Data access control for AI agents, user by user

Your agents pull from the same documents, tools and databases for every user. Protecto checks who is asking and masks what that user should not see.

SOC 2 Type II · ISO 27001 · HIPAA, BAA available

Trusted by regulated enterprises & agentic platforms
Inovalon
Automation Anywhere
Citrix
Bank Of Muscat Logo
Dell
Case study result

Automation Anywhere protects customer data in its AI agents

5,000+

Live enterprise tenants

Isolated

Tenants, each with its own policy and retention

Logged

Every reveal of original values

Why source permissions stop working inside AI agents

A junior analyst asks the company’s AI agent a routine question about team budgets. The agent searches the shared drive, pulls a paragraph from a compensation file and puts it in the answer. The analyst could never open that file.

The permission sat on the file, and it did not travel with the text. Once a document is chunked and retrieved, the source system’s rule no longer applies.

The Sensitive Data Risk Map for Agentic AI shows where your access model breaks, layer by layer. 11 pages.

One agent serves every role.

Agents get broad access so they can answer a wide range of questions. Unless something passes in the user’s context, they run the same way for everyone.

Hard-coded checks do not last.

Role checks written into each agent take time to build, and every policy change means editing every agent again.

Nobody can show who saw what.

Without a record at the agent layer, you cannot tell an auditor what an agent retrieved or revealed for a given user.

One question, a different view for each user

The model receives tokens every time. Protecto then restores only what the person asking is allowed to see, and logs it.

Same agent, same questionIllustrative

The question

Summarize the complaint on case 4471.

Support contractorNothing revealed

[PERSON_7a1] reports a duplicate charge on [ACCT_c42]. [LEGAL_NOTE]

Billing specialistName, account

Dana Reyes reports a duplicate charge on AC-48213. [LEGAL_NOTE]

Legal counselAll revealed

Dana Reyes reports a duplicate charge on AC-48213. Settlement offered.

RevealedMasked by policy
Value levelNot the whole record
Every requestChecked at runtime
LoggedEach reveal

You write the rules. Protecto enforces them.

Your application passes this context to Protecto with an auth key. Protecto checks it before data moves.

Who is asking

The user’s identity and role, as your application knows them.

Which agent

Two agents working for the same user can get different access.

What task

Drafting a refund reply needs less data than an audit review, so each task gets its own scope.

What data

The entity types, context categories and documents this interaction may use.

Access checks at each step of an agent run

Your application or gateway calls Protecto before a model call, on retrieval, around tool calls and before the answer returns.

Real timeChecks retrieval, tool calls, agent memory and answers as data moves.
Architecture: retrieved documents, database records, tool results and agent messages pass through Protecto, which runs on-premises or in your private cloud. Protecto checks the user, agent, task and policy. Agent memory holds tokens, users without access stay masked, approved tools and authorized users get real values by policy, AI models see tokens only, and each decision is logged.On-premises or your private cloudModel providersWhat an agent readsRetrieved documentsRAG, SharePoint, wikisDatabase recordsCRM, warehousesTool resultsAPIs, MCP serversAgent messagesHandoffs, shared contextProtectoOne control point1DetectSensitive values in free text2Check accessUser, agent, task, policy3TokenizeConsistent, format-preserving4LogEvery mask and every revealToken mappings and keys stay with youAccess logWhat was masked and what was revealed.How each policy was enforced.Agent memoryTokensUsers without accessStays maskedApproved toolsReal ID, loggedAuthorized usersUnmask by roleTokens onlyAI modelsAny providerGoogleOpenAIAnthropicSees tokens onlyAnswers come back with tokens.Protecto restores real values only forthe people and tools your policy names.Dana Reyes[PERSON_7a1]AC-48213[ACCT_c42]AC-48213[PERSON_7a1]Architecture: retrieved documents, database records, tool results and agent messages pass through Protecto, which runs on-premises or in your private cloud. Protecto checks the user, agent, task and policy. Agent memory holds tokens, users without access stay masked, approved tools and authorized users get real values by policy, AI models see tokens only, and each decision is logged.On-premisesor your private cloudRetrieved docsDatabaserecordsTool resultsAgent messagesProtecto1DetectSensitive values in free text2Check accessUser, agent, task, policy3TokenizeConsistent, format-preserving4LogEvery mask and every revealAgent memoryTokensOther usersStays maskedApproved toolsReal ID, loggedAuthorized usersUnmask by roleAccess logWhat was masked and revealed,how each policy was enforced.Tokens onlyModel providersAI modelsGoogleOpenAIAnthropicSees tokens only. Answers come backwith tokens, and Protecto restores real values.
Real valuesTokensAnswers coming back

On-premises · Private cloud · Air-gapped · SaaS. When Protecto runs in your environment, token mappings, keys and logs stay on your servers.

Stop writing access checks into every agent

Hard-coded checks can work for one agent. They get harder to keep right as agents, roles and policies multiply.

