AI Implementation & Safe Usage Framework/ Policy

This Framework establishes how Legistify implements, governs and safely uses artificial intelligence across its products and internal systems. It replaces the informal Legistify AI Architecture note with a formally owned, version-controlled and approved policy, and it sets the mandatory security, privacy and responsible-use controls for all AI use at Legistify.

Document Name:	AI Implementation & Safe Usage Framework / Policy
Classification:	Internal
Document Owner:	AI Product & Engineering Lead
Document Approver:	CISO
Original Document Issue Date:	6/11/2026
Current Edition:	Version 1.0
Revision History:
S. No.	Description of Change	Date of Change	Version No.
1	Initial Release	6/11/2026	1.0

1. Purpose & Objective

The purpose of this Framework is to define how Legistify designs, implements, operates and uses artificial intelligence in a secure, privacy-preserving and responsible manner, and to assign clear accountability for doing so.

The objectives are to:

Establish a single, formally approved source of AI governance that supersedes informal documentation.
Ensure all AI features are built only on approved technologies and architectures.
Protect client, personal and confidential data processed by AI from unauthorised access, disclosure, retention or use, including use for model training.
Mandate preventive and detective security controls (guardrails, access control, encryption, network isolation, logging) for all AI processing.
Set the boundaries of acceptable and prohibited AI use, with human oversight over AI assisted outcomes.
Meet applicable legal, regulatory and contractual obligations, including India's data protection regime.
Define review, exception and enforcement processes so the Framework stays current and is consistently applied.

2. Scope

This Framework applies to all artificial-intelligence capabilities that Legistify builds, operates or uses, and to everyone who develops, administers, procures or uses them.

In scope

AI features embedded in Legistify products, including Codex (Contract Management), the Litigation Management Tool (LMT) and the Notice Management Tool (NMT), and any AI capability added to other products.
All AI processing performed through AWS Bedrock in Legistify's AWS environment, including foundation-model inference, guardrails, prompts, uploaded files and AI outputs.
The supporting infrastructure for AI: IAM roles, networking, encryption keys, storage and logging used by AI workloads.
All personnel who design, develop, review, operate, administer or use AI features, including employees, contractors and relevant vendors.

Out of scope

General IT systems and non-AI processing, which are governed by Legistify's other information security policies (Access Control, Cloud Platform, Network Security, Secure Development, Data Privacy and Data Protection).
Any use of external or public AI platforms, which is prohibited under this Framework (see Section 7).

Where this Framework and another policy both apply, the stricter control prevails.

3. Definitions

Term	Definition
Artificial Intelligence (AI)	Technology that performs tasks typically requiring human intelligence, such as understanding language, extracting information and generating text. At Legistify this is delivered through large language models.
Large Language Model (LLM) / Foundation Model	A pre-trained model, accessed via AWS Bedrock, that generates or analyses text in response to a prompt. Legistify does not build or train its own foundation models.
Term	Definition
AWS Bedrock	The managed AWS service through which Legistify accesses approved foundation models within its own AWS account.
Guardrail	An AWS Bedrock Guardrails control applied to model inputs and outputs to block or mask sensitive, harmful or non-compliant content.
Prompt	The input (instruction and context, including user text and content derived from uploaded files) sent to a model for a single AI request.
Client / Customer data	Any data belonging to or relating to a Legistify customer, including documents, contracts, notices, case data and the personal data contained within them.
Personal data / PII	Information relating to an identified or identifiable individual, such as names, addresses, email IDs and financial identifiers.
AI output	Content generated by a model in response to a prompt.
Model provider	The third party that supplies a foundation model made available through AWS Bedrock.
Responsible AI	The principles and controls that keep AI use accurate, transparent, fair, secure and subject to human oversight.

4. Governance, Roles & Responsibilities

Accountability for AI at Legistify is assigned as follows. Role titles may be held by named individuals in the information security role register; this Framework defines the responsibilities attached to each role.

Role	Responsibilities
CISO (Approver)	Approves this Framework and material changes to it; holds final accountability for AI risk; authorises exceptions.
Role	Responsibilities
AI Product & Engineering Lead	Ensures AI features are designed and built to this Framework; maintains the approved-model and architecture list; assesses new AI use cases before launch.
Engineering & DevOps	Implement and operate the technical controls; configure guardrails, IAM policies, encryption and network isolation; keep logging enabled.
Employees, contractors and vendors	Use AI only within this Framework; follow acceptable-use rules; report suspected misuse or incidents.

