Protecto DeepSight
Sensitive data detection that reads like a human reviewer
DeepSight, Protecto’s agentic data classifier, finds PII, PHI and PCI in typos, mixed languages and free text, then returns tokens your AI can still use.
99.74% F1 score in Protecto’s DeepSight datasheet (internal test, April 2026).
Why sensitive data detection misses what AI reads
A support agent logs a call. The NER model we built in our first two years checks the note before it reaches an AI model.
What the AI model receives
[PERSON] called about his claim #4456. He mentioned his SSN ends in 7821 and wants a callback at [PHONE].
Two sensitive values would have reached the model.
A missed value is gone.
It reaches the model, the logs and the search index. No later step can recover data you never detected.
A false flag hurts answers too.
Mark “Tesla” as PII and the sentence loses its meaning. Answers get worse, and reviewers stop trusting the alerts.
Rules never catch up.
Each new format, language or entity means more rules and more retraining. Fixing one case tends to expose another.
Six inputs that pattern tools miss
AI systems read noisy, real-world text. DeepSight weighs the words around a value before it labels it.
Missing spaces
Charge to 4228231102032030and confirmA card number runs into the next word, so a pattern stops matching.
Mixed languages
اتصل بي على 467-802-3822 بعد الساعة 5 مساءًA phone number sits inside Arabic text, where English-tuned models lose it.
Context only
Forward the $72K package to HR. Sarah accepted the offer.No word says “salary”. Only the sentence shows this is pay data.
Lookalikes
Reference code: 4228 2311 0203 2030Sixteen digits, and not a card. An Aadhaar number isn’t a PAN either. Context settles the type.
Obfuscated values
rahul(dot)mehta(at)gamil(dot)comTypos and spelled-out symbols slip past exact matches.
Partial values
His SSN ends in 7821Four digits look harmless until they sit next to a name and a claim number.
Agentic data classification: five stages for every span
One model prediction isn’t enough for messy data. DeepSight works through each span in stages, so ambiguous values get reasoned about instead of guessed.
- 01 IngestTake the input as it is
Text, JSON or document chunks. DeepSight detects language, script and structure without changing the format.
- 02 ParseKeep each span in context
Candidate values are segmented with the words around them.
- 03 ReasonLabel by meaning
The agent labels each span by what it means, with pattern as one signal.
- 04 VerifyCheck before output
Each label is checked against policy, format rules and edge cases, and gets a confidence score.
- 05 ReturnSend back tokens
Reversible, format-preserving tokens, with audit metadata for every value.
Worked example: a German patient record
Benchmark case TC-45
Input
Patient: Klaus Wagner, geb. 14.09.1965. SSN: 789-01-2345. Adresse: Schillerstr. 7, Hamburg 20095. Tel: 040-9876-5432.
Detected: 7 of 7 entities
Returned to your AI
Patient: <PER>Xucvd eQuhGE</PER>, geb. <DOB>RDZn8</DOB>. SSN: <SSN>681-952-1079</SSN>. Adresse: <ADDRESS>JkZ5AIDrc5. 0Bj</ADDRESS>, <CITY>dcxB157</CITY> <PIN>30948</PIN>. Tel: <PHN>180-4494-4825</PHN>.
SaaS · VPC · On-premises · Air-gapped. One classifier, one API and one audit model in every mode. When DeepSight runs in your environment, your data stays there.
Scored two ways, including the stricter one AI needs
Standard scoring counts a value as protected once it is found and masked. AI needs more. A masked person, place or ID must keep its role in the sentence, so the type has to be right too.
Source: Protecto’s DeepSight datasheet (internal test, April 2026). 75 cases, including hard inputs in Arabic, French, German, Spanish and Japanese. Results by test suite are in the datasheet.
