Pragatix AI FAQ Page
Pragatix: Expanded FAQs
EXZOS is dedicated to GDPR compliance and transparent data handling, prioritizing user privacy and trust. Unlike Microsoft, which is criticized for its lack of transparency and monopoly abuse, NEXZOS offers a fair, user-focused alternative. Positioned as both a technology platform and a trusted partner, NEXZOS aims to create a more ethical and transparent digital landscape. If you’d like more details, you can explore the overview or additional linked articles.
Table of Contents
Table of Contents
Data Brokers and European Digital Legislation
Architecture (1–18)
1. What is the overall architecture of Pragatix?
Pragatix is an end-to-end AI security & enablement platform composed of an AI Security Suite (Prompt Guardian, Model Guardian, Guardian Agent, Risk Engine), a Private AI Suite (RAG Knowledge Chatbot, Knowledge Assistant,multi-model support), and an AI Behaviour Suite (Behaviour Intelligence,Security Awareness). It supports SaaS (AWS), single-tenant VPC, and onprem/air-gapped deployments.
2. What are the primary runtime components and how do they interact?
Key runtime components: Prompt Guardian (real-time proxy enforcement), Guardian Agent (runtime governance for agents), Model Guardian (model vetting & scanning), Risk/Policy Engine, Proxy/Browser extension or Service API topologies. All AI interactions flow through these layers for inspection, classification, policy enforcement and auditing.
3. What topologies are supported for traffic capture and enforcement?
Browser extension (Chrome/Edge) for user-level coverage, network proxy for broad traffic coverage (browsers and native apps), and Service API topology for programmatic integration with existing AI services (e.g., Copilot integration)
4. How does Pragatix handle multi-model routing and model hosting?
Multi-model support: routing to cloud-hosted models (OpenAI, Google Gemini, Anthropic Claude, AI21, Mistral), AWS Bedrock, and self/privately-hosted models. The AI Gateway routes prompts according to policies and tenant configuration.
5. How is the Guardian Agent positioned in the architecture?
Guardian Agent secures and monitors interactions between AI agents and connected systems. It enforces runtime policies, inspects prompts/responses for sensitive data and toxic content, and controls which connectors/tools an agent may use
6. What is the role of the Prompt Guardian?
Prompt Guardian is the real-time proxy that inspects prompts/responses, applies DLP & sensitivity classification, blocks or tokenizes sensitive content, detects shadow AI, and defends against prompt injections.
7. What does Model Guardian do?
Model Guardian performs multi-pillar scans (provenance & identity, static analysis, behavioural vetting, legal/compliance). It checks for backdoors, dependency CVEs, license validation, hash/GPG verification, and provenance checks.
8. How is intent and behavior analysis implemented at runtime?
Behavioral baselines are built per-user and per-agent at runtime in the customer tenant to detect anomalies; combined with NLP classifiers and deterministic rules for richer intent and behavior understanding.
9. How are policies evaluated and enforced?
A hybrid risk & policy engine combines deterministic rules (RegEx, allow/block lists, identity/group conditions) with NLP semantic classification. Policies are enforced at the proxy/agent runtime and recorded in the audit log
10. Where is telemetry and audit data stored?
For SaaS, data is stored on AWS in configurable regions (default eu-central-1 for EU customers). Customers can configure retention windows per data category; on-prem deployments keep data within customer infrastructure.
11. How are connectors and external servers controlled?
Administrators define per-agent scope: which agents may access which connectors/tools. Guardian Agent enforces these access controls and can block/allow connections based on prompt inspection.
12. How are prompt masking and tokenization implemented?
Sensitive tokens (names, IDs, account numbers) can be replaced with surrogates before reaching upstream LLMs; restored client-side when allowed. Blocking, redaction or tokenization is policy-driven.
13. What is the interaction flow when an AI agent attempts a risky action?
The action is intercepted by Guardian Agent/Prompt Guardian, evaluated against policies, possibly blocked or routed to human-in-the-loop approval; every step is logged immutably
14. How does Pragatix integrate with upstream LLM providers securely?
Pragatix configures enterprise/API tiers for providers, invoking zero-dataretention modes where supported, and routes traffic through the AI Gateway with policy enforcement.
