Enterprise AI Is an Organizational Design Problem in Disguise

In August 2025, MIT's NANDA initiative published a number that should have ended a two-year argument. Roughly 95% of enterprise generative AI pilots produced little to no measurable impact on the bottom line. When I first read that, I assumed it meant the tools were not ready yet. Then I went looking at which pilots worked and which ones quietly died, and the pattern sat somewhere I did not expect. The failures had almost nothing to do with weak models and almost everything to do with how the companies were built.

Buying the technology and changing nothing is the most reliable way to get nothing from it

Here is a story that has nothing to do with AI, which is why I hope it helps you.

When factories first electrified in the late 1800s, most of them barely got faster. Owners ripped out the central steam engine, dropped in an electric motor of about the same size, bolted it to the same overhead shafts and leather belts, and waited for a productivity boom that did not come. For roughly forty years electricity underdelivered so badly that economists still call it a paradox. The economic historian Paul David told this story in his 1990 paper "The Dynamo and the Computer," and the resolution is so interesting: the gains only arrived once owners stopped treating the motor as a drop-in swap and rebuilt the factory around it, with small motors on individual machines and a floor laid out by the flow of work rather than the reach of a driveshaft.

The technology had been sitting there for decades and what was missing was the willingness to rearrange the organization around it.

Most enterprise AI today is a steam-era factory with an electric motor bolted on. The model is genuinely capable. The org chart it plugs into was drawn for a slower world, and the model just runs that old design faster.

Which surfaces something that sounds wrong until you say it out loud: pointing AI at a struggling process usually makes things worse, not better. A slow approval chain becomes a slow approval chain producing confident output at every broken step. Bad customer data does not get cleaned, it gets acted on at speed. Capability amplifies whatever it touches, so if that is dysfunction, congratulations, you have automated the dysfunction. (Also read: The Dangerous Myth of 'Just Add AI' to Your Data Warehouse)

The question is shifting from how many people we employ to how many things we manage

There is a shape to how AI shows up inside a company, and it quietly changes what a manager does all day.

The assistant made individuals a little faster and changed nothing structural. The agent reshapes how a single function operates. The coworker, if it arrives the way the trend suggests, reshapes the org chart itself, because a manager is suddenly running a blended team of people and systems and answering for the output of both.

From the top of the article, the 95% is measured. This three-stage arc is not, so treat it as informed extrapolation. The direction feels solid to me, but the timeline is a guess, and it is the piece I would least want to bet the house on.

Nobody owns the real question, so the projects quietly die

Most large companies have a CIO, a CTO, often a Chief Data Officer and a Chief Digital Officer. Very few have anyone whose actual job is the question underneath all of it: how does AI change the way this company is organized and what it rewards? That work has grown too big to survive as someone's side quest.

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, and their own analyst is fairly blunt:

"Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied." -- Anushree Verma, Senior Director Analyst, Gartner (2025)

That is an org-design failure wearing a technology mask. When no one owns governance, standards, and the budget, you get a hundred disconnected experiments, each launched by a department chasing the same hype, none accountable to a system that makes them add up to anything.

The right structure is neither one central AI team nor everyone for themselves

The best structure for AI in a large company is neither centralized nor decentralized, and choosing either pure option is setting you up for failure.

Think about whales. Sales, marketing, finance, HR, operations: each function needs its own AI team that knows its process cold and carries the context nobody ever bothered to write down. Those teams are the whales, large and expert and adapted to their own patch of ocean. A whale on its own does not get far, though. They move in pods that share what they have learned and hold to common behavior.

A company needs both animal and pod at the same time. The function teams (whale) innovate locally while a central AI organization plays the pod, owning shared standards, security, architecture, and the components everybody reuses. Kill the whales and you get generic tools that miss the real work. Kill the pod and you get the hundred-experiment mess from the section above. The structure is the strategy, which is not a sentence most technology vendors want you to sit with.

The durable advantage is the boring knowledge you never wrote down

People say AI is becoming a commodity, and on the model side they are right. Foundation models keep getting cheaper and more available, so raw capability is arriving on everyone's desk at once.

Which sets up the part that sounds backward: your most valuable AI asset is probably the least impressive thing you own. Not the model. The undocumented knowledge of how your company actually works. Why that particular customer churns. How an exception really gets approved, as opposed to how the policy claims it does. Which rules are unwritten and quietly load-bearing.

