Your company has the licences. People have taken the training. A steering committee is meeting. Several pilots are underway. The roadmap extends into next year.
All of that can be true while the organization falls further behind.
The most dangerous AI gap may no longer be technological. It may be institutional.
Institutional lag is the distance between what the technology makes possible and how quickly your organization can recognize, authorize and operationalize it.
That gap rarely appears on an AI dashboard. It hides in incentives, approvals, silos, budget protection and the unspoken fear of what greater productivity will mean for people's roles. If those conditions remain unchanged, adding more AI can make the organization look active without making it more capable.
A seven-year-old warning about institutions
In 2019, Peter Thiel joined Eric Weinstein for the first episode of The Portal, titled “An Era of Stagnation & Universal Institutional Failure”. It was not a discussion about corporate AI adoption, and it does not offer an AI transformation method. Its value here is as a lens.
Thiel and Weinstein examine why institutions can maintain a story of progress while becoming less willing to question their own assumptions. At about 39:50, Thiel describes the institutional rule as “no polymaths allowed.” Specialists may understand their own domains deeply, yet no one is encouraged—or authorized—to connect the whole system.
Later, at about 2:01:14, Thiel offers the principle that matters most for this discussion: “when we can't talk about things, we can't solve them.”
AI has made both problems harder to ignore.
Meaningful AI transformation cuts horizontally across technology, operations, data, finance, customer experience, human resources and risk. Most organizations still divide authority vertically. IT owns the platform. Operations owns the process. Finance owns the business case. Risk owns the constraints. HR owns the roles. Each group can perform its assigned part correctly while nobody owns the complete result.
The technology has moved. The institution may not have.
Adoption is not transformation
The current evidence supports a distinction between using AI and changing how an organization works.
The 2026 Stanford AI Index reports that 88% of surveyed organizations used AI in at least one business function in 2025, and 70% used generative AI. Yet agent deployment remained in the single digits across most functions. AI activity has become widespread; scaled operational integration has not.
Canadian productivity evidence points in the same direction. Statistics Canada found that 12.2% of Canadian firms used AI in 2025. AI adopters initially appeared 16.8% more productive, but that relationship weakened or disappeared after researchers accounted for pre-existing productivity and complementary capabilities.
The conclusion is not that AI fails to improve productivity. It is that the tool is only one part of the system. Statistics Canada identifies organizational change, digital infrastructure, skills, research and development, and broader innovation as important complements.
Economists have seen this pattern with other general-purpose technologies. The NBER paper The Productivity J-Curve explains that major technologies often require new processes, products, business models, human capital and ways of organizing production before their gains become visible.
A licence is access. A pilot is a test. Transformation is a change in operating capability.
A faster production line cannot fix an unchanged operating model
AI behaves the same way when it is added to an unchanged operating model. A team may complete one task in minutes, yet the result can still wait days for permission, cross several departmental boundaries or enter a process that measures activity instead of business value.
The useful question is not whether one piece of technology became faster. It is whether the organization redesigned the complete flow of work around what is now possible.
That is why AI adoption has to be management-led.
A CEO or president does not need to become an AI engineer. But leaders do need enough firsthand experience to understand what the technology can and cannot do, ask useful questions and recognize where an existing assumption is no longer true. If leaders will not use AI in their own work, employees receive a clear signal: this is another initiative being delegated downward, not a change in how the company intends to operate.
Leadership also has to confront a second issue: trust.
Organizations often tell employees to innovate while requiring permission for every meaningful experiment. Leaders may fear losing control. Staff may fear that demonstrating a better method will threaten their jobs. The result is a careful performance of adoption in which people attend training and run harmless pilots while avoiding the changes that would affect authority, budgets or roles.
Trust does not mean leaving people unsupervised. Leaders remain involved, define boundaries and own the consequences. Employees, in turn, need enough autonomy to step up, test responsibly and lead a better way of working.
NIST's AI Risk Management Framework supports this more useful form of governance: clear roles, continuous oversight and decisions grounded in business context and use-case risk. A reversible internal experiment should not automatically face the same process as a consequential production system handling sensitive information or making high-impact decisions.