Standard scoring: found and masked
- 191 of 192 entities found
- 0 false positives
AI scoring: found and typed correctly
- 98.6% precision
- 90%+ response accuracy retained
How DeepSight compares with pattern and NER tools
Rules and NER models worked when data was clean and labeled. AI inputs are neither.
| Capability | Pattern and NER tools | Protecto DeepSight |
|---|---|---|
| Free text | Tuned for clean fields. Brittle on notes, tickets and chat. | Reads each span with the words around it. |
| Typos and obfuscation | Miss a value once the format breaks. | Handles typos, partial values and novel formats. |
| Lookalike values | A 16-digit reference code can read as a card number. | Resolves the type from surrounding context. |
| Context-only sensitivity | No keyword or entity to match, so nothing is flagged. | Weighs meaning, so “$72K package” reads as pay. |
| Mixed languages | Often tuned for one language at a time. | Officially tested in seven languages, with locale-specific IDs. |
| Custom entities | New types mean more rules or retraining. | Add claim numbers or internal codes in days, with no ML stack to run. |
| Model upkeep | You pick, tune and retrain the model. | The agentic layer picks the models for each task, behind one API. |
Independent benchmark: UT Dallas and DataXpert measured a higher F1 for Protecto than for AWS Comprehend and Microsoft Presidio.
Detection is where control starts
DeepSight names the data. The rest of Protecto acts on those names, value by value and role by role.
Tokens that keep AI working
Tokens keep format, type and consistency across records, so joins, embeddings and analytics still work.
AI chat and tool calls
Protecto Gateway reveals real values to authorized tools, so a lookup finds the record while the model only sees tokens.
Real values back by role
Context-Based Access Control (CBAC) decides who sees real values. Everyone else keeps seeing tokens.
A log for every label
Each classification is logged with its confidence, policy path and reversible token ID, so reviewers can check any decision.
“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.”
SVP Product, Automation Anywhere
Built for regulated data at enterprise scale.
SOC 2 Type II
ISO 27001
HIPAA
GDPR
DPDP
CCPA / CPRA
Go deeper.
Protect Sensitive Data Without Breaking AI Workflows
How to mask sensitive values for the model and still give an approved tool the real value it needs, with a worked tool-call example and a checklist for production.
Why we stopped building NER models
What broke our own NER models, and how agentic data classification replaced them.
Agentic data classification: a new architecture
How the agentic layer picks models by context, data type and policy, behind one API.
Why near-100% recall is the standard for AI data
Why a missed value costs more than a false flag, and how to read recall and precision.
Questions about DeepSight
What is Protecto DeepSight?
DeepSight is Protecto’s agentic data classifier. DeepSight finds sensitive data in free-form text, structured fields, chat logs, documents and tickets, then returns policy-aware tokens that are safe for LLMs, RAG and agent workflows.
Is DeepSight a separate product?
No. DeepSight is a core feature of Protecto Vault, not an individual product.
What is agentic data classification?
The classification is a structured workflow instead of one model prediction. The system examines the input, reasons about context, identifies sensitive values, applies policy and checks the output.
What data can DeepSight detect?
50+ entity types across PII, PHI and PCI, plus custom entities you define, such as claim numbers and internal codes. India’s PAN and Aadhaar numbers are built in.
Which languages does DeepSight support?
Detection is officially tested in seven languages: English, German, French, Spanish, Dutch, Arabic and Japanese. Other languages are available on request and tested on your data.
How accurate is DeepSight?
In Protecto’s internal test from April 2026, DeepSight scored 99.74% F1 with 0 false positives across 75 cases. Under stricter scoring, where the type must also be right, recall was 98.8%. DeepSight also retained 90%+ response accuracy, a measure that compares AI answers before and after masking. Source: Protecto’s DeepSight datasheet.
Does my data leave my environment?
It doesn’t have to. DeepSight can run in your VPC, on-premises or air-gapped, with the same classifier and API as the SaaS option. The classifier runs there too, and external models receive only tokens.
How do we integrate DeepSight?
Through a REST API for real-time, batch and streaming work: POST /vault/scan, /vault/mask and /vault/unmask.
Does DeepSight replace DLP?
No. DLP keeps guarding email, endpoints and file transfers. Protecto covers where it falls short: AI traffic, context and unstructured data moving to models and agents.
See DeepSight on your data
Bring a sample document or ticket, with test data if you prefer. We will show you what DeepSight finds and what your AI receives.
- SOC 2 Type II certified
- ISO 27001
- HIPAA compliant, BAA available
- On-premises or private cloud
