Perplexity Enterprise Controls: How to Govern Perplexity (and Comet) at Work Without Blocking It

Answer first: Perplexity is safe for work only on the account that is set up to be safe. Perplexity's own help center says Enterprise data is never used for AI training and uploaded files are kept for just seven days, while Free, Pro and Max accounts have AI data retention for training switched on by default. Same domain, same app, opposite data policy. The control that matters is not "allow Perplexity" or "block Perplexity." It is knowing which account each employee is signed into and enforcing the right one. dope.security does that from the endpoint: it discovers Perplexity use with a personal-versus-enterprise account breakdown, blocks sensitive prompts and uploads with Dopamine DLP, and applies the policy on every laptop in under a minute.

Perplexity has become the second or third AI tool in most companies without anyone deciding it should be. It is fast, it cites sources, and it now ships a browser called Comet that can act on pages for the user. IT teams that spent 2025 getting ChatGPT and Claude under control are finding a third door open. This guide covers what Perplexity's enterprise controls actually do, what they cannot do, and how to govern Perplexity at work without becoming the team that blocks useful tools. If your starting point is the same question for ChatGPT, read how to block personal ChatGPT while allowing the corporate tenant; the Perplexity version of that problem is below.

Every claim about Perplexity here comes from Perplexity's published help center and pricing pages as of early November 2026. Where third-party sources disagree, we say so and tell you what to verify.

What Perplexity's enterprise plans actually promise

Perplexity sells two business tiers, Enterprise Pro and Enterprise Max, on top of its consumer Free, Pro and Max plans. The help center is explicit about the data commitments for Enterprise: query information is never used to train Perplexity's models, uploaded files are retained for only seven days, organizations above a seat threshold (50 or more Enterprise Pro seats, or a single Enterprise Max seat) can set custom retention periods and force deletion at any time, and Perplexity holds zero-data-retention and zero-data-training agreements with the model providers it routes to, including OpenAI and Anthropic.

Those are good terms, and they are the reason an Enterprise seat is defensible to a security review. Two caveats. First, Perplexity's own page describes the zero-retention commitment as an agreement between Perplexity and its model providers, not a setting your admin controls, and third-party reviews note that zero retention on the API is available only under a written agreement. Get it in your contract. Second, the governance features that make enterprise AI auditable (SCIM provisioning, audit logs, usage dashboards, retention configuration) are gated. Perplexity's pricing page has tied them to the 50-seat or Enterprise Max threshold, and a packaging change reported in August 2026 moves several of them behind a signed contract. Published Enterprise Pro pricing has sat around $40 per seat per month; check the current page before quoting it.

The personal account problem, in one paragraph

Here is the sentence from Perplexity's help center that should drive your policy: for Free, Pro and Max users, "AI Data Retention is enabled by default," and users must opt out individually in their account settings, with opt-outs applying only to data collected after the opt-out date. In other words, an employee who signs into Perplexity with a personal Gmail account and pastes a customer contract has, by default, contributed that contract to model training. The employee next to them on an Enterprise seat has not. Nothing on the network distinguishes the two sessions. Both go to perplexity.ai over TLS.

This is the same shape as the ChatGPT and Claude problem we have written about for two years, and the same mistake gets made: a URL filter sees the domain, a DNS filter sees even less, and the policy becomes a blunt allow or block. Blocking loses the productivity. Allowing loses the data. The only precise control is one that can see the account, and that requires inspecting inside the TLS session on the device where the user is signed in.

Comet makes it an agent problem too

Perplexity's Comet browser turns the question from "what did the user paste" into "what did the agent do." An agentic browser can read the pages a user has open, fill forms, and carry out multi-step tasks, which means it can move data out of a SaaS app without a human ever clicking upload. Our guide to governing ChatGPT Atlas, Comet and agentic browsers covers the browser-specific risks in detail. The short version for Perplexity: if you allow Comet, you have allowed a second browser with its own identity model, and your controls need to work at the process level, not the tab level.

