10 min read<\/p>\n
One third of DACH companies with production\u2011grade AI projects report cost overruns compared with the original business case. The Bitkom 2026 study cites 33\u202fpercent. For management this is not a procurement quibble but a governance issue. Whoever does not establish a solid AI investment model by Q4\u202f2026 will have the topic slammed in the investment committee \u2013 or flagged by the auditor in the audit\u2011report discussion.<\/strong><\/p>\n The Bitkom 2026 study was released in April and provides the first reliable data set on productive AI use in German mid\u2011size firms and large corporations. 41\u202fpercent of surveyed companies say they have AI in production. Of those users, 33\u202fpercent report cost overruns against the original business case. 19\u202fpercent say they have cut jobs and attribute it to AI efficiencies.<\/p>\n Three figures, three narratives, one common thread: AI has reached the stage of an honest balance sheet. The hype curve has run its course, and reality checks are happening in corporate accounting. What started 18\u202fmonths ago as an innovation initiative now sits in the investment committee and must prove its value contribution.<\/p>\n What is AI investment governance?<\/strong> AI investment governance is the separate leadership and reporting logic for initiatives with model\u2011 and data\u2011dependent value contribution. It differs from classic IT investment governance through stage\u2011gate approvals, value\u2011contribution allocation between IT and business units, and a dedicated risk\u2011reporting line to the audit committee. Without this separation, two investment classes with different scaling dynamics get mixed in a single report.<\/p>\n The 33\u202fpercent are not a procurement phenomenon. They are the symptom of a strategy gap. Treating the cost overrun as a supplier problem is tightening the wrong screw.<\/p>\n Traditional IT investments follow a simple logic: requirement, budget, contract, implementation, acceptance. AI investments break that logic. The value contribution is created not during implementation but in ongoing operation \u2013 and it shifts with every model change, every data enrichment, every process adjustment in the business unit.<\/p>\n Anyone who runs an AI initiative like a licensing project ends up with a licensing\u2011project forecast. It lasts 14 months. Then the first real scaling step arrives. The model meets reality: more tokens than planned, new pipelines, additional business units, external data sources with their own licensing costs. The forecast was never wrong \u2013 it was built for the wrong investment class.<\/p>\n From a C\u2011level perspective this means: AI belongs in its own investment class with its own governance logic. Not as an IT cost line item, but as a scaling\u2011sensitive growth investment. Mixing the two logics in one report dilutes the insight.<\/p>\n Industry observers from the CIO community report consistently: the first cost\u2011overrun discussions with the CFO are not about model prices but about who books which value contribution. When sales uses an AI\u2011powered lead score and reports an 8\u202fpercent higher conversion rate, the effect flows into the sales P&L. The costs, however, sit with IT. Without value\u2011allocation, the investment committee sees only the cost side.<\/p>\n The next four quarters will decide whether AI adoption in DACH companies continues in a controlled manner or faces a full brake in the supervisory board. Four phases, four decisions.<\/p>\n Anyone who wants to have a serious cost\u2011overrun discussion needs three levers \u2013 and the willingness to turn any of them against your favourite initiative when necessary.<\/p>\n Lever\u202f1: Value\u2011contribution logic before tool selection.<\/strong> The question isn\u2019t \u201cwhich model\u201d or \u201cwhich vendor,\u201d but \u201cwhich decision should be improved and how much that improvement is worth per year.\u201d Starting with the tool choice means you\u2019ve already lost the proper sequence.<\/p>\n Lever\u202f2: Stage\u2011gates instead of an annual approval.<\/strong> AI initiatives need phased approvals with a right to abort. If you can\u2019t plausibly quantify the value contribution after 90\u202fdays, you won\u2019t receive a second tranche. That prevents the costly 14\u2011month cost overruns that are now becoming visible.<\/p>\n Lever\u202f3: Risk reporting in the audit committee.<\/strong> AI belongs in the same reporting stream as ERP risks or cloud risks \u2013 a standing agenda item in the quarterly review. Three indicators are enough: run\u2011costs versus plan, value contribution versus plan, supplier concentration. It\u2019s lean but audit\u2011ready. Anyone looking for a reference framework for strategic anchoring will find it in the Merck decision story on Agentic AI<\/a>.<\/p>\n \u201cHonest AI discussions now happen between the CFO and the CIO, not between marketing and the innovation lead. Walk in without a clean value\u2011contribution logic and you\u2019ll lose the next investment round \u2013 no matter how well the pilot performed.\u201d<\/p>\n Tone from the CIO round\u2011table in the DACH mid\u2011market, spring\u202f2026<\/p>\n<\/div>\n Today\u2019s cost overruns are tomorrow\u2019s path dependence. Scaling AI initiatives in the first 18\u202fmonths without a clear architecture decision creates lock\u2011ins that become expensive later. Three of them are so common they barely register any more.<\/p>\n First: data\u2011stack dependence. Models are interchangeable, training and context data are not. Building the training\u2011data workflow in a hyperscaler tool without cleanly specifying data extraction means you won\u2019t switch later. That\u2019s not a tech question; it\u2019s an investment question.