15. How are change/versioning and configuration managed?
Detection rules and classifier rulesets are versioned, tested and released in scheduled cycles. Customer-defined rules are versioned and audited separately; on-prem instances can pin model versions and updates require client consent.
16. How is identity and access enforced across the architecture?
Identity via OIDC/OAuth2 or SAML 2.0 integrations (e.g., Entra ID, Google Identity). RBAC scopes admin and analyst roles to tenant-limited views.
17. How is network-level enforcement achieved?
Through a proxy firewall that captures browser and app traffic, combined with browser extensions for client-side visibility and enforcement.
Security & Compliance (19–38)
19. How does Pragatix perform DLP and sensitivity detection?
Hybrid stack: deterministic RegEx for structured identifiers, NLP classifiers for contextual categories (financial, health, legal, trade secrets), and behavioral baselines. OCR is applied to files/images/PDFs for content inspection.
20. Are the DLP/NLP models proprietary and how are they trained?
The NLP layer uses standard pre-trained foundation models with Pragatix’s prompt-engineered classification logic and in-house rule taxonomies. Labels/evaluation data are synthetic or public and labeled in-house; customer data is not used to train these models.
21. Does AGAT use customer data to train or fine-tune models?
No. By default, customer prompts, responses, audit logs and telemetry are not used to train or fine-tune Pragatix models. This can be contractually guaranteed in the DPA.
22. How does Pragatix support GDPR data subject rights?
Access: filterable audit exports by user. Erasure: configurable retention windows and targeted record deletion via Admin console. Portability: structured exports to SIEM. Restriction/Objection: policy flags can suspend processing for identities.
23. Is a DPA available and is Pragatix GDPR-aligned?
Yes. Pragatix is built on GDPR principles and an Article 28-aligned DPA is available during onboarding.
24. How does Pragatix help customers meet EU AI Act obligations?
Pragatix provides logging, transparency (user notifications when actions are blocked), human oversight controls (overrides, approvals), post-market monitoring inputs, and technical documentation to support deployer conformity assessments when required.
25. Does Pragatix itself fall under Annex III high-risk categories of the AI Act?
As shipped and recommended, Pragatix is a governance layer and does not make autonomous consequential worker-management decisions; therefore it does not fall under Annex III high-risk categories. Customer configurations that use it to make worker-management decisions may trigger high-risk obligations for the deployer.
26. How are breaches and incident notifications handled (GDPR Article 33)?
AGAT commits to notifying customers without undue delay to support customer 72-hour notification requirements; initial notification within 24 hours of confirmed breach detection is part of contractual commitments in the DPA
27. What audit and compliance reporting capabilities exist for regulators (DORA, NIS2, AI Act)?
The Activity Auditing console captures who/what/when/where and policy decisions per interaction. Reports exportable to formats suitable for GDPR, EU AI Act logs, DORA (ICT third-party evidence), and NIS2 incident reporting. Exports to SIEMs are supported.
28. Does Pragatix provide DPIA support for high-risk processing?
Yes. AGAT provides technical/organizational descriptions, data flow diagrams, retention info and mitigation controls to help the controller perform DPIAs. Features such as anonymization, scoped dashboards and granular RBAC support proportional deployments.
29. What certifications and third-party audits does AGAT hold?
AGAT is SOC 2 audited and aligned with ISO 27017 and ISO 27018. Annual external penetration testing and vulnerability scanning are performed; summary reports and SOC 2 attestations are available under NDA.
30. Are penetration test reports and independent audits available to customers?
Yes — a summary letter and remediation report from recent tests (latest Jan 2026) and audit artefacts are available under NDA.
31. How is multi-tenancy data isolation achieved in SaaS?
Logical isolation with per-tenant data stores, separate encryption keys, tenant identifiers enforced on every request. Dedicated single-tenant VPC or onprem deployments available for stronger isolation.
32. What encryption and key management mechanisms are used?
AES-256 at rest and TLS 1.2+ in transit. SaaS uses AWS KMS; on-prem supports BYOK
33. How are third-party LLM provider commitments enforced contractually?
Upstream providers are configured on enterprise tiers with zero-dataretention options. Sub-processor schedules and contractual terms in the DPA bind providers; the list is available and can be tailored.
34. How does Pragatix handle privileged or legal communications?
Built-in legal/advice usage classification rules can detect privileged content and apply guardrails: block, redact, or tokenise before reaching external LLMs; allow only internal private assistants where required.
