Use AI, but not too much: the mandate is backfiring

For two years, large employers pushed one message: use AI, or risk falling behind. Accenture and Coinbase told staff that failing to embrace AI could jeopardize their careers. Starbucks tied a quarter of its technology bonuses to AI adoption. The instruction was simple: get usage up.

Now, as Bloomberg reported this week, some of those same companies are pulling back. Walmart has capped employee use of its internal AI agent. Uber is limiting spending on AI coding tools after exceeding budget. Amazon shut down an internal leaderboard after employees began "tokenmaxxing" to climb the rankings rather than improve outcomes.

The message has changed. Use AI... just not too much.

That's the whiplash. And it's a leadership problem before it's a cost problem.

CEO and Founder of Reejig, Siobhan Savage put it plainly to Bloomberg:

Every enterprise is deploying AI. Almost none can see the work they're deploying it into. AI capability is compounding. Work visibility is not.

That gap explains why so many organizations are discovering that usage alone was never a strategy.

Why the usage mandate is backfiring

Usage was the easiest thing to measure, so it became the goal. The problem is that usage is an input, not an outcome.

A team can consume millions of tokens and change nothing about how work actually runs. The spend gets counted. The work stays the same. Amazon's tokenmaxxing problem illustrates the issue: when leaders reward consumption, employees optimize for consumption, not better work.

Now the bill has arrived. Agentic systems consume far more compute than the chat tools that came before them. Budgets are under pressure. Executives are asking harder questions about return. So organizations are capping access.

But limiting usage does not solve the underlying problem. It only limits the cost of a program that was never tied to work outcomes in the first place.

According to Bain, many enterprises still struggle to connect AI spending to measurable business value. MIT research found that most organizations have yet to realize meaningful returns from AI investments. McKinsey reports that only a minority have fundamentally redesigned how work operates around AI.

Tools were deployed. Work stayed the same. The meter kept running.

The part employees notice is not the spend. It is the contradiction. "Use AI or fall behind" and "stop using so much AI" are coming from the same leadership teams, often in the same quarter. That is what erodes trust. Not the technology. The mixed signal.

What leaders should measure instead

Stop measuring AI usage. Start measuring whether work is running differently.

That shifts the conversation from consumption dashboards to operational outcomes: hours recovered and redeployed, capacity created, cycle times reduced, risk removed, and revenue generated. If a workflow looks identical before and after an AI rollout, there is no return to report, regardless of how much usage increased.

This is where many enterprises hit a second problem. They cannot see the work clearly enough to prove what changed. Org charts describe reporting structures. Job descriptions describe roles. Skills frameworks describe capabilities. None of them describe the tasks AI is actually changing.

You cannot measure a change to work you cannot see.

Work Architecture is the foundation. It provides a structured view of how work actually runs across tasks, subtasks, workflows, and skills. Without that foundation, organizations struggle to demonstrate what changed, who changed it, and where value was created.

The question is not whether people are using AI. The question is whether work is operating differently because of it.

The order matters: work first, tools second

Many organizations inverted the sequence. They deployed tools first, then started looking for value. The sequence that survives board scrutiny works the other way around.

Make work visible. Identify which tasks should remain human, which should be augmented, and which should be automated. Redesign the workflow. Then connect AI into that redesigned work. Then measure what changed.

Now there is a baseline. Now there is evidence. Now there is a return to prove.

That is the difference between an adoption program and a redesign of how work runs. One drives a usage metric. The other changes work and measures the outcome.

What the whiplash should teach leaders

Organizations are not wrong to control runaway costs. But cost controls are not a strategy. Pull back too aggressively and the organization loses value alongside waste.

The lesson is not to do less AI. It is to stop managing AI through usage metrics, because usage was never the thing creating value in the first place.

A usage target asks employees to perform adoption. A work-based target asks the organization to prove impact. Only one survives a budget review. Only one builds trust.

The enterprises that get this right in 2026 will look unremarkable from the outside. No usage leaderboard. No adoption contests. No all-hands meeting celebrating token counts. Just work that runs differently every quarter, with measurable evidence to prove it.

Every enterprise is deploying AI. The question is whether work is changing with it.

Reejig is the Work Operating System for AI-powered work. The platform enterprises use to see how work runs, build AI workflows, orchestrate agents, and drive adoption.

It makes work visible at the task level, redesigns how work runs across humans and AI, and measures actual changes to work, not consumption metrics.

If your AI program is still being measured on usage, that is the conversation worth having now.