<\/p>\n Second: supplier concentration. Hyperscaler stacks look like smart consolidation in the first 12\u202fmonths. By month\u202f24 they become a single point of failure for your negotiating position. Several board members in the DACH region learned this during cloud migrations \u2013 the AI generation is repeating the mistake right now. Those who want to avoid it systematically can find the architecture\u2011decision logic in the analysis of three observations from the Constellation Enterprise Intelligence Report<\/a>.<\/p>\n Third: skill concentration. When AI expertise sits in a central team that speaks only one platform\u2019s toolchain, the switching\u2011cost corridor isn\u2019t a licensing issue. It\u2019s a people issue. This path dependence arrives with a time lag \u2013 but it arrives.<\/p>\n Ignore all of this and you\u2019ll face a strategy problem on the table in\u202f2027, not just a cost\u2011overrun problem. From a supervisory board perspective, the latter is uncomfortable; the former is liability\u2011relevant.<\/p>\n The dividing line is not speed. Both columns describe companies that take AI seriously. The dividing line is the question of whether steering is ahead of scaling or lagging behind. Those who lag will face a tough discussion in 2027 \u2013 with the auditor, the supervisory board or the investor, depending on ownership structure.<\/p>\n The CIO provides the inventory: which AI initiatives are actually running, what their operating costs are, and against which approved plan. Without this list, every further discussion is speculation.<\/p>\n The CFO brings the governance model: AI as its own investment class, stage\u2011gates, value\u2011contribution allocation. That is his original playing field \u2013 and the point at which most mid\u2011size corporate groups need to rethink the architecture of their reporting.<\/p>\n The CEO raises the strategic question: Where do we want to be in 2028? Which AI investments will get us there? Without this anchoring, every stage\u2011gate debate becomes a tactical squabble. With anchoring, it becomes strategic steering.<\/p>\n And the chairman of the supervisory board asks the honest question: Can we defend the investment approvals of the past 18 months to the investor? If the answer is \u201cI don\u2019t know,\u201d the topic will be on the agenda by the spring meeting of 2027 at the latest. Better to have it there in a controlled way beforehand.<\/p>\n The Bitkom figures are a snapshot. Upcoming surveys will show whether 2026 was the year DACH companies matured their AI governance \u2013 or the year the first CEOs were dismissed because of uncontrolled AI investments. Both scenarios are plausible. Which one materialises will be decided not by the IT roadmap but by the investment logic.<\/p>\n A complementary CFO perspective on the architecture and procurement side of the same Bitkom finding can be read in the MBF Media network: AI more expensive than planned: What the 33\u2011percent cost\u2011overrun rate means for mid\u2011size CFOs<\/a>. Those who want to deepen the value\u2011contribution logic will find the strategic lines in the analysis AI Governance 2026 at System Level<\/a>. And the data on productive AI use are expanded in the three figures from the Deloitte State of AI April 2026<\/a>.<\/p>\n The AI value contribution is generated during ongoing operation and changes with data, model, and process. Classic IT investments have a fixed scope of delivery. Therefore AI requires its own investment class, stage\u2011gates instead of an annual approval, and a clean value\u2011contribution allocation between IT and the business unit.<\/p>\n<\/details>\n Run costs versus planned value, measured value contribution versus forecast, supplier concentration per application area. Narrow enough for quarterly reporting, audit\u2011ready for the external auditor, and meaningful enough for the supervisory board.<\/p>\n<\/details>\n When strategic options are lost without documented decisions. Supplier lock\u2011in or data\u2011stack dependency without an explicit investment justification is regarded in a dispute as a gap in due\u2011diligence. An early explicit architecture decision protects against this discussion.<\/p>\n<\/details>\n Source cover image: Pexels \/ Vitaly Gariev (px:36713442)<\/p>\n Bitkom 2026: 33 percent of DACH companies are exceeding their AI plan. What C\u2011level executives now need to change in the governance model before the\u2026<\/p>\n","protected":false},"author":114,"featured_media":21315,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_yoast_wpseo_focuskw":"AI Cost Overrun DACH","_yoast_wpseo_title":"AI Cost Overruns as a Governance Issue: What the 33% Figure from Bitkom\u2019s 2026 S","_yoast_wpseo_metadesc":"Discover 33% cost overruns in AI projects. Master C\u2011level governance, stage\u2011gates and risk reporting for DACH boards before the 2027 supervisory meeting.","_yoast_wpseo_opengraph-image":"","_yoast_wpseo_opengraph-image-id":0,"_yoast_wpseo_twitter-image":"","_yoast_wpseo_twitter-image-id":0,"featured_post_sortierung":0,"featured_post":0,"pre_headline":"","bildquelle":"","teasertext":"","language":"de","_evm_slot_owner":"","_evm_translation_lang":"","_wp_old_slug":[],"footnotes":""},"categories":[700,694],"tags":[],"class_list":["post-21489","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ict-infrastructure","category-new-work-leadership","entry"],"wpml_language":"en","wpml_translation_of":21316,"yoast_head":"\nKey Takeaways<\/h2>\n
\n
What the Numbers Really Mean<\/h2>\n
Why this changes the governance logic, not the procurement calendar<\/h2>\n
The twelve months that matter now<\/h2>\n
Three Levers That Carry Weight in the Executive Board Meeting<\/h2>\n
The Path Dependence No One Shows on the Slide<\/h2>\n
What Controlled Steering of Audit\u2011Stop Differs From<\/h2>\n
\n
\n
Who Should Put What on the Table Now<\/h2>\n
What\u2019s Next on the List<\/h2>\n
Frequently Asked Questions<\/h2>\n
How does AI investment governance differ from classic IT investment governance?<\/strong><\/summary>\n
Which three indicators are sufficient for AI risk reporting to the audit committee?<\/strong><\/summary>\n
When does AI path dependency become liability\u2011relevant from a C\u2011level perspective?<\/strong><\/summary>\n
Read more<\/h3>\n
\n