Most Companies Are Implementing AI Wrong—Here's the Framework That Fixes That

Why most AI implementations fail before they start

There is a version of AI adoption that looks productive from the outside: a company subscribes to a handful of AI tools, runs some workshops, and declares itself AI-enabled. Usage metrics go up. Productivity metrics do not move.

The reason is almost always the same. The implementation started with the tool rather than with the problem.

Gunnar Brune, who we spoke to on a recent Tech Effect episode, has spent more than two decades in business strategy and marketing consulting, most recently focused on how organizations actually create value with AI rather than how they appear to adopt it. His book, AI Today, was published in Germany the same week ChatGPT launched—a timing coincidence that briefly made him a fixture on German television and, more importantly, confirmed his view that most people's mental model of AI was already too narrow by the time they were paying attention to it.

His core argument is simple: the reason most AI implementations underdeliver is that they are driven by tool availability rather than strategic diagnosis. Companies see what the tool can do and ask how to use it. The companies that build lasting capability ask a different question first—where does value actually get destroyed or created in our business, and can AI address that?

The ChatGPT problem: Why the public mental model of AI is limiting companies

The 30th of November 2022—when ChatGPT launched publicly—created a new challenge for anyone trying to have a serious business conversation about artificial intelligence. Overnight, "AI" became synonymous in most people's minds with a chat interface that answers questions.

Brune is direct about why this matters:

"My experience, especially since the 30th of November 2022, is that everybody—a lot of people who have not been familiar with artificial intelligence before—take ChatGPT as the core of AI and large language models as the AI. But they're so much more than that."

This is not a minor semantic point. When a leadership team believes AI means a chatbot, their AI strategy becomes a chatbot strategy. The much larger space of AI—predictive analytics running in supply chains, computer vision on production lines, machine learning models embedded in the phones and devices already in use, digital twins predicting the behavior of physical systems—remains invisible and therefore unexplored.

The practical consequence shows up when companies assess their AI readiness. If the question is "should we give our team access to ChatGPT?", the answer is easy and the impact is limited. If the question is "where in our value chain are there inefficiencies, bottlenecks, or decisions that AI could improve?"—that question is harder, requires actual business analysis, and tends to identify significantly higher-value opportunities.

The data problem: Why digitization comes before AI

Ask Brune how companies should assess their readiness for AI and his first question is not about which tools they have considered. It is about whether they have data at all.

"The first thing would be: do they track data? How many sensors, how much data do they actually collect? Is it the right data? Can they combine the silos to a network? Can they train on it? Do they understand what they can do with AI and what they can't?"

This sequence matters because AI does not generate insight from nothing. Every model, whether it is predicting equipment failure, optimizing a supply chain, or personalizing a customer experience, learns from historical data that represents the business accurately. A company whose operations are not yet digitized—whose decisions are made on intuition and spreadsheets rather than on structured, connected data—has nothing to train on and nothing to validate against.

The more dangerous version of this problem is not the company that knows it lacks data. It is the company that has data but data that does not represent the business accurately. When a model is trained on incomplete or unrepresentative data, it produces confident predictions that reflect the gaps in the training set rather than the reality of the business. Decisions made on that basis will be wrong in ways that are difficult to detect because the AI produces its errors with the same apparent certainty as its correct answers.

Brune's 3x3 strategy framework, developed across his consulting practice and documented in his book, addresses this directly. It covers three dimensions: understanding of AI, the people and business model of the company, and what needs to be done to be ready. Data maturity sits at the foundation of all three, not because data is the most interesting part, but because without it, every other AI investment is premature.

Finding the bottleneck: The right starting point for an AI strategy

Once data foundations are in place, Brune's diagnostic framework shifts to a specific set of questions that he runs in client workshops. They are deceptively simple:

What are the bottlenecks—the problems that have resisted solutions for years? What are the innovations that were always wanted but never achievable with available resources? What are the business model changes that could not previously be imagined?

The logic behind this structure is that durable AI value comes from addressing problems that genuinely could not be solved before, not from automating tasks that were already being handled adequately. The low-hanging fruit argument—start with quick wins—has a legitimate place in building organizational confidence, but it should not drive the strategic agenda.

"We enter a time where AI gives us so many opportunities. Maybe do it with a strategy and not just pick the next low-hanging fruit."

