10,000 tomorrow, does cost 10x? Double? Stay the same?
What happens if we want to switch vendors? Is the data yours? Is the model transferable? Or are you locked in?
What is the cost of downtime? If the system goes down, what does it cost in terms of SLA, penalties, operational impact?
Who pays if the model underperforms? Are there performance guarantees? Exit clauses if metrics aren't met?
ROI Must Justify TCO, Not the Quote
The most common mistake is calculating ROI by comparing expected benefit with initial development cost. The correct calculation compares expected benefit with TCO.
If your AI system saves 150,000 euros per year in operational costs, and TCO is 130,000 euros per year (after the first year), ROI is positive but thin. If the benefit is 300,000 euros per year, ROI is solid.
Do the math with real numbers, not the vendor's optimistic estimates. And do it before starting, not after.
GRAL always presents the complete TCO to its counterparts. Not because it's common practice — it isn't — but because a project built on realistic expectations has far better chances of success than one built on an attractive quote.