87% of U.S. Workers Use AI Weekly — and 44% Break the Rules

By Grace Kim

KPMG's 2025 American Worker Survey found that 87% of U.S. workers use AI at least weekly, and 51% use it daily. Yet a separate global study from KPMG and the University of Melbourne — covering more than 48,000 people across 47 countries — found that 44% of employees had used AI in ways that contravened their organization's policies. Another 66% had relied on AI output without evaluating its accuracy.

Those numbers define the problem for business owners. AI adoption has already happened inside most companies, sanctioned or not. What hasn't happened, according to an Entrepreneur contributor analysis, is management catching up.

Employees are drafting emails with ChatGPT, analyzing spreadsheets with Copilot, researching competitors with Claude and experimenting with tools their employers have never heard of. Some of that usage is official. Some almost certainly isn't.

Why the old question is the wrong question

For the first few years of generative AI, the debate centered on whether employees should be allowed to use it at all. The analysis argues that question is obsolete. "Employees are using it, and the much more important issue is whether anybody is actually managing how."

The shift matters because AI's role in work is escalating. The piece lays out what it calls the AI Management Ladder, five levels of delegation:

• Level 1: No AI; the employee does the work.
• Level 2: AI assists the employee.
• Level 3: AI handles defined parts of a workflow.
• Level 4: An AI agent executes multiple steps toward an objective.
• Level 5: A human-and-agent team, where employees manage outcomes while AI executes most of the work.

At Level 1, an employee might use ChatGPT to summarize a document. Six months later, that same employee's AI might research potential customers, qualify prospects, draft personalized outreach and update the CRM. The employee still owns the outcome. The machine now does much of the work.

Each step up the ladder raises productivity — and raises the stakes, because technology gains greater responsibility for work that affects customers, employees and the business.

What should every manager ask?

Managers don't need to master every AI model or build a complicated governance structure, the analysis says. But they should be able to answer six questions:

• Who is using AI?
• What are they using it for?
• How are they using it?
• What company, customer or employee data are they putting into it?
• Who checks the output?
• Who owns the result?

The last question carries the most weight. Consider a scenario the analysis raises: an employee uses AI to prepare a customer proposal, the AI includes an outdated price or invents a product capability, the employee misses it, and the proposal reaches the customer. "Who owns the mistake? The answer cannot be ChatGPT."

Accountability has to sit somewhere inside the organization — and ownership needs to be explicit before something goes wrong, not afterward.

What does a workable policy look like?

The proposed fix is what the analysis calls an AI Operating Agreement, splitting AI use into three categories:

• Green — use freely. Brainstorming, formatting, summarizing non-sensitive information, first drafts, ideation. Low-risk activities where employees should experiment with approved tools.
• Amber — human in the loop. Client-facing materials, financial analysis, important research, hiring-related work, recommendations — anything where an error could materially affect the business. An accountable person checks before anything ships.
• Red — don't delegate it. Sharing prohibited confidential information, uploading sensitive customer or employee data into unauthorized systems, or letting AI make consequential decisions it hasn't been authorized to make.

The boundaries differ by company. A five-person marketing agency shouldn't have the same rules as a financial services firm with 5,000 employees. What matters is that everyone understands where the lines sit.

The framework came out of a real engagement. While reviewing employee roles with the owner of a growing service business, the author found traditional job descriptions explained what each employee was responsible for — but not which parts of the job AI could perform, where it needed supervision, or where it shouldn't be used at all. Recoding roles as Green, Amber and Red turned job descriptions from task lists into accountability maps.

Why a policy isn't management

Here's where owners develop a false sense of security: write an AI policy, email it around, collect acknowledgments, declare victory. The KPMG and University of Melbourne data shows the gap — 44% of employees violated those very policies, and widespread rule-breaking persisted regardless.

The alternative is making AI part of normal management conversation. How did you produce this? Where did AI help? What did you check? What information did you give it? Could AI safely take on more of this workflow? Those questions shouldn't be gotchas. They should be as routine as asking who owns a project.

Restricting AI outright tends to backfire, the analysis warns. Banning tools doesn't stop usage; it makes usage invisible, driving employees to hide what they're doing.

The better target is an organization where employees are comfortable saying when they use AI, understand the boundaries, and stay accountable for results. Businesses have spent years teaching employees how to use AI. As AI moves from answering questions to completing workflows, teaching managers to manage AI-using people becomes a fundamental management skill — one most businesses haven't developed yet.