How Protecto works, layer by layer

Protecto classifies, tokenizes and controls sensitive data as it moves through your AI systems. Your models keep the context they need. Your real values stay protected.

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

Bank Muscat masked Arabic and English data so its teams could use Gemini

99%

Recall on sensitive data, incl. Arabic

96%

Precision on the same data

85%

Semantic accuracy retained after masking

Why data moves differently in AI

A company connects an internal assistant to its shared drive, HR files and CRM. An analyst asks a routine question. The assistant pulls a passage from a compensation document the analyst could never open, adds it to the prompt and answers.

Traditional controls were built around boundaries: tables, files and services. AI assembles context on the fly from databases, PDFs, knowledge bases, SaaS systems and external tools. Once data joins that context, those boundaries disappear. Sensitive data can sit anywhere in the prompt, the retrieved documents or the response.

Protecto sits between your enterprise data and your AI systems and controls how sensitive data moves through them at that moment.

One prompt, four sources

context window4 sensitive spans, 0 boundaries

Synthetic example.

The Protecto architecture

Protecto is a layered platform. Each layer has one job: analyze the data, understand its context or apply your policy during AI workflows.

Protect the values

Find sensitive data, replace it with tokens and keep the originals apart. Layers 1 to 3 →

Decide who sees what

Connect identity, read the context and make an access decision on each request. Layers 4 to 6 →

Run it at scale

Protect RAG pipelines and run on infrastructure built for high-volume AI work. Layers 7 and 8 →

Layer 1

Agentic data classification

Sensitive data rarely sits in clean fields. It appears in documents, PDFs, emails, scanned records and conversational text.

DeepSight, a core feature of Protecto Vault, reads the meaning around a value and goes beyond pattern matching, so it finds sensitive entities in structured and unstructured content. Protecto identifies:

  • Personally identifiable information (PII)
  • Protected health information (PHI)
  • Financial and payment card data (PCI)
  • Confidential business information, through CBAC in layer 6

The result is a persistent, structured record of where sensitive data sits in your content.

Independent benchmark: UT Dallas and DataXpert measured a higher F1 for Protecto than for AWS Comprehend and Microsoft Presidio.

support_ticket.txt
pls refnd jhon smtih, card 43 21
Bitte rufen Sie Anna Keller back about the claim
contact: tom.reyes@example.com
The patient is Tom Reyes.
Last week he was told he has type 2 diabetes.
Synthetic example

Why this is hard

Sensitive meaning rarely fits one clean field. It spreads across sentences, mixes languages and arrives with typos and inconsistent formatting. Pattern detection and basic NLP miss these cases.

Layer 2

Context-preserving tokenization your AI can still read

Conventional masking swaps values for random characters or generic placeholders. That works for storage. It breaks AI, because the structure language models depend on is gone: reasoning degrades, retrieval fails and responses stop making sense.

Protecto replaces each sensitive value with a structured token that keeps its type, its position and its links to every other mention, across documents and datasets at high throughput.

Summarization, extraction and question answering keep working, and the model only receives tokens.

used a credit card ending in .

has requested a refund.

Send it back to card .

Is the cardholder the person asking?

Why this is hard

Traditional masking breaks AI reasoning because the links between entities disappear. Tokens must keep semantic structure and referential integrity across whole documents while still protecting the values. Get this wrong and model output degrades.

Layer 3

Vault infrastructure

Each token maps to its original value inside Protecto Vault. Your AI processing stays strictly separate from your sensitive data storage.

AI processing path

PromptGHJ6M7
Model callGHJ6M7
Response logGHJ6M7
Prompt cacheGHJ6M7

Protecto Vault

GHJ6M7enc:9f2c…a71e
HWE12Kenc:04be…c3d0
Q19K02enc:7ad1…08f4
Encrypted at restkeys managed

Encrypted storage

Sensitive values are encrypted at rest, with key management.

Complete audit logs

Every token access event is logged for compliance, forensics and traceability.

Processing isolation

AI systems work on tokenized data. Real values stay out of the inference path.