Significant new AI use cases are reviewed by the AI Product & Engineering Lead and the Information Security team before release, and high-risk personal-data processing follows the DPIA process referenced in Section 16.

5. Approved AI Architecture & Technologies

All AI features at Legistify are delivered through a single, controlled architecture. No other AI technology may be used without approval under the change process in Section 13.

AWS Bedrock only. All AI features are powered by large language models accessed through AWS Bedrock within Legistify's own AWS account. Legistify does not build, host or train custom AI models outside AWS Bedrock for these use cases.
No external or public AI platforms. There is no direct integration with external or public AI platforms (for example, public chatbot services or third-party AI APIs).
Approved foundation models only. Only foundation models on the approved list may be enabled. The list is maintained by the AI Product & Engineering Lead; as of this version it comprises Claude Opus 4.6, Claude Opus 4.5 and Claude Haiku 4.5. Adding or changing a model follows Section 13.
Data residency in India. Models are deployed in the AWS Mumbai region (ap-south-1). No AI processing occurs outside India.
Processing is request-scoped. Client data is processed only for the specific AI request initiated by the user and is not retained by the model or the Bedrock service after the request (see Section 8).

The security controls that protect this architecture — guardrails, access control, encryption, network isolation and logging — are mandatory and are set out in Sections 9 to 11.

Reference: Legistify AI Architecture and AI Architecture FAQs, Information Security collection, help.legistify.com.

6. Acceptable & Prohibited Use of AI

Permitted use. AI may be used within Legistify products and workflows for the business purposes it is designed to serve, such as extracting and summarising information from contracts, litigation and notice documents, drafting assistance, question answering over authorised data, and similar tasks — provided the processing goes through the approved AWS Bedrock architecture, stays within the user's authorised data, and keeps a human responsible for any decision or external communication.

Prohibited use. The following are not permitted:

Sending Legistify or client data to any external or public AI service, browser extension, or personal AI account.
Using AI models, providers, regions or integrations that are not on the approved list. Using AI to make a final legal, compliance or contractual decision, or to send an external communication, without human review and sign-off.
Attempting to bypass, disable or weaken guardrails, access controls, encryption, network isolation or logging.
Feeding data a user is not authorised to access into an AI request, or using AI to circumvent access controls.
Using AI outputs as authoritative legal advice without verification by a qualified person.
Processing special-category or excessive personal data through AI beyond what the task requires (data minimisation applies).
Using client data to train, fine-tune or evaluate any model.
Misuse is handled under Section 17 (Enforcement).

7. Data Protection, Privacy & Residency

Data processed by AI is protected to the same standard as the rest of Legistify's platform, with additional AI-specific controls.

No training on client data. Customer prompts, uploaded files and AI outputs are never used to train, fine-tune or improve AWS or third-party models.
No model-side storage. Customer data is not stored by the foundation models or the Bedrock service. No indexing, vectorisation or embedding storage is performed in the AI request flow. Data is processed only for the specific request initiated by the user.
Provider isolation. Model providers cannot access customer data, the deployed inference environment, or logs within Legistify's infrastructure.
Storage of inputs. Files uploaded for AI processing are stored in Legistify's Amazon S3 under the encryption and access controls in Sections 9 and 10.
Data residency. All AI processing and storage remain in India (AWS ap-south-1). No AI processing occurs outside India.
Data minimisation. Only the data needed for a given AI task is processed, consistent with Legistify's Data Privacy and Data Protection Policy.
Retention and deletion. AI-related data follows Legistify's retention schedule. Customers may request deletion; on contract termination or a verified request, a hard deletion is performed, which may include cryptographic erasure by deleting the relevant KMS keys.
Data subject rights. Requests relating to personal data processed by AI are handled through Legistify's data-protection process (Section 16).

Reference: Legistify AI Architecture FAQs and Data Privacy and Data Protection Policy, help.legistify.com.

8. AI Guardrails & Content Safety

AWS Bedrock Guardrails are implemented as an additional safeguard layer on AI processing. They are configured and managed within Legistify's AWS Bedrock account and are not a replacement for encryption or access controls.

