Most AI budget requests fail before they reach a decision. Not because the initiative lacks merit, but because the proposal reads like a technology pitch instead of a business case. The Geschäftsführung does not evaluate AI on technical sophistication. It evaluates it on three things: how much, how long, and what happens if it fails. Answer those three questions with arithmetic and you have a business case. Dodge them with adjectives and you have a slide deck that stalls.
The macro evidence has caught up with this instinct. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls. That is not a technology problem. Those are exactly the three failure modes a disciplined business case is built to close off before a euro is committed. After structuring business cases across DACH engagements, we have seen the same pattern hold: approval is an exercise in evaluability, not enthusiasm.
Why most AI proposals fail
The typical AI proposal arrives as a slide deck with a use-case description, a reference to "efficiency gains," a vague timeline, and a budget range so wide it communicates uncertainty rather than planning. Stripped of its diagrams, it reads as a single sentence: we think this could work, and we need money to find out. That framing triggers every risk instinct a Geschäftsführer has, because it sounds like R&D wearing the costume of a business investment — and R&D is precisely what gets cut first when the quarter tightens.
The proposals that get approved do something structurally different. They frame AI as a process improvement with known inputs, measurable outputs, and a defined payback period. They answer the question the executive is actually asking — what do I get for this, and when — rather than the question the engineer finds interesting, which is how the thing works. The shift is not cosmetic. It moves the proposal from the "experiment" mental category, where the default answer is no, into the "operational improvement" category, where the default answer is show me the numbers.
The structure that works
A business case that survives executive review has five sections. Not ten, which signals that nobody has decided what matters. Not three, which signals that the hard parts have been skipped. Five.
The operational problem in numbers. This is not "our customer service could be better." It is a baseline you can defend: the support team handles a known volume of tickets per week, the average first-response time is a measured figure, a definable share of those tickets follow predictable resolution patterns, and the current cost per ticket is calculable from headcount and volume. If you cannot populate those numbers from systems you already run, you are not ready to write a business case — you need discovery first. The absence of a baseline is itself the finding, and it is cheaper to learn it here than at the go/no-go gate.
The proposed intervention. State precisely what the system will do. Not "enhance customer service with AI" but something a non-technical board member can picture: classify incoming tickets by urgency, route them to the correct queue, and draft initial responses for the predictable share. The specificity you can muster here is downstream of how clearly your workflows are mapped, which is why workflow readiness is the dimension that most often determines whether this section reads as a plan or a wish.
The cost model. Mittelstand AI initiatives sit at a different order of magnitude than the hyperscaler case studies that dominate the trade press, and pretending otherwise destroys credibility instantly. A focused single-workflow deployment is a five-figure to low-six-figure commitment; legacy integration work and a second workflow push it higher. Your business case should name the tier, with line items for engineering, infrastructure, and — the one most proposals omit — change management, because the model is rarely what fails. Disclose every cost upfront. Hidden costs surfaced later erode trust faster than high costs disclosed honestly.
The payback calculation. This is where most proposals collapse into speculation, and it is the section the Geschäftsführung reads first. The discipline is simple and unforgiving: calculate payback using only the operational metrics from the first section. If the system removes a measured number of hours of handling time per week at a known fully-loaded cost per hour, the weekly saving is arithmetic, not narrative. No "potential revenue uplift," no strategic optionality, no second-order benefits you cannot defend. Just measured baseline to projected output. The full metric framework lives in measuring AI ROI; the rule here is that every number in this section must trace back to one in the first.
The risk mitigation plan. What happens if it does not work? The answer must never be "we lose the entire investment." It should be a staged commitment: Phase 1 funds a working prototype on real data within a defined window, with a hard go/no-go decision at the end and a capped amount at risk; Phase 2 proceeds only if Phase 1 hits its threshold. Staged funding with explicit kill points is the single most effective pattern for executive approval, because it converts an irreversible bet into a sequence of small, reversible ones — and it directly neutralises the "escalating costs, unclear business value" failure mode that Gartner flags as the reason most projects die.
