How we govern AI agents and your data.
AI should take on the busy work without taking away accountability. This page sets out the commitments behind every Mirror.lk engagement, the frameworks we use as references, and what we do — and don't — claim.
Six things every engagement includes.
A named human for every workflow
Autonomy earned on evidence
Least privilege
Everything logged
Cost under control
You own the build
Your data, mapped and minimised.
Before anything goes live, you receive a data-flow map for your workflow and a list of every service that processes your data.
- —We map every data flow — your systems, our tools and any model provider — before go-live, and give you the diagram.
- —We list the third-party services that process your data in each engagement, so you can meet your own disclosure duties.
- —Model API calls may be processed outside your country. We tell you where, choose business terms that exclude training where available, and avoid sending data a task doesn't need.
- —Access ends when the engagement ends; we confirm removal in writing.
Mapped to the NIST AI Risk Management Framework.
The NIST AI RMF is a voluntary framework organised around four functions. There is no certification for it; we use it to structure how we manage AI risk.
Source: NIST, Jan 2023(opens in a new tab)| Function | What it covers | How we apply it |
|---|---|---|
| Govern | Policies, roles and accountability for AI risk. | Named owners per workflow, written autonomy level, approval rules and an AI register for each client. |
| Map | Context, purpose and potential impacts of each AI system. | Workflow audit: process map, data classification, affected people and decisions, data-flow diagram. |
| Measure | Testing and tracking of performance and risk. | Evaluation sets before launch, sampled QA in production, escalation and correction rates, cost per task. |
| Manage | Prioritising and acting on risks over time. | Exceptions queue, incident playbook, monthly review, autonomy demotion when quality drops. |
Every agent is checked against the OWASP Top 10 for LLM Applications.
| ID | Risk | Our control |
|---|---|---|
| LLM01 | Prompt Injection | Untrusted input isolated from instructions; tool allow-lists; output checks before actions. |
| LLM02 | Sensitive Information Disclosure | Data classification, redaction, and only the data each task needs. |
| LLM03 | Supply Chain | Established providers and maintained libraries; dependencies reviewed and pinned. |
| LLM04 | Data and Model Poisoning | Curated knowledge sources with an owner; changes reviewed before they go live. |
| LLM05 | Improper Output Handling | Outputs validated and escaped before reaching other systems or customers. |
| LLM06 | Excessive Agency | Scoped permissions, approval steps for consequential actions, reversible operations. |
| LLM07 | System Prompt Leakage | No secrets in prompts; credentials kept in managed secret stores. |
| LLM08 | Vector and Embedding Weaknesses | Access-controlled knowledge bases; retrieval scoped per client. |
| LLM09 | Misinformation | Answers grounded in approved sources; hand-off to a person when unsure. |
| LLM10 | Unbounded Consumption | Per-task token budgets, rate limits and cost alerts. |
Dates our clients are preparing for.
Automated-decision transparency (APP 1.7–1.9)
Privacy policies of APP entities must disclose where computer programs make, or substantially and directly contribute to, decisions that could significantly affect individuals. We document where AI is used in each workflow so you can describe it accurately. Australian clients also remain accountable under APP 8 when personal information is disclosed overseas — our data-flow maps support that assessment.
Source: Office of the Australian Information Commissioner, 2026(opens in a new tab)Personal Data Protection Act, Parts I and III
Core obligations for processing personal data, and for controllers and processors, come into operation. As a Sri Lanka-based processor, we are preparing our own processes for these obligations.
Source: Government of Sri Lanka, 22 Jul 2026(opens in a new tab)Guidance for AI Adoption
The National AI Centre's guidance sets out essential practices including clear accountability, managing risk, testing and monitoring, and maintaining human control — the same principles behind our commitments.
Source: National AI Centre, 21 Oct 2025(opens in a new tab)General information, not legal advice. Confirm your obligations with a qualified adviser.
Straight about our status.
- —We are not ISO/IEC 42001 certified. We use it, the NIST AI RMF and the OWASP LLM Top 10 as reference frameworks.
- —We are not an official partner of GoHighLevel, n8n or any AI model provider. We recommend tools on fit.
- —We don't describe any system as “fully autonomous” or promise zero errors. We design for errors to be caught.
- —We don't publish client results without permission, and we don't quote research findings as if they were our own results.
Questions about security or data handling? Contact us — we'll answer in writing.
See where AI can return hours to your team.
Book a free workflow audit call. We'll look at one recurring workflow with you and tell you honestly whether automation or an agent is worth it.