Applied to inputs and outputs. Guardrails are applied to both model input and model output.
PII and sensitive-content detection. Guardrails use machine learning to identify personal and sensitive information such as names, addresses, email IDs and financial identifiers (for example, masking a PAN).
Block and mask modes. Guardrails can run in Block Mode, which stops a request containing sensitive information, or Mask Mode, which redacts PII with tags such as {NAME} .
Coverage of uploaded content. Guardrails also scan content derived from uploaded files for PII, harmful content and policy violations.
Versioned and tested. Guardrails are configured, tested and then deployed as versions. Changes are controlled and recorded.
Responsible-AI alignment. Guardrail configuration is maintained to align with the responsible-AI principles in Section 12.

The Information Security team reviews guardrail configuration periodically and after any significant change to AI features or data flows.

Reference: Legistify AI Architecture and AI Architecture FAQs, help.legistify.com.

9. Security Controls — Access, Encryption & Network

AI workloads are protected by the same information security controls as the rest of the platform, applied specifically to Bedrock, the inference environment and AI data stores.

10.1 Access control

Role-based access control (RBAC) is enforced using AWS IAM.
Least-privilege access policies apply to AI resources, with controlled access to Bedrock APIs and the supporting infrastructure.
Access follows the need-to-know and least-privilege principles of Legistify's Access Control Policy, including periodic access reviews and session controls.

10.2 Encryption

Data in transit is encrypted using TLS 1.2 or higher.
Data at rest (including S3 storage and logs) is encrypted using AWS KMS (SSE-KMS), with AWS-managed or customer-managed keys.
Encryption keys are governed by least-privilege IAM policies; key deletion can be used for cryptographic erasure (Section 8).

10.3 Network security & isolation

AI processing occurs inside a private AWS VPC.
AI endpoints are not publicly exposed.
Communication between services is over secure internal channels.

These controls operate together with the broader Cloud Platform and Network Security policies; where controls overlap, the stricter applies.

Reference: Legistify AI Architecture, AI Architecture FAQs and Access Control Policy, help.legistify.com.

10. Logging, Monitoring & Audit

AI activity is logged and monitored so that use can be audited and incidents detected.

Activity logging. User activity logs for AI features are retained for over one year.
Technical logs. Infrastructure logs are retained in CloudWatch, typically for around 15 days, for operational monitoring and troubleshooting.
Model invocation logging. Model invocation logs are captured to support auditability. Sensitive data and PII are not intended to be captured in logs; the Information Security team verifies this in guardrail and logging configuration.
Encryption and access. Logs are encrypted (SSE-KMS) and access to them follows least-privilege IAM.
Audit. The Information Security team reviews AI logs and configuration as part of periodic security audits and during incident investigations.

Logging requirements are consistent with Legistify's broader monitoring and observability controls.

11. Responsible AI Principles

Legistify uses AI in a way that is accountable and safe for a legal-technology context. All AI features are designed and operated to these principles:

Human oversight. AI assists people; it does not make final legal, compliance or contractual decisions autonomously. A qualified person reviews AI-assisted outcomes before they are relied on or sent externally.
Accuracy and verification. AI outputs are treated as drafts or aids, not authoritative legal advice, and are verified against source documents and authorised data.
Transparency. Customers are informed, through product documentation and the published AI architecture, about how AI is used and how their data is handled.
Fairness and non-discrimination. AI features are designed to avoid producing unfair or discriminatory outcomes, and issues raised are investigated.
Privacy and security by design. Data minimisation, guardrails, encryption, access control and network isolation are built into AI features from the outset.
Purpose limitation. AI is used only for its intended business purpose and only on data the user is authorised to access.
Accountability. Responsibility for AI use is assigned under Section 5, and misuse is addressed under Section 17.

12. Third-Party & Model Provider Management

AI relies on AWS Bedrock and the model providers it offers. These relationships are managed to keep data protected and changes controlled.