35. Does Pragatix support healthcare regulations (EHDS, NEN 7510)?
Pragatix supports EHDS-aligned processing patterns and ships healthcare classification rules. AGAT is aligned with ISO/27001 controls but does not currently hold formal NEN 7510 certification; on-prem/dedicated deployments are recommended for strict compliance.
36. How does Pragatix address financial sector regulations (DORA) and NIS2?
Provides ICT third-party inventories, immutable logs, incident evidence exports, operational resilience testing features, and contractual addenda (DORA clauses) upon request.
37. How are employee privacy and monitoring balanced with security?
The platform supports anonymization in logs, configurable retention, scopelimited dashboards and RBAC to enable proportional monitoring. These controls help meet works council/data-protection obligations.
38. Is transparency provided to end users when automated enforcement occurs?
Yes. Users receive in-context messages describing the policy that blocked/flagged the prompt and are offered in-context training/guidance. Admins have full policy chains and decision logs
Deployment & Operations (39–54)
39. What deployment models are available?
SaaS multi-tenant on AWS, SaaS single-tenant / dedicated VPC, onprem/private cloud, and air-gapped deployments
40.What are recommended deployment topologies by use-case?
Fast visibility / low friction: Browser extension. Broad coverage across apps: Network proxy. Programmatic governance (Copilot, custom AI systems): Service API topology. Highly regulated environments: On-prem or dedicated VPC.
41. How are updates and model versioning handled in different deployment models?
SaaS receives scheduled product releases (every few weeks); on-prem deployments use pinned model versions and receive updates only with client consent. AWS Bedrock model versioning controls and explicit model IDs available.
42. What is the typical onboarding process and timeline?
Pilot tenants are provisioned on AWS with IdP integration; baseline policy sets pre-loaded. Browser extension pilots can be stood up quickly; full enterprise rollouts depend on topology and integrations and are agreed during onboarding
43. How are customer-defined policies and rules managed operationally?
Admins create and edit rules in the console. Rule creation is LLM-assisted but requires admin review. Rules are versioned and audited; tenant-specific rules are not shared across customers.
44.What logging and monitoring integrations are supported?
Exports to SIEMs (Splunk, Azure Sentinel etc.), webhook integrations, and dashboards for activity auditing, policy triggers, and behavioural metrics.
45.How is rollback and change control performed for rules and classifiers?
46.How are exceptions and overrides handled operationally?
Authorized admins can perform real-time overrides, temporarily disable policies, or approve blocked high-risk actions via integrated approval workflows (Slack/Teams/webhooks). All overrides are logged.
47. How is high availability and resilience addressed?
SaaS on AWS and single-tenant VPC deployments leverage cloud availability patterns. On-prem deployments can be architected per customer HA/DR requirements. The platform supports export of logs for offline analysis.
48.What operational metrics and KPIs are available?
Shadow AI detection counts, top data classification triggers, policy violation rates, risky-user/department metrics, adoption and ROI indicators from
Behaviour Intelligence.
49.How are agent runtimes (AI agents) discovered and monitored?
Discovery via network traffic inspection, API monitoring and runtime agent gateways. Guardian Agent provides visibility into agent activity, connectors used and actions taken.
50.How are credentials and secrets handled for connectors?
Connectors and credentials are managed with secure storage patterns (encrypted at rest). Access is scoped via policies and RBAC; the Guardian
Agent mediates connector usage.
51. What support is provided for logging retention and purging?
Retention windows configurable per data category; logs are purged automatically at expiry. Support for targeted deletion for data subject requests is provided.
52. Are there operational playbooks for incident response?
AGAT maintains an incident response process aligned with SOC 2/ISO frameworks. Customers receive breach-notification SLAs in the DPA and can obtain incident response support details during onboarding.
53. How is shadow AI discovery operationalized across cloud estates?
The platform inspects network traffic and cloud footprints across AWS/Azure/GCP to detect unapproved AI services and personal account usage; results are surfaced in the Activity Auditing console.
54.What operational controls exist to prevent misconfiguration and drift?
Versioned policy templates, pre-packaged sector templates, configuration auditing, and release notes with advance notification to prevent unintended enforcement changes.
