Ask three vendors what an AI initiative costs and you will get three answers that differ by a factor of five. Ask an internal team and you will get a number that covers the obvious parts and misses the rest. The result is a budget that is either so padded it never gets approved, or so lean it runs out mid-project — at which point the half-built system quietly joins the majority that never reach production.
That majority is the real backdrop to any budget conversation. McKinsey's State of AI survey found that while 88% of organisations now use AI in at least one function, only 39% can attribute any EBIT impact to it — and for most of those, the impact is under 5%. The money is being spent. The value, mostly, is not landing. The single strongest correlation McKinsey found with actual EBIT impact was not model choice or vendor — it was fundamental workflow redesign, which only around a fifth of adopters had attempted. Budgeting for an AI initiative is therefore not about pricing software. It is about pricing the change.
What follows is a structural breakdown of where the money actually goes, and which cost categories organisations consistently underestimate. The euro figures are illustrative planning bands for a DACH Mittelstand context — single-site or small-multi-site, not hyperscaler scale — not a quote and not survey data. Use them to interrogate a proposal, not to sign one.
The cost anatomy
Every AI initiative has six cost layers. Most budget proposals cover two of them.
Layer 1: Discovery and scoping
Before any engineering begins, you need to establish workflow readiness, validate that the data you need is actually accessible, and define what success measurably looks like. This is the work that determines whether the initiative should proceed at all — and at what shape.
Skipping discovery is the most expensive mistake in AI deployment, precisely because the failure mode is invisible until late. A discovery engagement that kills a doomed initiative early saves the entire build budget. One that sharpens scope and de-risks the integration pays for itself in the first sprint. The organisations stuck in what the industry now calls pilot purgatory are, overwhelmingly, the ones that went straight to building something impressive before they had established whether it could connect to anything or whether anyone would use it.
As a planning band, expect roughly €8,000–25,000 for a serious discovery on a single workflow. Spending nothing here does not save that money — it defers it into the build, where scope, data, and integration surprises are discovered at the most expensive possible moment.
Layer 2: Engineering and build
This is the cost everyone budgets for. It covers model selection and configuration, prompt engineering, workflow orchestration, API development, and testing — call it €20,000–120,000 depending on complexity.
The range is wide because of one variable, and it is not the one most people assume. A single-workflow deployment with clean data and modern APIs sits at the low end. The same workflow with legacy integration, custom data pipelines, and gnarly business logic sits at the high end. A multi-workflow deployment with cross-system orchestration sits above it.
The key driver is not AI complexity — it is integration complexity. The model configuration for most Mittelstand use cases is genuinely straightforward; the frontier work has already been done by the model providers. The engineering effort goes into getting data out of legacy systems, handling the edge cases your business logic actually contains, and building pipelines that do not break at 2am. This is why data accessibility is the strongest predictor of engineering cost, and why two superficially identical projects can differ in price by a factor of three.
Layer 3: Infrastructure
Cloud and model-API infrastructure for AI workloads is cheaper than most organisations expect — typically €200–3,000 per month for a Level 1 deployment, covering compute, storage, model-provider API calls, monitoring, and logging.
Crucially, the cost scales with volume, not complexity. A workflow processing a thousand items a week costs a manageable multiple of one processing a hundred — and the per-item cost at this scale is trivial next to the human time it displaces. The arithmetic that surprises people runs the other way: the marginal cost of doing more is low, which is exactly why the value case depends on volume and adoption rather than on squeezing the infrastructure line.
Where infrastructure does bite is during the build, in data staging and transfer from legacy systems. If the pipeline requires heavy transformation or a large historical back-load, expect a temporary spike that has nothing to do with steady-state running cost. Budget it separately so it does not contaminate your ongoing-cost expectations.
Layer 4: Integration and data access
This is the layer that blows budgets. Integration is the bridge between the AI workflow and the systems it must read from and write to. For an organisation with a modern, API-first architecture, it is minimal and folds into the engineering layer. For an organisation running the kind of long-lived ERP and bespoke databases typical of the Mittelstand, it becomes a project within the project — somewhere between €5,000 and €40,000 on its own, depending on what you are bridging.
A thin API wrapper around an existing database view sits at the low end. A batch pipeline out of a legacy ERP sits in the middle. A real-time event pipeline from a system that was never designed to stream, or a multi-system integration that requires reconciling data that does not agree across sources, sits at the top.
The decisive insight is that integration cost is knowable before the project starts. A readiness assessment maps the integration landscape in days. The expensive version of this layer is the one discovered mid-build, when the team hits a system that cannot be read the way the architecture assumed — and the fix has to be retrofitted around work already done.