A species does not win its niche through raw metabolism. It wins by being adapted to an environment no competitor understands as well, and your business context is that adaptation. A rival can buy the identical model tomorrow afternoon. What they cannot buy is twenty years of knowing where the bodies are buried in your own operation. AI without that context gives you generic answers, and AI with it gives you answers that fit your business, which is the only kind worth paying for. (Also read: Context is the Fuel Every AI System Runs On)

Waiting to buy software might be the riskiest move on the board

A lot of companies are quietly slowing their software purchases, waiting for AI to settle before committing to a platform that could look obsolete a year from now. It feels prudent, and it might be the riskiest thing on the board.

The likelier future is not AI replacing enterprise software but dissolving into it. CRM does not vanish, it grows AI coworkers. The analytics platform stops merely drawing the chart and starts explaining it and running the workflow behind it. If that is the direction, then sitting on the sidelines means falling behind on the exact systems about to get their capability upgrade, while competitors compound their lead inside tools they never stopped improving. (Also read: Incumbent SaaS Companies Are Misreading the AI Moment)

The work ahead is organizational, and that is the hard part

For a century, companies were designed around people and the limits of what people could hold in their heads. The next design problem is a company built around people and capable systems working the same problems together, and the firms that get there first will not be the ones with the best model subscription. They will be the ones that redesigned how decisions get made, how knowledge moves before it walks out the door, and what the organization actually pays people to do.

Go back to the factory one last time. Buying the motor was the easy part. Rewiring the building around it was the hard part, and it took most owners forty years to work that out. AI is sitting at the same stage right now, which means most of the difficulty ahead is not artificial intelligence at all. It is organizational intelligence, and there is no vendor selling that one.

Key Takeaways

• Roughly 95% of enterprise generative AI pilots showed no measurable bottom-line impact, and MIT traced the failures to organizational integration rather than model quality.
• Capability amplifies whatever process it lands on, so adding AI to a broken workflow tends to speed up the dysfunction instead of fixing it.
• The best AI structure is neither one central team nor fully independent departments: functions operate as expert whales, a central AI organization keeps them coordinated as a pod.
• The model is becoming a commodity while undocumented business context stays scarce, which makes that context the real competitive advantage.
• AI is more likely to dissolve into enterprise software than replace it, so pausing purchases can quietly become the bigger risk.
• The hard problem is organizational design and incentives, not the technology, the same way rewiring the factory, not buying the motor, was the hard part of electrification.

FAQ

Why do most enterprise AI projects fail?

The evidence points to organization, not technology. MIT's 2025 research found about 95% of enterprise generative AI pilots produced little to no measurable bottom-line impact, driven by a gap between the tools and how work actually flows. Gartner separately expects more than 40% of agentic AI projects to be canceled by the end of 2027, largely because they are hype-driven experiments applied to problems that never needed them.

What is the whale and pod model for AI teams?

It is a structure where every major business function runs its own AI team that knows its process deeply, and all of those teams roll into one central AI organization that owns shared standards, governance, security, and reusable components. The function teams are the whales, expert in their own environment. The central organization is the pod that keeps them coordinated. It avoids both the bottleneck of a single central team and the duplication of everyone building alone.

Should a company really create an AI leadership role?

Deciding how AI changes a company's operating model, setting governance, and preventing a sprawl of disconnected experiments has grown too large to bolt onto an existing executive's job. Whether the title is Chief AI Officer or something else, the function is becoming hard to skip, because the leaderless projects are the ones getting canceled.

About the Author

Martin Goetzinger has spent his career in enterprise software sales, helping large organizations such as Apple, Microsoft, and Verizon connect data, insight, and action. His work focuses on transforming how businesses measure success and create customer value through technology.

Outside the enterprise world, he writes about the five forces he believes are reshaping everything: AI, blockchain, energy, personalized health, and robotics. Not from a purely technical lens, but from a human one as to how these technologies will redefine work, wealth, and well-being.

He is based in the U.S. and publishes at www.MartinGoetzinger.com.

Disclaimer

The views expressed in this article are the personal opinions of the author and are provided for informational and educational purposes only. Nothing in this article constitutes investment advice, financial advice, legal advice, or any other form of professional advice. Do not make investment or financial decisions based on the content of this article. Always consult a qualified professional before making decisions that affect your finances, business, or livelihood.