The banana ripening example from Brune's work in food supply chains illustrates what this looks like in practice. Ripening is a multi-variable control problem—temperature, ethylene levels, carbon dioxide concentration, timing—that has historically required rare specialist expertise to manage well. As grocery supply chains have become more complex and demand more variable, the gap between what human ripening masters can handle and what the system requires has widened. A company called SoftDrive in Germany developed an AI-driven ripening system that manages this complexity more precisely than human operators alone can, reducing waste, improving taste consistency, and extending shelf stability. The AI did not replace the experts, it extended their capability into environments too complex for manual control.

This is the structure of a well-targeted AI application: 

• A problem that resisted solution because of complexity, variability, or the scarcity of the required expertise. 
• An AI system that addresses exactly that constraint.
• Measurable outcomes in cost, quality, or sustainability that justify the investment.

The hybrid team challenge nobody is preparing for

The conversation about AI and work tends to polarize between two positions: AI will take jobs, or AI will free people to do more meaningful work. Brune's view is more specific, and more demanding, than either.

He describes what actually happens when a team's routine work is absorbed by AI: the humans in that team are left doing only the work that AI cannot handle. That work is, by definition, the most complex, most ambiguous, and most cognitively demanding. The team's workflow becomes a series of edge cases and exceptions, with occasional system failures requiring rapid human intervention at full cognitive capacity.

"Our work is to react when AI goes down. So suddenly we are in this incredibly intense situation where we as humans have to do all the work that AI did before but we cannot do. So we are kind of like a fire brigade. We have these phases where we are not needed at all, where we are at the brink of being 'how long will I do this job where I'm obviously not needed'. And then we come into a situation where I'm very much needed but being overwhelmed about what's happening."

This is not a hypothetical. Brune describes a recent training session where Cloudflare went down, taking multiple AI tools offline simultaneously. The participants, professionals who had integrated AI into their daily workflows, were immediately exposed. Some tools failed completely. Others slowed under load as free and lower-tier accounts were deprioritized. The room's dependency on AI infrastructure became instantly visible.

The leadership challenge this creates is real. Teams that are performing well under normal conditions, because AI is handling the routine, may be significantly underprepared for the periods when AI fails or produces unreliable output. Building that resilience back in, while keeping the productivity benefits of AI integration, requires deliberate workforce planning that most organizations have not yet done.

The shadow AI problem

There is a related issue that Brune raises which most organizational AI conversations avoid: employees are not waiting for company-approved AI tools. They are already using the tools they trust from their personal lives—and applying them to company data.

"I think at the moment we don't have to tell people to trust AI. They trust AI far more than companies would like them to trust AI, because what people use in their private lives in their little risk environments can be disastrous for a company if they apply this usage to the company."

The result is a structural gap. Employees work with familiar, capable tools from their personal accounts, while company-sanctioned tools lag behind. The company does not know what data is being processed where, what model providers are storing inputs, or what compliance obligations are being inadvertently triggered.

This is not a problem that can be solved with a policy memo. It requires a combination of providing genuinely competitive tools within a governed environment, and explaining clearly—in terms employees find credible—why the distinction between personal and professional AI use matters.

AI in marketing: From synthetic audiences to talking with data

Brune's professional origin is marketing, and his view of how AI changes that discipline is more specific than the usual commentary about content generation.

The most significant shift he identifies is not automation of production, it is the ability to interact with data through natural language. For most of marketing's history, understanding a target group meant commissioning research, running focus groups, analyzing surveys, and extracting patterns from large data files. The insight was in the data, and accessing it required significant analytical effort.

"Now we do not have to read an Excel file with 100 columns to understand a target group. We can feed this file into an LLM, create a synthetic target group based on this data, and start chatting with them."

This compresses a research cycle that previously took weeks into something that can happen in hours. The strategic insight is not just speed, it is accessibility. Marketers who were previously dependent on data analysts to interpret research can now interrogate their own data directly, in natural language, and iterate on hypotheses in real time.

The authenticity question—whether AI will eventually replace the emotional and creative dimensions of brand storytelling—Brune answers carefully. Current AI systems replicate rather than understand. They produce outputs that can be indistinguishable from human creativity without any underlying comprehension of what they are producing. Whether that distinction will eventually matter less, as simulation becomes indistinguishable from the real thing, is, in his view, genuinely philosophical rather than technical.