Pilot & Reseller (55–66)
55. Is a sandbox or pilot environment available?
Yes. Structured pilots under NDA are supported. A pilot tenant is provisioned on AWS, integrated with IdP, and pre-loaded with baseline policies.
56.What are typical pilot topologies and which one is fastest to deploy?
Browser extension pilots are fastest (low friction). Network proxy and Service API provide deeper coverage but take longer. Pilot topology selection is based on desired coverage.
57. What is a recommended pilot scope and duration?
Typical reseller trial parameters discussed: 2–4 weeks with up to 3 named users per lead as a starting point. Specific scope and KPIs are defined jointly for each pilot.
58.What visibility does a pilot provide out-of-the-box?
Immediate dashboards showing which AI services are used, top data classification triggers, top usage topics, and detected shadow AI across the
environment.
59.How are pilot success criteria defined?
Success criteria commonly include visibility gained (shadow AI inventory), number of policy violations detected/blocked, reduction in risky behaviors, and user adoption metrics.
60. Can trial environments be branded per-lead (subdomains, expiry)?
AGAT is open to branded trial URLs and per-lead provisioning; feasibility depends on provisioning, DNS and certificate automation and must be scoped with the product team.
61. Are reseller-specific commercial terms and trial programs available?
Yes. AGAT is open to defining reseller programmes including trial parameters, support touchpoints, and go-to-market alignment during reseller agreement finalization.
62. What support does AGAT provide during pilots?
Technical onboarding, baseline policy templates (sector-specific), demo sessions, and a working session to align on success metrics and pilot tuning.
63. Can sector-specific policy packs be included in a pilot?
Yes — pre-packaged sector templates (healthcare, finance, legal, public sector, education) can be pre-loaded and tuned during the pilot
64.What are common pilot data residency and privacy arrangements?
Pilots on AWS can be provisioned in region-specific tenants (e.g., eu-central-1) to meet data residency requirements. Retention policies and zero-dataretention upstream settings can be applied.
65.How does AGAT support reseller enablement (materials, NDA artifacts)?
AGAT can share technical documentation, control mapping, and NDA-bound artefacts (pen-test summary, SOC 2 attestation) with resellers and prospects under NDA.
66.What next steps are recommended after a successful pilot?
Move from browser extension to broader proxy/API coverage, finalize deployment topology (single-tenant or on-prem for strict compliance), sign DPA/contract addenda for sector needs, and operationalize RBAC and retention policies.
Product Features (67–75)
67. What is Behaviour Intelligence and how does it work?
Behaviour Intelligence monitors human and agent interactions to detect risky patterns (sensitive data sharing, risky prompts). It provides in-context training, targeted nudges and adoption metrics to improve security and productivity.
68.What is the Security Awareness / In-context Training capability?
When risky behaviour is detected, users receive immediate in-flow guidance (short training cards, tips). Repeated occurrences can trigger more extensive content (video, podcast or course) delivered in-context.
69.How does Pragatix perform sensitive file and document processing?
Intelligent Document Processing extracts structured content from unstructured documents, preserves layout where needed, applies
classification, and integrates with RAG pipelines.
70. What smart search and knowledge assistant capabilities exist?
Smart Search combines keyword & semantic retrieval across company sources. Knowledge Chatbot supports RAG, pre-defined domain bots, ad-hoc file Q&A and adjustable answer style/depth.
71. What developer productivity features are included?
AI Code Assistant provides codebase autocomplete, code Q&A, and codefocused search to assist developers across languages and repositories.
72. How does the platform defend against prompt injection and OWASP AI threats?
Prompt Guardian includes prompt-injection defenses, toxicity filtering, output/input validation and a risk-based policy engine aligned with OWASP AI Threat Protection
73. How are models assessed for legal and licensing compliance?
Model Guardian includes license validation, provenance checks and legal/compliance scanning as part of its four-pillar vetting (Provenance &
Identity, Static Analysis, Behavioural Vetting, Legal & Compliance).
74. Can administrators restrict answers to company-only data sources?
Yes. The Private AI Suite supports constraints to restrict responses to company data, define domain expert chatbots, and enforce grounding and scope rules.
75. How does Pragatix help measure AI adoption and ROI?
Adoption Intelligence tracks which teams/use cases generate value, monitors usage patterns and intent, identifies leading users and training needs, and provides metrics to measure productivity uplift and ROI.