Why AI Projects Don't Fail on Technology

When people talk about artificial intelligence (AI) and its use in enterprises, reports range from complete disillusionment to demonstrable success. According to a study published by Gartner in spring, only 39% of IT leaders worldwide expect their company's AI projects to have a positive impact on financial performance in the foreseeable future. The study surveyed 353 Data & Analytics (D&A) and AI leaders between November and December 2025. At the same time, according to Gartner, organizations with successful AI initiatives invest up to four times more in data foundations than those disappointed by their AI initiatives. These foundations include everything than falls under the umbrella term governance – topics such as data quality, access rights, approval and decision-making processes, as well as data accountability and ownership.

In June, Gartner followed up and identified six trends:

Alongside the growing importance of souvereign AI, most of these trends are, to a greater or leasser extent, about governance: decision governance for transparent and traceable outcomes, the rising importance of governance platforms, and – also a topic within the data foundation space – the rise of GraphRags, the combination of knowledge graphs and language models, for accurate AI-generated statements.

Why Governance Is the Foundation for AI Success

The importance of good governance also becomes clear time and again in our client projects. AI PoCs often lead, for the first time, to KPIs being centrally defined and standardized across the organization. AI required this, as it otherwise cannot operate reliably.

Alongside data quality itself, semantic modeling is decisive here. A semantic layer gives the language model the context of the data: what a metric means from a business perspective, how it is calculated, at what level of granularity it exists, how entities relate to one another, and who is allowed to see which data. This exact knowledge isn't contained in the raw data – so far, it has mostly existed only in the heads of business departments.

By contrast, how powerful the chosen language model is turns out not to be decisive. On the contrary: when semantic context is clean and well-defined, even smaller and significantly cheaper models deliver reliable results. Without a solid data and semantic foundation, on the other hand, even the strongest model simply cannot understand the data, doesn't know which data may be used by whom, and starts to hallucinate.

But how do companies arrived at good governance, where effort and benefit are well balanced? How much governance is needed for good AI outcomes, and which topics matter how much? Here, we want to present a framework that we successfully apply in our client projects.

The path to Good AI Governance in Four Steps

It Starts with a Concrete Use Case

Governance doesn't need to be reinvented for every new AI use case. What matters is building a structure that can be reused. Starting with one very concrete use case has proven effective.

The framework follows M2' four consulting stages – AI Clarity, AI Build, AI Operate, AI Control – because, in exactly this order, they answer the four questions Gartner identified as the cause of failed projects: Who is allowed access, who is accountable for quality, who grants approval, and how does the decision behind it become traceable?

Phase 1 – AI Clarity: Naming Responsibilities

The goal of this phase is not a governance concept on paper, but a list of names. For every AI use case that is in production or being planned, three roles are documented: who grants access to which data, who is accountable for the quality of the outputs, and who gives final approval. All three must be name individuals or functions – not departments.

Completion criterion: No active or planned use case without these three names. For a single use case, this can be completed in one to two weeks.

Phase 2 – AI Build: Embedding Governance in the Architecture

What was documented in Phase 1 is now technically enforced. Access rights are mapped as a role model within the system, and approval processes are implemented as workflows with fixed criteria and escalation paths – this is what Gartner means by decision governance: decisions that are explicitly modeled in advance so they no longer need to be renegotiated case by case. This allows quality checks to happen before the model responds.

Completion criterion: A system may not be released if an access rule, approval workflow, or quality check is missing. Governance is thereby firmly anchored in the process all the way to go-live.

Phase 3 – AI Operate: Verifying Governance in Live Operations

Governance that existed only at rollout quietly erodes during operation. This phase continuously checks whether approval processes are actually being followed or bypassed, and wheter the responsibilities named in Phase 1 still match actual usage. New real-time data flows – Gartner also names the growing need for agentic data streaming among its trends – are only introduced where the controls from Phase 2 are already in place. Otherwise, speed grows faster than control, which is exactly what Gartner is currently observing in practice.

Completion criterion: Deviations – such as unreviewed or AI-hallucinated content reaching stakeholders – are caught before they cause damage, not after.

Phase 4 – AI Control: Making Governance Reusable

In this phase, the case-by-case solution becomes a standard. New use cases go through the same criteria from Phases 1 to 3 without governance being discussed from scratch each time. Every agent decision remains traceable – Gartner puts a figure on the effect of explicitly modeled decisions: five times higher trust and an 80% shorter decision time compared to unmanaged decisions.

Completion criterion: A new AI use case can be classified using the existing toolkit instead of triggering a new governance discussion.

What the Framework Doesn't Replace

It doesn't replace a decision that a business department must make itself – who ultimately grants approval, for example, remains an organizational question, not a technical one. And it doesn't work retroactively: buying a model first and retrofitting governance afterward is exactly the pattern described in this article, and it's what delays projects.

M2 as Architect and Builder of Governance

The four phases described here don't come from theory. We build them, again and again, in client projects. If you're currently working on a concrete use case and aren't sure where your organization stands, we'll review the starting position together with you and point out which steps make sense next.

^Sources

^{Gartner: Organizations with Successful AI Initiatives Invest Up to Four Times More in Data and Analytics Foundations, 16.04.2026 — gartner.com/en/newsroom/press-releases (abgerufen: 2026-07-20).} 

^{Gartner Newsroom: Gartner Identifies the Top Trends for Data and Analytics, 16.06.2026 — gartner.com/en/newsroom/press-releases (abgerufen: 2026-07-20).} 

^{BARC-Magazin: Die Formel für AI-Ready Data, 19.06.2026, frei zugänglich (abgerufen: 2026-07-20).} 

^{dbt Labs: State of Analytics Engineering Report 2026, 14.04.2026 — getdbt.com/resources/state-of-analytics-engineering-2026 (abgerufen: 2026-08-14; Herstelleraussage, deckt sich inhaltlich mit unabhängigen Gartner-/BARC-Zahlen).} 

Contact us