Endpoint AI FAQ: controlling AI on employee devices

Anything that sees or limits what AI tools do on, or from, an employee's device: an on-device control, a browser extension, a network or SSE proxy, an AI or MCP gateway, or a written policy. See control types.

Chat tools take text in and send text back. Coding agents act: they read files, run shell commands, install packages and start MCP servers on the laptop, with the permissions of the person who started them. See the on-device guide.

The Model Context Protocol connects AI applications to outside tools and data. Each MCP server adds tools an agent can call, so the list of servers on a laptop is a list of things an agent can do. See MCP.

AI tools, agents, extensions or MCP servers used without the company's approval or knowledge. Finding them is usually the first job of any AI control. See shadow AI.

No. EDR looks for malware and intrusion. AI endpoint controls decide what trusted AI tools may do. They work alongside EDR, and several products here are delivered through it.

Not always. With a handful of users, general information and no coding agents, a written policy, an approved-tools list and the settings your tools already offer may be enough for now. See policy only.

On-device control, because coding agents act on the device and local MCP servers never touch the network. A gateway helps with remote MCP servers. See the on-device guide.

A browser extension on managed devices, or a network or SSE proxy if your traffic already runs through one. See the browser extension guide.

Not much on the device itself. None of the 19 products describes coverage for unmanaged devices on the pages we read. A gateway for traffic you control and a clear policy are the realistic options. See the guide to unmanaged devices.

Most companies start by watching or warning and block only a short list of actions. Warnings keep people on approved tools; blanket blocks tend to push use somewhere you cannot see. See coaching versus blocking.

Only the traffic that passes through it, and only as far as it recognises AI services. Ask your SSE vendor the questions in the network or SSE proxy guide.

19: AI Access Security (Palo Alto Networks), Akamai Workforce Protector, Bay, Bloom Security, Cloudflare AI Gateway, CrowdStrike Falcon Guardian, Ent, Glow Security, Harmonic Security, Koi (Palo Alto Networks), Lasso Security, Neo, Noma Security, Onyx Security, Pluto Security, Prompt Security (SentinelOne), Sun Security, Zenity and Zscaler Generative AI Security. See product notes.

No vendor site here publishes a list price. AWS Marketplace lists the Noma Security Platform at $200,000 for 12 months and Lasso Security's chatbot product at $20,000 for 12 months; Harmonic Security names its tiers without prices. Budget with vendor quotes.

On the pages we read, these match that check fully: Bay, CrowdStrike Falcon Guardian, Lasso Security, Neo, Onyx Security, Sun Security and Zenity. Others describe it partly. See line up products.

Fully: Akamai Workforce Protector, Bay, Ent, Harmonic Security, Neo and Onyx Security. Partly: Glow Security, Lasso Security, Noma Security and Zenity.

Fully: Bay, CrowdStrike Falcon Guardian, Harmonic Security and Noma Security. Partly: Koi (Palo Alto Networks), Lasso Security, Neo, Pluto Security and Prompt Security (SentinelOne).

Palo Alto Networks completed its acquisition of Koi on 14 April 2026 and sells it as Cortex Agentic Endpoint Security. SentinelOne announced its agreement to acquire Prompt Security on 5 August 2025, and its blog post of 3 September 2025 is titled SentinelOne Acquires Prompt Security. Akamai sells the former LayerX as Akamai Workforce Protector, and Cato Networks acquired Aim Security on 3 September 2025.

Bay is an endpoint-first control for AI agents, coding assistants and MCP servers. It decides Allow, Ask or Deny on each action before it runs and rolls out through the EDR or MDM you already run. See Bay's product notes.

It covers the endpoint only, with no browser extension, network proxy or gateway described and nothing on SaaS or cloud AI apps. It has one of the smallest public footprints here and publishes no price, and its headline figures are its own statements.

No. There are no scores, totals or rankings. Products are listed alphabetically or in order of fit to your finder answers, and equal fits are shown as equals.

Public vendor pages, documentation, marketplace listings, blog posts and press releases, read on 27 September or 11 October 2026. Every fact carries its source. We did not test the products. See how the finder works.

We did not find that fact on the pages we read. It may exist; it counts zero in the finder rather than against the product.

Use the contact page with the topic Correction and a public source. We check it and update the data.