What mattersChecks inside each agentProtecto
Where the logic livesIn each agent’s codeIn one layer every agent calls
When a rule changesEdit and retest each agentUpdate the policy context your application passes. Agent code stays the same
Data a user may not seeOften removed, so answers lose contextConsistent tokens, so the model can still follow the text
Data with no patternCustom logic for each caseContext-Based Access Control (CBAC) applies your policy to passages
Evidence for auditsLogging you build and maintainLogs of masking, unmasking, filtering and retrieval

A SaaS company cut cost by 90% compared with building this in-house.

Control each user’s view, and prove it

Access rules only help if the agent keeps working and you can show what happened.

01

Masked by default, revealed by role

Sensitive values stay tokens, and real values appear only for the people and tools your policy names.

02

Retrieval that respects permissions

GPTGuard MCP returns only the document chunks this user, agent and task may see.

03

Business context with no pattern

CBAC applies your policy to discount terms, legal notes and HR context, where pattern rules have nothing to match.

04

A record of each decision

Logs cover masking, unmasking, filtering and retrieval, and reports show how policies were enforced.

“Protecto enabled us to deliver data security, privacy, and compliance for our AI features at scale across 5,000+ enterprise clients, strengthening customer trust in our AI automation solutions.”

SS
Steve Shah
SVP Product, Automation Anywhere
Security & compliance

Built for regulated data at enterprise scale.

Protecto combines data protection, audit logs, and ready-to-use policies for regulated data, whether deployed as SaaS, on-prem, in a private cloud, or air-gapped.

SOC 2 Type II

Audited controls

ISO 27001

Certified ISMS

HIPAA

BAA available

GDPR

EU data ready

DPDP

India privacy law

CCPA / CPRA

US privacy
Independently verified to outperform AWS Comprehend and Microsoft Presidio · audit logs on every scan, mask, and unmask.

Go deeper on data access for AI agents

Data Protection Architecture For Agentic Ai Whitepaper Cover
Whitepaper · 14 pages

Data Protection Architecture for Agentic AI

A reference architecture for where data controls sit in agentic AI, from prompts to tool calls.

Questions about data access control for AI agents

What is data access control for AI agents?

It decides what an agent can use and reveal for each request, based on the user, the agent, the task and your policy. A user without access sees masked values, and the agent keeps working.

When it retrieves documents or database rows, when a tool returns more fields than the user needs, when it stores data in memory across turns, and in the final answer. Data leak prevention for AI maps each path.

Protecto is API-first. You call it from LangChain, LlamaIndex or your own code, with any model provider, at the steps in your agent run that handle sensitive data.

No. Those tools decide who a user or agent is and what systems it can reach. Protecto works on the data itself, using the identity and policy context your application passes in.

Yes, through Privacy Gateway. You keep each user’s masking and unmasking access in AD or Okta groups, and Gateway issues a Protecto auth key with that access at login. If you use Protecto Vault directly, it’s a headless API, so your team can build the integration around it.

It works where the data is stored. Once data is chunked, embedded, retrieved into a request or passed between agents, those rules often no longer apply.

It can, when placeholders carry no meaning. Protecto uses consistent, format-preserving tokens, so the model can still follow the text.

You do. Your application, agent framework or orchestration layer decides what each user, agent and task may see. Protecto enforces those decisions at runtime.

Yes. Protecto reveals real values to the people and tools your policy names, and logs it. Everyone else keeps seeing tokens. Protecto Gateway does this for authorized tool calls automatically. Without the Gateway, your application handles that step using Protecto’s APIs.

Yes. Context-Based Access Control (CBAC) applies your policy to passages such as discount terms, legal strategy and HR context.

On-premises or in your private cloud, air-gapped if you need it. A SaaS option is also available. When Protecto runs in your environment, token mappings, keys and logs stay on your servers, and external models only receive tokens.

Teams use Protecto to support GDPR, HIPAA and CCPA programs. It masks sensitive data before it moves and produces reports on how your data policies were enforced.

See what your agents reveal to each user

Bring one agent workflow, with test data if you prefer. We will show what each role receives and what gets logged.

Cover of the Protecto Privacy Vault datasheet
Datasheet

Protecto Vault: sensitive data protection for AI workflows

Features that safeguard your data across agents, AI applications, RAG pipelines, MCPs, data lakes, and enterprise systems.

By submitting you agree to Protecto’s privacy policy.

Cover of the whitepaper The Sensitive Data Risk Map for Agentic AI
Whitepaper

The Sensitive Data Risk Map for Agentic AI

A framework for finding where sensitive data moves through your AI systems, before a security review or an auditor finds it for you.

By submitting you agree to Protecto’s privacy policy.

Data Protection Architecture for Agentic AI whitepaper cover
Technical whitepaper

Data Protection Architecture for Agentic AI

A reference architecture for protecting sensitive data across agent workflows, for architects and AI platform engineers.

By submitting you agree to Protecto’s privacy policy.

New: run any AI model without exposing real values.
See how it works