Layer 5: Change management
The cost almost no one budgets for, and the reason many technically successful deployments quietly fail in practice. Change management covers team training, process documentation, redefining the KPIs the team is measured on, and the support required to move people from a manual workflow to an AI-assisted one. As a band, €5,000–20,000 for a single workflow — small against the engineering line, decisive against the outcome.
This is not a soft footnote; it is the layer McKinsey's data points directly at. Workflow redesign — not the model — was the strongest correlate of measurable impact, and the four-in-five organisations that layered AI on top of unchanged processes are largely the same ones reporting no EBIT effect. Without this layer you reproduce the pattern described in operating model clarity: the system works technically, and is unused within weeks, because no one redefined how the team actually operates around it.
Layer 6: Governance and compliance
The newest line on the budget, and the one most legacy cost models omit entirely. The EU AI Act is now in force on a phased timeline, with the obligations for high-risk systems — those listed in Annex III, covering use cases such as recruitment, worker management, credit scoring, and critical-infrastructure operation — among the heavier ones arriving over 2026–2027. For German firms this lands on top of an already-cautious mood: in Bitkom's 2025 survey, legal uncertainty was tied for the single most-cited barrier to AI adoption, named by 53% of companies.
For the typical Mittelstand workflow, the practical cost here is not a six-figure conformity programme — most internal-productivity use cases are not high-risk under Annex III. It is the documentation discipline: knowing how your system is classified, keeping records of data sources and model behaviour, ensuring a human stays meaningfully in the loop, and being able to show your work if asked. Built in from discovery, that discipline is a modest line. Bolted on after a system is live — or after a regulator or a customer's procurement team asks — it is a retrofit, and retrofits are where compliance gets expensive. Treat classification as part of scoping, not as a phase-two surprise.
Total cost by deployment type
Combining the layers into honest planning bands:
Level 1 — single workflow, modern stack: roughly €40,000–80,000 to build, €500–1,500/month to run. Payback typically inside a year when adoption is real.
Level 1 — single workflow, legacy integration: roughly €70,000–150,000 to build, €800–2,000/month to run. The premium is almost entirely integration cost, which is why it is worth knowing before you commit.
Level 2 — multi-workflow, cross-system: roughly €150,000–300,000 to build, €1,500–3,000/month to run. Longer payback, and worth attempting only after a Level 1 deployment has proven the organisation can actually absorb the change. Most failures at this tier are not technical; they are organisations buying scale before they have bought adoption.
Where organisations overspend
Over-engineering the AI layer. The model and prompt work for most Mittelstand use cases is the cheap part. If it is dominating your budget, you are either solving a problem that does not need AI or building custom models where a configured commercial one would do. See build vs. buy for the decision framework.
Under-investing in discovery and redesign. This is the costliest pattern, and the survey data is unambiguous about why: impact tracks with redesigning the work, not with deploying the tool. Money saved by skipping discovery reappears, larger, as mid-build rework and as a finished system nobody adopts.
Treating change management as optional. A heavily engineered deployment with no adoption budget reliably delivers less value than a leaner one with structured training and KPI redefinition behind it. The technology only generates value when people use it — and people use it only when the work around them has been redesigned to assume they will.
How to budget
Start with the diagnostic to assess your readiness across all six dimensions. The dimension scores tell you which cost layers will dominate before you spend anything: strong workflow readiness and data accessibility push you toward the lower end of engineering with minimal integration; weak data accessibility means integration becomes the largest single line; no operating-model clarity means change management is non-negotiable rather than discretionary. Then build the business case using the template structure, with Phase 1 gates that cap the capital at risk until the first workflow has earned the next.
The goal is not to minimise cost. It is to spend knowingly — with a clear, defensible expectation of what each layer delivers, when the investment pays back, and which of the six layers, left unbudgeted, would have quietly sunk the whole thing.
A Fit Call pressure-tests your initiative against this six-layer anatomy — so you find the integration, change-management and compliance costs in a 30-minute conversation, not three months into a build.
References: McKinsey, "The State of AI: How organizations are rewiring to capture value," 2025 (https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai); EU Artificial Intelligence Act, "High-level summary" and Annex III, artificialintelligenceact.eu (https://artificialintelligenceact.eu/high-level-summary/); Bitkom, "Durchbruch bei Künstlicher Intelligenz," September 2025 (https://www.bitkom.org/Presse/Presseinformation/Durchbruch-Kuenstliche-Intelligenz).