1. Purpose
Enable useful AI-assisted development and product features while protecting customer information and keeping financial results reliable. AI may help write, explain, classify non-customer material, and propose actions; people remain accountable for its use. This policy supports the Information Security Policy.
2. Scope
This policy covers AI features in Moogul, coding assistants, internal AI tools, agents, and model integrations used by employees, contractors, or vendors. Product model hosting and internal development tools have different approval boundaries below; neither is exempt from data protection or access rules.
3. Customer Data And Redaction
Raw customer identities, financial records, transaction histories, private conversations, and credentials must remain in Moogul’s authorized internal systems. Product AI may receive only an approved, minimized and redacted summary necessary for the requested feature, through Bedrock. Do not send customer data, including redacted customer-derived summaries, to coding or general internal AI tools. Do not use customer data for model training, fine-tuning, or development experiments.
Use synthetic examples, public material, and approved company information for AI work. Minimize and redact inputs before sending them: remove credentials, identifiers, private content, and unrelated context. Redaction must happen within Moogul-controlled processing before a model request, rather than asking a model to remove the data after receipt. Removing a name does not make financial records anonymous. Redacted summaries remain protected customer information and require an approved purpose, fields, and processing path.
Apply these limits to attachments, images, audio, embeddings, conversation history, tool results, logs, and error reports as well as prompt text. Block requests that cannot meet the data boundary. Keep identity mappings and unnecessary underlying values internal. Do not reintroduce removed data in later steps. Free-form customer input must be filtered and minimized before inference; block it when that cannot be done reliably.
4. Calculations And Forecasting
Internal non-AI tools must perform financial calculations, conversions, projections, and forecasting using tested code and explicit assumptions. A language model must never be the source of financial arithmetic or forecast values.
AI may explain an approved redacted summary of verified results or propose a structured operation. Internal code must validate the operation, check authorization, and perform the calculation. Render authoritative figures from internal results, not model-generated arithmetic; validate any AI explanation against those results. Do not send underlying customer records merely to help a model calculate. A model’s proposed action or narrative is not evidence that a calculation or change succeeded.
5. Product Models And AWS Hosting
Product model inference must use approved Amazon Bedrock deployments operated within AWS. Prompts and outputs must not be routed to external model-provider APIs. AWS documents that model providers cannot access the AWS-managed deployment accounts or their prompts, outputs, and logs. Bedrock still processes the submitted input through a model; AWS hosting does not remove the restrictions in section 3.
Use least-privilege access, encryption, approved processing regions that meet applicable data-residency requirements, and private connectivity through AWS PrivateLink for production inference. Bedrock is an AWS-managed service outside Moogul’s VPC; PrivateLink provides private access to it. Do not describe the model itself as hosted inside Moogul’s VPC.
Review each model’s terms, retention and abuse-review behavior, routing, optional features, and logging settings before use or material changes. Choose configurations that prevent reuse for training; do not enable optional data sharing. Do not claim zero retention without verifying the exact model and configuration. Prohibit silent fallback to direct provider APIs or unapproved regions.
6. Development And Internal AI Tools
Approved coding and business AI tools may operate outside AWS for work that excludes customer data. Record permitted uses and data categories in the vendor inventory; an existing software account does not automatically approve its AI features.
Developers may use approved tools for routine coding, research, documentation, and synthetic testing without approval for each prompt. Check generated work through normal review and relevant verification. Agents may act within approved task permissions and sandbox boundaries; protect credentials through controlled tool access rather than including secrets in model context. Production access and consequential actions remain subject to the Access Control Policy.
7. Review And Accountability
The CEO and Security Review Team review new AI features and material model/data-flow changes for security, privacy, customer impact, misuse, and failure modes. Keep an owner, intended use, approved inputs, hosting/retention settings, evaluation results, and a way to disable or roll back each feature. Test redaction, data isolation, authorization, prompt-injection resistance, and calculation boundaries before release and after relevant changes.
Identify AI-assisted features clearly. Do not present generated text as a verified financial result or allow it to authorize access or consequential customer actions. Report suspected disclosure or unsafe behavior promptly and follow the incident procedures.
8. ISO/IEC 42001 Direction
Moogul’s target framework for AI governance is ISO/IEC 42001:2023. We are developing our AI practices toward its management-system requirements, beginning with ownership, risk and impact assessment, data controls, evaluation, monitoring, and improvement.
This is a development direction, not a claim of conformity or certification. The CEO must maintain a gap assessment and track remaining work, including the management-system scope, objectives, internal audit, management review, and operating evidence, before making a stronger claim. Policy review, exceptions, and enforcement follow the parent Information Security Policy.