Token limits are becoming a Shadow AI problem.

Recent improvements in efficient open models have changed the economics of my own AI work. I can run multiple local agents, route tasks between larger and smaller models, and reserve paid frontier-model capacity for the work that benefits from it most.

That is useful for an individual experimenter. It also points to an emerging enterprise challenge.

Giving employees an AI subscription is a start, but it is not an AI strategy.

When cost controls collide with productivity

As adoption grows, organisations need token budgets, usage limits, and other controls to keep AI spend predictable. The problem is that a single limit rarely reflects the value or complexity of every task.

An employee who has redesigned a workflow around AI can hit a limit precisely because the tool is working. If approved capacity disappears halfway through the day, the immediate cost problem may be solved while a larger productivity problem is created.

Highly engaged users do not necessarily stop. They may switch to personal subscriptions, local models, browser extensions, or routers connected to unmanaged providers. Data can then move into services with different retention, training, access-control, and audit rules.

Shadow AI is often a signal

Unauthorised tools are a governance risk, but they are also evidence. They can reveal where approved services are too constrained, where employees do not understand the controls, or where the organisation has failed to provide a safe path for a valuable use case.

The answer is not simply tighter control of the people getting the most value from AI. A durable response combines enablement, governance, and architecture.

1. Help employees get more from approved capacity

Many teams still treat prompting as the main AI skill. The larger opportunity is to teach reusable workflows: structuring context, separating research from generation, preserving useful artefacts, selecting the right tool, and avoiding repeated high-cost work.

Training should help people understand why limits are being hit and how to use approved systems efficiently. It should also make the boundary clear: which information can be used, where it can go, and what must never leave controlled environments.

2. Build a secure portfolio of models

Not every task requires the largest frontier model. Classification, extraction, formatting, routine code changes, and other bounded work may be handled by smaller or lower-cost models when they meet the required quality and risk threshold.

Organisations should prepare approved access to a portfolio that can include frontier subscriptions, secure deployments of open models, and specialised tools. The right mix will depend on the sensitivity of the data, the consequences of error, latency, quality, and total cost.

3. Route work according to value and risk

A model router can direct work to an appropriate capability tier instead of treating every request alike. But routing is more than a cost optimisation. It needs policy.

A production approach should consider:

• The sensitivity and permitted location of the data
• The capability and reliability required for the task
• The consequence of an incorrect or incomplete result
• Latency, availability, and unit economics
• Retention, training, logging, and audit requirements
• When escalation to a stronger model or a human is necessary

This creates a more useful control than a flat token ceiling. Expensive capacity can be preserved for high-value work while routine tasks remain inside an approved, observable environment.

Manage the system, not only the subscription

AI strategy now spans procurement, employee capability, product design, security, and operational governance. Leaders need to understand which work is generating value, where people are constrained, and which alternatives they reach for when the approved path fails.

The objective is not unlimited usage. It is a managed system that gives employees enough room to improve their work without forcing the organisation to choose between runaway spend and unmanaged risk.