Skip to content
Endpoint AI Finder

CONTROL TYPE · POL

Policy-only AI control: when you may not need a product yet

Short answer

For a handful of people using AI chat tools with general information, or for devices you cannot install anything on, a written AI use policy, an approved-tools list and the settings your AI tools and security stack already offer can be the right first step. Policy costs little and every company needs it anyway. It stops being enough when coding agents arrive, when sensitive data is in reach, or when AI use spreads beyond what you can follow by hand.

By The Device Desk · Endpoint AI Finder · Published 2026-10-11

Desk research from public vendor pages read on 27 September and 11 October 2026. Profiles follow our published method; we did not run the products.

Vendor pages read September and October 2026 · updated 2026-10-13

What does policy only do?

A policy-only approach has four parts: a list of approved AI tools, with who may use them and for what data; a rule for adding new tools and MCP servers, with a named reviewer; the administrator settings your AI tools offer, such as permission prompts and allowed MCP servers, where the vendor documents them; and a question to your existing EDR, device management and DLP vendors about what they already see of AI tools.

The MCP specification puts part of this into the tools themselves. Its tools page says there should always be a human in the loop with the ability to deny tool invocations, and that clients should ask for confirmation before sensitive operations. Using those prompts as intended is a policy decision, not a purchase.

The OWASP GenAI Security Project's 2026 Top 10 for LLM Applications, announced on 1 September 2026, ranks Excessive Agency third: agents with more tools, permissions or autonomy than the task needs. A policy that limits which tools and permissions agents get is the first defence against it.

When do you need it?

  • A handful of people, or one or two teams, use AI tools.
  • The data at stake is mostly public or general internal information.
  • Most devices are unmanaged, so nothing installed would reach them.
  • You do not yet know which AI tools are in use and want to start with a survey and a list.

What can it not see?

Policy does not see anything. It relies on people following it and on the settings of each tool. It cannot tell you that a developer added an MCP server last night, or stop an agent that was tricked by content it read into running a command. When those risks are real, a product that sees and decides on the device or at a gateway becomes worth its cost.

Which finder answers point to POL?

  • +2if the answer to "Which AI tools do employees use at work?" is Not sure yet
  • +2if the answer to "Do you manage the devices people use for AI?" is Mostly unmanaged devices
  • +3if the answer to "How sensitive is the data AI tools can reach?" is Mostly public or general internal information
  • +1if the answer to "Which of these do you already run?" is None of these
  • +2if the answer to "When an AI action looks risky, what do you want to happen?" is Only watch for now
  • +3if the answer to "How widely is AI used?" is A handful of people
Try the finder

What does policy alone cover, and what does it miss?

What policy covers: which tools are approved; which data may go into them; who reviews new tools; how exceptions are asked for; what the tools' own permission prompts should be set to.

What policy misses: AI tools nobody declared; MCP servers added from a terminal; agents acting on instructions hidden in content they read; any record of who or what did what.

Signs it is time to add a product: coding agents with MCP servers are in daily use; agents can reach secrets or production systems; AI use has spread beyond a few teams; an auditor or customer asks for a record of AI activity.

What should you ask?

  • Which AI tools are in use today, and who approved them?
  • Which administrator settings do our AI tools offer, and are they on?
  • What do our EDR, device management and DLP products already report about AI tools?
  • Who reviews a request for a new AI tool or MCP server, and how fast?
  • What would make us revisit this decision: more users, coding agents, or more sensitive data?

Sources

Source: MCP specification: tools (2026-07-28) · Read 27 Sep 2026

Source: OWASP GenAI Security Project: 2026 Top 10 for LLM Applications announcement · Read 27 Sep 2026

Learn more

Questions people ask

Is it responsible to have no AI security product?

For some companies, for now, yes: few users, general information, and a clear policy. The finder points there only when your answers do, and it lists the signs that should make you revisit the decision.

What should an AI use policy include?

An approved-tools list, the data allowed in each, who reviews new tools and MCP servers, how to ask for an exception, and how the tools' own permission settings should be configured.

When should a policy-only company look at products?

When coding agents and MCP servers arrive, when agents can reach secrets or production systems, or when AI use spreads beyond a few teams. Run the finder again at that point.