Provider no-access. Model providers cannot access Legistify customer data, the inference environment or internal logs, and client data is not used to train or improve their models.
Processing within AWS. All model access is through AWS Bedrock in Legistify's AWS account in ap-south-1; no data is sent to external or public AI platforms.
Contractual controls. Third-party processing is governed by the applicable AWS terms and Legistify's processor/DPA clauses, consistent with the Data Privacy and Data Protection Policy.
Change management for AI. Enabling a new foundation model, region, guardrail change or AI integration requires review by the AI Product & Engineering Lead and the Information Security team, a check against this Framework and data-protection requirements, and recording in the approved-model list and revision history.
Periodic review. The approved-technology list and provider arrangements are reviewed at least annually, or sooner on a material change.

13. AI Incident Management

AI-related incidents are managed through Legistify's established security incident response process, with AI-specific triggers.

An AI incident includes, for example: data exposure through an AI feature; a guardrail bypass or failure; use of an unapproved model, region or external AI service; unauthorised access to Bedrock, keys or AI data stores; or an AI output that causes material harm or a compliance breach.

Report. Anyone who suspects an AI incident reports it promptly to the Information Security team through the standard incident channel.
Respond. The Information Security team triages, contains, investigates and remediates using the incident response process, including preserving relevant logs.
Notify. Where personal data is affected, the Data Protection Officer assesses notification obligations under applicable law (Section 16).
Learn. Root cause is identified and controls, guardrails or this Framework are updated as needed, with changes recorded in the revision history.

This section operates alongside Legistify's general incident response and breach handling procedures.

14. Legal & Regulatory Compliance

AI use complies with applicable law and with Legistify's information-security and privacy commitments.

Data-protection law. AI processing of personal data is carried out in line with India's Digital Personal Data Protection Act, 2023 (DPDP Act) and other applicable dataprotection laws, including principles of lawful processing, purpose limitation, data minimisation and data subject rights.
Data-protection policy linkage. This Framework operates with Legistify's Data Privacy and Data Protection Policy, which covers consent, data subject rights, dataprotection impact assessments (DPIAs) for high-risk processing, and processor obligations.
DPIA for high-risk AI. Where an AI use case involves high-risk processing of personal data, a DPIA is conducted before launch.
Data residency. Processing stays within India (ap-south-1), supporting datalocalisation expectations.
Certifications. Legistify's information-security programme is maintained in support of its certifications (for example SOC 2 Type II and ISO/IEC 27001); AI controls form part of that programme.
Contractual obligations. AI use honours customer contractual commitments on confidentiality, security and data handling.

Reference: Data Privacy and Data Protection Policy and Security Certifications, help.legistify.com. Specific DPDP applicability should be confirmed with Legal.

15. Training, Awareness & Communication

This Framework is communicated to all relevant stakeholders so that it is understood and followed.

Publication. The approved Framework is published on Legistify's centralised policy portal and is accessible to employees and contractors.
Communication. It is communicated to relevant stakeholders — employees, contractors and, where applicable, vendors — on publication and on each material update.
Acknowledgement. Personnel in scope acknowledge the Framework as part of onboarding and policy-update cycles, in line with Legistify's information security practice.
Awareness. AI acceptable-use and data-handling expectations are included in security-awareness training.
Customer-facing transparency. Customer-facing information on AI architecture and data handling is maintained in the help centre.

The document owner ensures communication occurs and records are kept.

16. Policy Review, Enforcement, Exceptions & Non-Compliance

Review and maintenance. This Framework is reviewed at least annually (next review date in Section 1), and additionally whenever there is a significant change to AI technologies, data flows, architecture or applicable law. The document owner conducts the review, the Information Security team reviews the content, and the CISO approves changes. Every change is recorded in the revision history with a new version number.

Exceptions. Any deviation from this Framework requires a documented exception, riskassessed by the Information Security team and approved by the CISO, with a defined scope, compensating controls and expiry date. Exceptions are logged and reviewed.

Non-compliance. Non-compliance may expose Legistify and its customers to security, privacy, legal and reputational risk. Breaches are handled under Legistify's disciplinary and HR processes and may lead to withdrawal of access, and, for contractors or vendors, contractual action. Suspected breaches are reported as incidents under Section 14.

Policy Revision History

Date	Version	Author	Reviewer	Approver	Comments
6/11/2026	1.0	AI Product & Engineering Lead	Information Security team	CISO	Initial formally governed Framework. Consolidates and supersedes the informal Legistify AI Architecture note by adding document ownership, roles and responsibilities, approval authority, review frequency, version control and revision history, together with the security, privacy and responsible-use requirements governing AI at Legistify.