How to govern Perplexity at work: four layers

The approach that works is the same three-layer AI governance stack we apply to ChatGPT and Claude, plus DLP on the content itself. Here is how it maps to Perplexity on the dope.security platform.

• Discover. AI Analytics in dope.console lists every AI app seen on your endpoints, including Perplexity, with per-user transactions and data volume and a breakdown of personal versus enterprise-licensed accounts. Ask the Dopamine agent "is anyone using Perplexity with a personal account?" and get the answer from live console data, not a survey.
• Decide. SWG policy sets Perplexity to allow, warn or block per group. A warn page that says "sign in with your company Perplexity account" fixes most personal-account use without a ticket.
• Restrict. Cloud Application Control locks ChatGPT and Claude to your enterprise tenants and blocks personal logins on the same domain. For Perplexity, use SWG policy to allow the app only for the groups that hold Enterprise seats, and block it for everyone else, so an unlicensed personal account never becomes the default path.
• Inspect. Dopamine DLP reads the prompt and any attached file on the device, classifies it by meaning (PII, PCI, PHI, source code, intellectual property) and blocks the sensitive ones before they leave the laptop. Perplexity is one of the AI tools Dopamine DLP covers, alongside ChatGPT, Claude, Copilot and Abacus. See our comparison of AI DLP tools for ChatGPT and Claude for how that classification differs from regex.

Perplexity's native controls vs endpoint governance: the head-to-head

Perplexity's enterprise controls and an endpoint governance layer are not competitors. They answer different questions, and a complete policy needs both.

• Training on your data: Perplexity Enterprise never trains on your data, by policy; consumer accounts retain data for training by default. dope.security cannot change Perplexity's policy, but it can see which policy each employee is under and block the wrong one.
• Retention: Perplexity Enterprise keeps uploads seven days and lets qualifying orgs set custom retention. dope.security stops the sensitive upload before there is anything to retain.
• Who is using it: Perplexity's usage dashboards and audit logs cover your Enterprise seats and are gated by seat count or contract. dope.security's AI Analytics covers every Perplexity session from every managed device, licensed or not.
• Personal accounts: Perplexity's admin console has no visibility into personal accounts that employees create on their own. dope.security discovers them and lets you warn or block.
• Content inspection: Perplexity's controls govern what Perplexity does with data after it arrives. Dopamine DLP governs whether it arrives, with a Dopamine Summary that explains in plain language why a prompt was blocked.
• Comet and desktop clients: Perplexity's controls apply inside Perplexity. dope.security's agent inspects every process on the endpoint, so Comet, the desktop app and the browser tab all meet the same policy.

How Perplexity's terms compare with ChatGPT and Claude

Security teams rarely evaluate Perplexity alone; it lands on the desk next to ChatGPT Enterprise and Claude Enterprise. The good news is that the enterprise tiers of all three now make the same core promise: business data is not used to train models, and admins get SSO, retention controls and audit visibility at some tier. The differences are in the details that a procurement review has to chase. OpenAI and Anthropic expose tenant identifiers (a ChatGPT Enterprise workspace ID, a Claude Enterprise organization ID) that a security control can key on, which is what lets Cloud Application Control allow the corporate tenant and refuse every other login on the same domain. Perplexity's enterprise controls are account-based inside Perplexity's own console, and its published materials do not describe an equivalent header or identifier for network-level tenant enforcement. That is why the practical Perplexity control is group-based allow plus DLP on content, rather than a tenant lock.

The consumer tiers are where the three diverge most. ChatGPT's and Perplexity's consumer tiers use conversations for model improvement unless the user opts out, Claude's consumer tier asks the user to choose, and all three put that choice in a personal settings page your IT team cannot see or enforce. Perplexity's help center is unusually direct about it, which makes the policy conversation easier: you can point at the sentence. The lesson generalizes. Whatever AI tool is next, the enterprise seat will be fine and the personal account will be the risk, and the only place to tell them apart is the endpoint.