Policy-controlled restoration

Access policies govern the token-to-value mapping. Only authorized workflows can restore original values.

If a model, prompt or response is logged, intercepted or cached, that copy holds tokens instead of the original regulated data.

Why this is hard

Tokens must be reversible for authorized workflows and fully isolated from AI pipelines, and Protecto Vault has to hold up at large data volumes.

Layer 4

Access control and identity integration

A model with access to a knowledge base can combine several sources in one response. That response may hold data the person asking should not see, even though the model was authorized to read the source.

Role-based access control (RBAC) governs systems and datasets: a user can open a table or can’t. Protecto adds a layer that applies to the information inside AI responses as well as the systems behind them.

  • Gets each user’s role from your application, so Protecto knows who is asking and what permissions they hold
  • Takes policy context through an auth key
  • Works with your identity tools. Protecto does not replace identity or agent identity tools. The Protecto Vault API is headless, so you can build it into your identity systems, and Privacy Gateway works with AD and Okta

Why this is hard

AI responses can combine several data sources, so access control has to reach the information inside each response as well as the datasets. RBAC was never designed for this.

Layer 5

Understanding the data and its context

Finding sensitive data is only the first step. The same data can be fine in one context and restricted in another, and discount details or intellectual property often spread across a document.

  1. Signal 1The prompt being submitted
  2. Signal 2The data sources involved
  3. Signal 3The identity of the user
  4. Signal 4Your organization’s policies

UnmaskedTokens onlySynthetic policy

Together these signals set the exposure risk of each request. A request to summarize a document can be fine for one user and restricted for another, so Protecto checks each interaction and gives each user the right level of visibility.

Why this is hard

The right decision depends on the prompt, the data sources, the user and your policy at once. That check has to run during each AI interaction, because a static rule table can’t cover it.

Layer 6

Context-Based Access Control (CBAC)

Two people ask the same assistant about the same account. One should see the discount terms. The other should get the same answer with those terms kept masked.

CBAC is the runtime enforcement layer, where classification, context, identity and policy meet in one access decision. You define the policy. Protecto enforces it inline during AI inference, with no batch review and no manual approval queue.

Why this is hard

Decisions happen in real time during AI inference, while identity, intent, data sensitivity and policy rules are all evaluated. The controls also have to adapt to new context, data and requests.

Same question, try each roleWhat discount did Acme get on renewal?

Synthetic example. You define the policy.

Layer 7

Secure retrieval-augmented generation (RAG)

Many enterprise AI applications answer questions from internal knowledge bases. This is where risk concentrates: retrieval pulls documents from storage, and generation shows their contents to users.

GPTGuard, Protecto’s secure RAG layer, protects sensitive data across this pipeline, so you can build assistants on internal knowledge without exposing the confidential information inside documents.

Why this is hard

Retrieval can pull documents with mixed sensitivity. Protecto has to protect the sensitive segments inside them without hurting retrieval accuracy or model reasoning.

Layer 8

Enterprise platform infrastructure

Protecto supports high-volume AI workloads and keeps the same security guarantees at enterprise scale.

Deploy as
Audit logYour appsreal valuesProtectoToken map, keysAI models

Protecto-hostedReal valuesTokens

Multi-tenant architecture

Strict data isolation between tenants, with policy and retention set per tenant.

Real-time APIs

Sync APIs for low-latency inline processing during AI workflows.

Async processing pipelines

Batch classification and tokenization to prepare large content stores. High-volume masking →

Compliance and audit

Logs for every classification, tokenization and policy decision, for regulatory traceability.

One SaaS company processes 13M+ long texts a day on Protecto. Automation Anywhere runs Protecto across 5,000+ live enterprise tenants.

Why this is hard

Classification, tokenization and policy decisions have to run across large document sets and live AI interactions at once, with low latency, strict tenant isolation and the same guarantees in every deployment model.

“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.

See Protecto on your data

Your AI keeps the context it needs, and your sensitive data stays protected. Bring a sample, with test data if you prefer, and we will walk through each layer.

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