The numbers executives actually want
Three financial figures close the conversation, and they are not the ones finance textbooks lead with. Payback period comes first — not IRR, not NPV, but the number of months until operational savings recover the investment. For a focused first deployment, anything inside a year reads as serious; a horizon past eighteen months signals that the scope is wrong and the case needs rework before it goes in front of anyone.
Cost per unit improvement comes next, because the Geschäftsführung thinks in operational units rather than aggregates: cost per ticket, per invoice, per claim. Show the current cost, the projected cost after deployment, and the delta. That single line does more to make "efficiency gains" concrete than any slide of benefits ever will.
Risk-adjusted investment closes it. Frame the decision not as the total project cost but as the maximum capital exposed before the first kill point. A six-figure programme gated by a five-figure Phase 1 is, from the board's seat, a five-figure decision — and framing it that way is the difference between a quick yes and a deferred maybe.
Common mistakes
Over-scoping is the most expensive. The business case tries to justify an enterprise-wide AI strategy when it should fund one workflow that proves the model. Start narrow; the second case is far easier to approve once the first has produced numbers.
Technical language quietly reclassifies the document. The moment a proposal leans on fine-tuning, retrieval architectures, or vector databases, it has become a technical artefact addressed to the wrong audience. Translate everything into process and money. The board does not buy mechanisms; it buys outcomes.
Missing the comparison is the omission that loses on the merits. Every AI business case competes against doing nothing, and doing nothing is never free. Calculate the cost of inaction explicitly — the ongoing operational drag, the compounding inefficiency, the competitive exposure — as set out in the cost of AI inaction. A case that does not name its alternative invites the board to supply the cheapest one.
No sponsor alignment kills otherwise sound proposals in the review cycle. If no named executive has committed to champion the case, it has no decision authority behind it and will be talked to death. Secure the sponsor before you write the document, not after.
A word on the regulatory column that increasingly belongs in the risk section. The EU AI Act's transparency duties under Article 50 — the obligation to tell people they are interacting with an AI system, and to mark AI-generated content — begin applying on 2 August 2026, and they catch far more ordinary use cases than most boards assume, including customer-facing chat and generated communications. The heavier obligations for high-risk systems — the tier that covers use cases touching hiring, creditworthiness, and similar functions — were originally tied to the same August 2026 date, but the Digital Omnibus deal reached on 7 May 2026 postponed them: standalone Annex III high-risk systems now have until 2 December 2027, and AI embedded in regulated products until 2 August 2028. The relief is real, but it is not an excuse for silence. A business case that names where its use case sits on that risk tier — and budgets for the documentation, disclosure, and oversight each tier implies, on the timeline that actually applies — reads as governed rather than naïve, and governed is what survives committee.
The approval pattern
In the Mittelstand the winning pattern is remarkably consistent. A named Geschäftsführer or Bereichsleiter champions a single-workflow initiative with a modest, capped Phase 1 budget and a clear decision gate a couple of months out. The case fits on three pages. The payback calculation runs entirely on operational data the company already holds. The risk framing is honest about what could go wrong and what it would cost.
That is not a sophisticated framework. It is a clear proposal from someone who understands that executives approve investments they can evaluate, not technologies they are asked to trust. Build the case around the process, not the technology. The technology is a means; the process improvement is the investment — and the investment is what gets funded.
A Fit Call pressure-tests your AI business case against the five sections above — before it goes in front of the Geschäftsführung and gets sent back for rework.
References: Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," 25 June 2025 (https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027); EU Artificial Intelligence Act, Article 50 — Transparency Obligations (https://artificialintelligenceact.eu/article/50/); EU Artificial Intelligence Act, Implementation Timeline (https://artificialintelligenceact.eu/implementation-timeline/); Gibson Dunn, "EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes" (https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/).