What to write into the acceptable use policy

Keep the Perplexity clause short enough to enforce. Name the approved account ("the company Perplexity Enterprise workspace"), state that personal accounts are blocked on managed devices, list the data classes that must never go into any AI tool regardless of account (customer PII, PHI, card data, source code, unreleased financials), and say that uploads and prompts are inspected by Dopamine DLP and may be blocked. Then make the policy true: the warn page, the group-based allow and the DLP rule are what turn a paragraph into a control. Our free AI acceptable use policy template has the full language and the enforcement mapping.

A policy you can ship this week

Start by looking, not blocking. Run AI Analytics for seven days and count Perplexity users and personal accounts. Then buy Enterprise seats for the people who use it heavily, set SWG policy to warn everyone else toward the enterprise account or a sanctioned alternative, and turn Dopamine DLP on in Block mode for PII, PCI, PHI and source code. Write the rule into your AI acceptable use policy in one sentence: "Use Perplexity only through the company account; personal accounts are blocked." We have seen this exact sequence work for Claude, and our Claude enterprise controls guide is the template. A mid-market energy company used it to give its AI committee the one answer its DNS filter could never produce: who is using which AI tool, with which account, and what left.

Safe on the right account, visible on every account

Perplexity's enterprise terms are strong enough to approve. Its consumer defaults are loose enough to worry about. The gap between the two is a signed-in account on a laptop, and that is where dope.security enforces. Discover every Perplexity session, steer users to the enterprise account, and stop sensitive data at the prompt, all from one agent and one console. Start a free trial or see how dope.security governs AI at work.

Frequently Asked Questions

Does Perplexity train on enterprise data?

No. Perplexity's help center states that Enterprise data is never used or retained for training its models, and that it holds zero-data-retention and zero-data-training agreements with the third-party model providers it routes to. Consumer Free, Pro and Max accounts are different: AI data retention for training is enabled by default and must be switched off per user. dope.security's AI Analytics shows which account type each employee is using so you can enforce the enterprise one.

How long does Perplexity Enterprise keep uploaded files?

Perplexity's help center says uploaded files on Enterprise Pro and Enterprise Max are retained for only seven days, and organizations with 50 or more Enterprise Pro seats or at least one Enterprise Max seat can set custom retention periods and force deletion at any time. Dopamine DLP from dope.security prevents sensitive files from being uploaded in the first place, so retention never becomes the question.

Can I block personal Perplexity accounts but allow the company account?

Perplexity's admin console cannot see or control personal accounts employees create themselves. dope.security discovers Perplexity use per user with a personal-versus-enterprise breakdown, lets you allow Perplexity only for groups that hold Enterprise seats and warn or block everyone else, and inspects every prompt with Dopamine DLP regardless of account. For ChatGPT and Claude, Cloud Application Control additionally restricts logins to your enterprise tenant on the same domain.

Is Perplexity Comet safe for work?

Comet is an agentic browser, so it can read open pages and act on them, which means data can leave a SaaS app without a user clicking upload. Treat it as a second browser with its own identity and apply controls at the process level. dope.security's on-device agent inspects Comet's traffic the same way it inspects Chrome or the Perplexity desktop app, so DLP and app policy apply no matter which client the user picks.

How much does Perplexity Enterprise cost?

Published Enterprise Pro pricing has been around $40 per seat per month, with Enterprise Max priced substantially higher, but Perplexity's packaging changed in 2026 and several governance features (SCIM, audit logs, retention configuration) are gated by seat count or a signed contract. Check Perplexity's current pricing page before budgeting. Whatever you pay Perplexity, dope.security's visibility and DLP apply to every seat and every unlicensed personal account at one per-user price.

Does Dopamine DLP work with Perplexity?

Yes. Dopamine DLP coverage spans ChatGPT, Claude, Perplexity, Abacus and Copilot. It extracts text from prompts and attachments on the device, classifies the content by meaning rather than by regex pattern, and can block, monitor or stay off per policy. Classification runs through an LLM with zero data retention, and the Dopamine Summary explains each detection in plain language.