What he is confident about is the practical answer for now. Specifically, the most valuable marketing work is produced in hybrid teams where AI handles research, synthesis, and production efficiency, and humans bring the creative judgment that decides what to say, how to say it differently, and why any of it should matter to the audience.

Digital twins and the future of physical infrastructure

The same logic that applies to supply chains and marketing teams applies at a much larger scale to physical infrastructure. Brune has written a forthcoming book on AI in construction and building management, and the digital twin concept runs through that work.

A digital twin is a computational representation of a physical reality—a building, a traffic system, an energy grid—that is precise enough to allow predictions, simulations, and optimization that would be impossible to run on the physical system itself. AI is what makes high-fidelity twins viable. The complexity of real-world environments exceeds what deterministic models can capture, and AI-based prediction fills the gap.

In Hamburg, where Brune has direct familiarity with the projects, the city is piloting digital twin technology on major infrastructure decisions, including the relocation of a train station, a project that involves rerouting subway lines, managing construction impact, and reconfiguring the surrounding urban space. The digital tools allow citizens to understand the complexity of the changes, model different scenarios, and contribute meaningfully to decisions that would previously have been too technical for effective public participation.

The practical implication for organizations managing physical assets is that AI-driven twins are not a future capability, they are already being deployed in demanding commercial environments. The adoption question is no longer whether the technology works, it is whether the organization has the data, the integration capability, and the change management process to extract value from it.

What the next five years actually look like

Brune's view of the near-term AI landscape is deliberately unspectacular, and more useful for that.

The years of rapid tool proliferation are giving way to a period of implementation. The tools are largely ready. The infrastructure is largely in place. What has not yet happened at scale is the organizational transformation: businesses actually changing their processes, their team structures, and their decision-making frameworks to reflect AI's capabilities rather than simply adding AI tools to workflows that remain unchanged.

"We've reached somewhat of a plateau and now it's time to apply these to have the time to really change our process. That'll be the thing to watch out for for the next five years."

For companies that have been in the experimental phase—testing tools, running pilots, building internal familiarity—the strategic question is how to move from experimentation to transformation. Brune's answer is the same diagnostic framework he applies at the start of any engagement: start with the bottlenecks, the unrealized innovations, and the business model assumptions that are now challengeable. The tools exist to address all of them. The work is deciding which ones to address first and building the organizational capability to actually do it.

Essentially, the companies that create durable value from AI are the ones that diagnose before they prescribe, identifying the specific bottlenecks, data gaps, and strategic opportunities where AI changes what is possible, and building the team culture and governance to operate safely in a world where AI is infrastructure, not a tool.

→ Listen to the full conversation with Gunnar Brune on the Tech Effect podcast

Key takeaways

AI is much larger than generative AI—and companies whose strategy is limited to ChatGPT use cases are missing most of the value. Predictive analytics, computer vision, edge AI in physical systems, and machine learning in operations represent mature, proven applications that predate and extend well beyond large language models.

Digitization is a prerequisite for AI, not a parallel track. AI cannot compensate for data that does not exist, is siloed, or does not represent the business accurately. Companies that have not digitized their operations will fall behind both digitization and AI simultaneously.

The right starting point for an AI strategy is the bottleneck, not the tool. Ask what problems have resisted solution for years, what innovations were always wanted but never achievable, and what business model assumptions could now be challenged. That analysis drives durable value; tool-first adoption drives short-term activity.

Hybrid human-AI teams require active management of what happens when AI fails. When routine work is absorbed by AI, humans are left with only the complex and ambiguous—which creates a cognitive load pattern that alternates between underutilization and crisis. Leaders need to plan for this explicitly, not discover it reactively.

Shadow AI is already inside most organizations. Employees trust and use personal AI tools on company data without formal approval. The gap between personal AI capability and company-sanctioned tools is a real governance and compliance risk that policy memos alone will not close.

Teams who talk with their data create insight that others cannot access. The ability to feed a customer dataset into an LLM and have a natural language conversation with a synthetic representation of the target group compresses research cycles from weeks to hours—and makes strategic insight accessible to people who were previously dependent on data analysts to extract it.