Briefing · November 18, 2026
Your AI Knows Everything. That's Exactly the Problem.
Centralising corporate knowledge into AI systems doesn't just change how work gets done — it rewires who actually holds power.

The most dangerous assumption in enterprise AI right now isn't that the technology will fail. It's that centralising corporate knowledge into a single AI system is a neutral, technical act — when in fact it is one of the most consequential power transfers an organisation can make.
Here is the core finding, stated plainly: the Financial Times (2025-07-14) argues that making corporate information machine-readable and centralising it in an AI system redistributes organisational power in potentially disruptive ways — meaning whoever controls the AI system effectively controls institutional memory, decision context, and informational advantage across the firm. That is not an IT architecture decision. That is a governance decision, and most executive teams are treating it like the former.
Why does centralising knowledge in AI shift organisational power?
Power in organisations has never been evenly distributed — it has always clustered around those who control information flows. The senior analyst who knows where the bodies are buried. The operations lead who understands which supplier relationships actually matter. The chief human resources officer (CHRO) who has the unwritten institutional history of every performance review. These people hold power precisely because their knowledge is not fully legible to the organisation at large.
When you train a retrieval system on every Slack thread, contract, and board memo, you do not simply make the organisation more efficient. You make previously tacit, human-held knowledge explicit and searchable — and then you hand the keys to whoever administers that system. As the Financial Times (2025-07-14) notes, this redistribution is rarely discussed in the business case for AI deployment, which tends to focus on productivity gains and cost reduction. The power dynamics are the hidden variable.
The question every executive team should be asking is not "how do we get the AI up and running?" but "who, inside this organisation, loses meaningful authority when the system goes live — and have we told them?"
What is actually at stake when AI systems absorb "information work"?
It helps to have a plain-language frame here. Information work, in the knowledge economy, is the category of tasks that involves gathering, synthesising, and acting on data — writing reports, drafting contracts, reviewing code, preparing board materials. The productivity case for AI is built almost entirely on this category. But Tim O'Reilly's Substack (2025-07-10) makes a sharp counter-argument: information work is actually responsibility work wearing an information costume. The report-writer isn't just formatting data — they are exercising judgment about what matters, what to surface, what to suppress. Strip out the human judgment and you have not automated a task; you have automated accountability out of the system.
This is the tension that MIT Sloan Management Review and Boston Consulting Group (BCG) (2025-07-15) describe as the defining challenge of responsible AI: designing systems that preserve human judgment when and where it counts. For the fifth consecutive year, the MIT Sloan Management Review and BCG international AI expert panel found that organisations consistently underestimate where in their workflows judgment — not just processing — is what actually produces value. The failure mode is not building AI that is wrong. It is building AI that is right on the metrics you can measure, while quietly eroding the judgment capacity you cannot.
Consider what this means structurally. If a law firm's AI can now absorb and replicate the lease-review expertise that previously required expensive outside counsel — as MIT Sloan Management Review (2025-07-15) describes happening at real estate investment firm Alturas Capital Partners, saving hundreds of thousands of dollars — then the expertise that once lived in specialist human practitioners is now embedded in a system the client controls. The practitioner's knowledge has been extracted, encoded, and transferred. That is not collaboration. That is a knowledge acquisition event.
The question for senior leaders is whether the AI systems they are currently deploying are genuinely augmenting human judgment — or whether they are conducting a slow-motion extraction of the firm's most valuable tacit knowledge, packaging it in a tool, and then creating an institutional dependency on that tool that no individual employee can challenge or walk away from.
What should leaders actually do with this frame?
Before the next AI deployment review, ask your team to map not just the workflow benefits but the knowledge flows: What expertise, currently held by specific people, will this system absorb? Who currently holds authority because of that expertise? Where does accountability sit after the system is live? If those questions cannot be answered, the deployment is not ready — regardless of what the pilot metrics say.
The organisations that will manage this transition well are not the ones moving fastest. They are the ones that understand they are not just buying software. They are redesigning the architecture of institutional accountability, one API call at a time.
Created with AI assistance. Editorial oversight: Juergen Ritzek. See our AI disclosure.