Why capable people may still resist
It is tempting to describe resistance as a failure of attitude. In many organizations, it is a rational response to the incentives already in place.
Consider two anonymized patterns Stephen has encountered more than once.
An IT group understands that AI could materially shorten development work. Acting on that knowledge could improve the company, but it could also raise uncomfortable questions about staffing, budgets and the department's future importance.
An administrative employee discovers that work which once occupied most of a day can be completed in a fraction of the time. The employee may see an opportunity to take on higher-value work—or evidence that the current role is in jeopardy.
Neither reaction is solved by another training session.
Leaders have to make the organizational consequences discussable. What happens to the time that is saved? How will people develop expertise if junior work changes? Which team owns a workflow that now crosses several specialties? Which reports, meetings or systems will stop? How will employees be rewarded for exposing a process that no longer makes sense?
This is where Thiel's warning becomes operational. If people cannot talk honestly about the effect of AI on jobs, authority and prior decisions, the organization cannot solve for those effects. Silence produces an artificially reassuring picture: everyone supports transformation in principle while resisting each concrete change it requires.
The OECD's research on AI adoption in firms identifies change management, leadership, skills, use-case understanding, data and infrastructure among the recurring conditions and barriers. The lesson is straightforward: employee reluctance is not separate from the operating model. It is information about the operating model.
How to measure institutional lag
Most AI scorecards count activity: users, prompts, licences, training completions and pilots. Those numbers can help manage a programme, but they do not tell a leader whether the organization is becoming more capable.
A more revealing diagnostic follows five intervals and decisions.
1. Recognition time
How long does it take the organization to turn a demonstrated AI capability into a specific business opportunity? If people discuss models more often than customer or operating problems, recognition is already slow.
2. Decision time
How long does it take to authorize a bounded, reversible experiment? If the approval process takes longer than the experiment itself, governance may be mismatched to the risk.
3. Operationalization time
How long does it take positive evidence to become a live workflow with a named owner, appropriate controls and a clear escalation path? Many pilots end here: technically successful, institutionally homeless.
4. Removal discipline
What will stop when the new workflow works? Adding AI without removing steps, reports, meetings or systems often makes the process more complicated rather than more productive.
5. Outcome evidence
What changed for the business? Look for cycle time, quality, cost, revenue, customer experience, staff capacity or controlled risk—not only adoption statistics.
This is not a universal scoring system. It is a way to expose the difference between activity and movement.
You should be concerned if your organization has many pilots but few production workflows, if every experiment enters the same enterprise governance process, if no one owns an outcome end to end, or if employees can privately identify obsolete work but cannot challenge it publicly.
The pressure on software businesses is a warning for everyone
This institutional gap may be especially visible in software-as-a-service businesses.
Broad AI platforms are expanding across functions that once required separate tools. OpenAI, for example, now describes workspace agents that can perform recurring workflows across sales, IT, finance and product operations. Anthropic publishes examples of Claude being used across marketing, engineering, strategic research and operations.
That does not mean every specialized application will disappear. Deep workflow knowledge, proprietary context, integration, accountability and control can remain highly valuable. But an undifferentiated point solution cannot assume its existing budget line will protect it as general platforms become more capable.
The same test applies inside any company. A system, vendor or role should remain because it continues to create differentiated value—not merely because removing it would expose an earlier decision.
Close the gap one complete workflow at a time
An honest reset does not begin with another company-wide AI announcement. It begins with one material outcome.
Choose a complete workflow with a meaningful business consequence. Give one cross-functional owner enough authority to examine the steps, information, decisions, risks and handoffs from beginning to end. Create a faster route for reversible learning while retaining stronger controls where the consequences justify them.
Then ask two questions that most transformation programmes avoid:
- What can AI help us do that was not practical before?
- If it works, what are we prepared to stop doing?
The first question finds technical possibility. The second tests institutional readiness.
Your company may already have AI. Whether it falls behind will depend on something harder to purchase: leadership willing to use it, people trusted to act, incentives that reward honest improvement and an organization prepared to remove the bottlenecks around a faster operating capability.
Written from Stephen Lau's experience and point of view, with AI-assisted research and editorial production. Reviewed by Salient AI before publication.