A temporary restriction can contain harm. It cannot replace the use cases, decision rights, competency, ownership, and quality measures adoption requires.
By Renee Cannon, Founder of eunoiaAI
August 27, 2026
![[Hero Alt Text]](https://cdn.prod.website-files.com/6a5e755fad03f8f3b4db484b/6a8f34f3e2e8c5d34be584ce_Paused%20access%2C%20clear%20ownership_ChatGPT%20Image%20Aug%2026%2C%202026%2C%2002_47_50%20PM.png)
In August 2026, Paysend chief executive Ben Chisell described why he restricted AI access for much of his workforce. Some employees were doing excellent work with AI. Others were producing shallow analysis, failing to follow through on commitments, and relying on AI-generated meeting summaries without doing the unglamorous work of confirming decisions, owners, and next steps.
The result was not merely disappointing output. Stronger employees had to absorb additional cleanup work.
Chisell’s response was selective. Most employees retained AI access for general research and could request broader tools with a justified use case. About 30 employees he described as top performers received unlimited access. He also said the restriction was not intended to be permanent.
As a circuit breaker, that decision is understandable. As a complete adoption model, it leaves a more important question open: was AI intentionally introduced into the work, or was software access distributed and expected to become useful on its own?
A license establishes availability. It does not define where the tool belongs, which work it may influence, what quality looks like, or who remains responsible.
When those decisions are missing, employees invent them individually. One person uses AI as a drafting assistant. Another treats it as an analyst. A third assumes an automated meeting summary has transferred ownership of follow-through to the software. Leaders then evaluate “AI use” as if those were comparable activities.
They are not.
The reported Paysend examples point to an adoption problem larger than prompting. A meeting summary is not a commitment-management system. A plausible analysis is not a completed management judgment. If the workflow does not require an owner to confirm actions, inspect the evidence, and answer follow-up questions, AI can make an existing accountability gap look more polished.
AI can accelerate strong habits. A disciplined employee may use it to compare evidence, find missing considerations, prepare alternatives, and reduce time spent on routine production. The same system can accelerate weak habits by generating an answer before the user understands the question.
That does not excuse poor performance. It does change the diagnosis.
If an employee repeatedly misses commitments, the root issue may be role clarity, management, workload, skill, or accountability. If analysis cannot survive follow-up questions, the organization may lack a defined evidence standard or review practice. Removing AI may expose those problems more clearly, but it does not resolve them.
This is why organizations should separate an AI control decision from a performance-management decision. They may interact, but treating access as the only variable can obscure what the work needed before AI arrived.
Restricting access is appropriate when the organization has evidence that use is creating material rework, unreliable decisions, privacy or security exposure, customer harm, or an unmanaged transfer of authority. A pause can contain the problem while leaders determine what failed.
It is especially useful when no one can answer basic questions: Which use cases are approved? What information may be entered? Which outputs require verification? Who owns the final decision? What is the escalation path? How will the organization know whether the tool improved the work?
In those conditions, proceeding at full speed is not adoption. It is uncontrolled experimentation inside operating work.
The pause becomes valuable when it creates space for diagnosis and repair. Without that next step, restriction can become a permanent substitute for organizational learning.
Broader access should be connected to demonstrated readiness, but readiness needs an observable definition. A controlled reintroduction can include:
The NIST AI Risk Management Framework organizes responsible AI work around governance, context, measurement, and ongoing management. That structure is useful here because it moves the conversation beyond “allowed” and “not allowed” toward evidence about how a particular use operates.
Giving broader access first to people with demonstrated judgment can concentrate experimentation among users most likely to produce value and least likely to create avoidable cleanup. It can also generate internal examples grounded in the organization’s own work.
But performance tier is an imperfect proxy for AI readiness. Strong employees are not immune to overconfidence, weak verification, or using AI outside their expertise. Meanwhile, employees who cannot practice in controlled settings cannot develop the very competency required to earn access.
A permanent two-tier system may widen the capability gap. Top performers accumulate tools, experience, and visibility while everyone else is left with abstract training and fewer opportunities to improve. It may also treat prior performance problems as fixed traits instead of asking which skills and operating conditions can be developed.
A stronger model is graduated access with a credible path forward: bounded tools, low-consequence use cases, supervised practice, evidence of quality, and expanded permission when competency is demonstrated.
Chisell’s most useful observation may be that tool usage is less important than impact. Login counts reveal activity. They do not reveal value.
Organizations can measure whether AI-assisted work improves cycle time without increasing rework, whether analysis withstands review, whether commitments are completed, whether exceptions are caught, and whether stronger employees spend less time repairing other people’s output. Measures should be specific to the workflow and compared with a baseline.
The 2025 study of AI-assisted knowledge work from Carnegie Mellon and Microsoft Research reinforces the need for this specificity. It found both reduced cognitive effort and a shift toward verification, integration, and stewardship. The question is not simply whether effort decreased. It is whether the right effort remained.
Not every employee needs unrestricted access. Not every task should use AI. Some uses should remain prohibited, some should require approval, and some work should remain human-led because the judgment, relationship, or consequence warrants it.
A responsible transformation can conclude that the organization should pause a use case, repair the workflow, improve data, narrow access, or keep part of the work manual. Those are not failures to adopt AI. They are evidence that the organization is making an operating decision rather than chasing usage.
Restricting access can stop the immediate problem. Adoption begins with what the organization learns and redesigns next.
eunoiaAI helps organizations define the workflow, use cases, human decision rights, evidence, and operating controls required for purposeful AI use.
AI adoption does not begin with access. Employees need to understand where AI fits, what remains human-led, what authority it has and who is accountable when something goes wrong.
As AI lowers the cost of prototypes and agents, the enterprise bottleneck moves upstream: deciding which workflows merit AI, what authority systems should hold, and what evidence justifies scale.
A demo proves that an agent can complete a prepared task. A business system must keep working when the data, users, tools, and circumstances are not prepared.