AI Usage Control: Why Employee AI Use Needs Its Own Category

According to Akamai’s State of the Internet/Security (SOTI) Enterprise AI Usage Risk Report 2026, nearly half of enterprise users (47.67%) have used AI tools, and close to one in five (18.24%) use them weekly. However, it’s not the quantity of enterprise AI use that stands out as much as its identity: 47.11% of all enterprise AI conversations happen through personal identities rather than corporate-managed accounts — and on leading large language models (LLMs) such as ChatGPT, Claude, and Copilot (the non-M365 variant), that figure climbs past 60%.

Moreover, it’s not just personal identities that are outside the scope of enterprise security teams: Approximately 14% of the conversations that use corporate identities (or ~ 7% of total conversations) use the corporate email via unmanaged personal AI licenses. These are employees using their work email on a consumer tier. To most identity-based controls, that traffic looks sanctioned. In practice, however, that traffic sits outside enterprise governance entirely, with data retention and model training terms your legal team never reviewed.

If this were only about employees pasting customer data into a chat window, the existing stack might eventually catch up. But it isn’t and it won’t.

The AI surface has moved from web interfaces to agents. Local assistants with skills and plug-ins now execute multistep, authenticated tasks. Coding agents run inside development environments, pull from external repositories, and increasingly operate with minimal human confirmation. Model Context Protocol (MCP) connections let agents reach local and external systems on a user’s behalf.

Our research found that nearly 75% of AI browser extensions request high or critical permissions. AI browser extensions are about three times more likely than ordinary extensions to request cookie access, and 16.31% carry known CVEs, compared with 10.80% across all extensions. These are background tools with elevated privileges, sitting in the browser where corporate credentials and sessions live.

Then there’s manipulation. A malicious prompt injected into an external web page can steer an autonomous agent into actions the user never authorized — and do it at machine speed. For example, in 2026, security researchers at LayerX (now part of Akamai) demonstrated a related technique against coding assistants, modifying a project’s instruction file to bend the AI assistant’s output into launching cyberattacks. This is just one example of how on-device AI agents can be harnessed for malicious purposes.

A network proxy can’t see any of that. Neither can a DLP policy written for files.

This is why a new, distinct market category has formed, specifically for the protection of AI interactions. Gartner, in the 2026 Market Overview for AI Usage Control, uses the term AI usage control (AIUC) to describe solutions that manage employees’ use of third-party AI, such as by providing discovery and inventory, risk analysis, policy-based controls, runtime protection, and incident response.

The framing that matters is that AIUC extends visibility and control beyond what endpoint protection platform  and SSE already provide, by adding inspection points and detection techniques built specifically for GenAI behavior.

In practice, an AIUC solution needs to do four things:

1. Discover AI use across the organization, including which tools are in use and who is using them

2. Track AI conversations, not just connections to AI domains

3. Monitor the data moving between users and AI tools

4. Enforce guardrails that block unauthorized tools and unsafe use

Those four capabilities sound straightforward. The difficulty is in the specifics: Which capabilities are genuinely mandatory, which are optional, and what coverage can each deployment approach deliver? That’s where the analyst guidance earns its keep.

Gartner’s own outlook gives a sense of the trajectory: By 2030, the firm expects more than half of enterprises to be using dedicated AI usage control technologies. It’s also worth noting Gartner’s finding that through 2027 the large majority of unauthorized AI transactions will come from employees violating internal policy rather than from attackers. This is, first and foremost, an insider visibility problem.

A new category attracts a crowded field. Gartner is currently tracking more than 50 vendors in AIUC, roughly 80% of them early-stage start-ups. Gartner is approaching the problem from different architectural starting points, which makes the problem of vendor selection particularly challenging.

This is why Gartner’s guidance for buyers is worth taking seriously. That guidance includes:

• Establish your top-level AI policy before you evaluate tools, and be explicit about which data and systems are off limits to AI. 

• Use discovery data to inform that policy rather than writing it in the abstract. 

• Prioritize vendors that integrate with the controls you already run.

• Run proofs of concept with more than one vendor, specifically to find your real visibility gaps. 

• Keep contract terms short, because this market will look different in 18 months.

Akamai offers an AI usage control solution through Akamai Workforce Protector (formerly LayerX), and we’re listed as a Representative Vendor in Gartner's 2026 Market Overview for AI Usage Control . Unlike heavy enterprise browsers that force employees to give up their preferred tools, Akamai Workforce Protector operates via a lightweight browser extension that inspects prompts in real time right at the point of interaction.

But the more useful thing we can do right now is put the analyst’s own framing in your hands. This is why we have licensed a complimentary copy for you to download.

Gartner’s Market Overview for AI Usage Control is the reference document for this category. It covers what AIUC is; why SSE, DLP, and EDR/XDR leave gaps; the mandatory and optional capabilities to evaluate against; the deployment form factors available; and the vendors currently in the market.

_{Gartner, Market Overview for AI Usage Control, Ashish Suraj Bhan, Peter Firstbrook, John Watts, 20 August 2026}