With all the talk of demos and pilots these days, it’s worth remembering the efforts of pilots from a different era of experimental engineering.
In the 1940s, there was a problem. Pilots described encountering “an invisible brick wall in the sky”, what we now know was the sound barrier (side note: if you don’t know the science of how and why it works, it’s cool).
The first destructive encounters with this force were a surprise. But even after engineers knew how it worked, overcoming it was harder than anyone expected.
Adopting AI at enterprise scale presents an eerily similar comparison because the power and speed of the technology expose and interact with “invisible” forces in unexpected ways . From conflicts in org incentives and ownership, ethical concerns and simple inertia of human habit, the AI scale barrier is becoming all too clear.
Obviously, we know how the story of the sound barrier ends. Through bravery, ingenuity, and perseverance, humanity eventually went supersonic. As the story of breaking the barrier to AI at scale is being written, here are the forces enterprises will encounter and need to break through.
How hard are enterprises hitting the wall? MIT reports that of the $30 to 40 billion invested in GenAI, 95% of organizations haven’t seen a penny in actual P&L return. Experts predict only 50% will succeed by 2030.
To be fair, we’re still in the early days. Expecting too much return too soon from initial investments would be unrealistic.
But here’s the damning truth: the scale barrier isn’t a natural phenomenon or even a technical limitation. It’s a manmade consequence of underinvestment, and poor collaboration to overcome blockers that are well known, but still unsolved.
People come to AI for authoritative, accurate answers. What they often get are overconfident fabrications. That’s because the fuel for enterprise AI is full of pollutants.
The top two problems are:
Inconsistencies and contradictions. As content libraries and knowledge bases sprawl, outdated and incorrect information persists.
On a human scale, the consequences of this widely open secret are small and often negated by people’s intuition and good judgement. AI removes those safeguards. The resulting speed of production and deluge of output overwhelms human capability for review and oversight, letting more and more slop seep through.
When it all goes into the model, you can’t know what will come out. In McKinsey’s State of AI 2025, 51% of organizations using AI have seen negative consequences related to AI inaccuracy.
Quality and readiness. Accurate content doesn’t make it good or usable. Content has to be structured and readable for both humans and AI models. It’s something we’re intimately familiar with through our Agentic Knowledge Management engagements.
Here’s a before-and-after example of how we drove AI improvements in a knowledgebase:
The blocker to AI at scale is the scale of the content itself. Even after you cull your libraries and knowledge bases by 70–80% (a pretty typical reduction in content debt), reviewing, improving, and managing that content appropriately is no small task.
Case in point: In a recent knowledge management engagement, a US Fortune 500 client was seeing a meager 30% success rate for accuracy and relevance in its AI tests. By optimizing content on just their top 10 topics, it saw a 90% success rate. With the value of a rigorous approach to content proven, the client was able to reduce content debt to the tune of 100K+ articles, helping them achieve scale.
It’s relatively simple to set up AI for any one task. The problem is: enterprises run on complex workflows made of hundreds of tasks. Trying to insert AI into only a few of them is like trying to go supersonic by adding an extra blade to your propeller plane.
The attempts to achieve AI at scale run into two blockers:
One example that will make sellers and marketers groan in recognition: finding relevant case studies and customer evidence. When you need an answer to nuanced questions like “Do we have any case studies for how our product was used in a transportation company and integrated with SAP?”, the answer is “Maybe, start reading.”
One of our clients suffered from “too much of a good thing”—an overwhelming library of over 5,000 assets. This resulted in sellers rarely searching beyond five well-known, but overused examples.
Even though the domain was limited and the data was there, the integration was still complex. Why? Basic AI tools like Microsoft Copilot studio can only review data so deep and wide without specific engineering and infrastructure. And because data lives across multiple fragmented sources, the AI first needs a significant data engineering effort to normalize, clean, and consolidate everything before it can deliver accurate, trustworthy results at scale.
Getting it working involved:
Was all that effort worth it? 240%. That’s how much time the average user saved searching, helping them find more relevant examples to share.
In pilots, security is mostly an afterthought. The stakes are low. The surface area is small. For AI at scale, both of those things change dramatically, and so do the cost of managing them.
Even in the clearly defined, high-stakes domain of legal practice, vigilant oversight is still required. In 2024, Stanford found that general purpose AI models hallucinated 58-82% of the time when asked for legal answers. But even with specifically trained legal models in 2025, hallucination rates were still between 17-34%.
BCG cites organizations that successfully scale AI devote 70% of their effort to people and processes. That means embedding oversight in all aspects of AI systems, business processes, and culture for people to find and fix issues. And this isn’t an occasional function. It has to match the always-on speed and scale of AI output.
Fortunately, the problem contains the solution: Building workflows and systems using AI to monitor the AI. An Agentic AI solution for RFP responses we developed for a mid-sized transportation and logistics company. The client adopted a two-pronged solution to cover output and oversight through:
Every technology challenge above is solvable with enough time and budget. What’s lacking is courage and commitment to actually take to the skies, vs just wanting to be seen trying.
The organizations succeeding at AI scale share a few specific characteristics.
As with the sound barrier, every scientific breakthrough was achieved through endless hours of meticulous, painstaking, unsexy effort. The fact is most solutions to scale AI aren’t complicated. They’re just hard work. And they take time.
That’s not an answer most companies want to hear, but the longer people try to chase the quick win, the further they’ll be from reaping the benefits of enterprise AI at scale.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。