Home / Services / AI

AI governance and agentic AI: agents that act, within limits you set

By Mirav Labs, led by Deepak MittalUpdated
Short answer

AI governance is the set of controls that decides what your AI systems may see and do, who approves the risky steps, and how you know they are working. We do it as engineering: we inventory your AI use, assess risk, and build agents with scoped permissions, approval queues, evaluation suites and an audit trail, so governance runs in the platform rather than in a policy file.

Key takeaways

  • An agent that can send, update or pay is a new kind of user in your systems, and needs its own identity and limits.
  • Governance works best when it is built into the agent platform: permissions, approvals and logs in code.
  • Roll agents out in stages: watch, suggest, act with approval, then act within limits.
  • Evaluate before launch and after every change to the model, prompt or tools.
  • India has no AI-specific statute today; existing laws, including the DPDP Act where personal data is involved, apply to AI systems.

Most organisations are past the question of whether to use AI. The harder questions now are which systems may see which data, what an agent is allowed to do on its own, and how you would know if it went wrong. Answering those in a policy document is a start. Answering them in the platform the agents run on is what makes them hold.

We do both halves of the work: we implement agents that take real work off your teams, and we build the governance into them from the first release.

Governance and implementation are one job

When governance sits apart from delivery, one of two things happens. Either the rules are so broad that teams route around them, or they are so loose that nobody can say what the agents actually do. Building them together avoids both:

  • Every agent is registered before it gets credentials, so the inventory stays complete.
  • Permissions and limits are set in code, so the policy and the system cannot drift apart.
  • Approval steps are part of the workflow, not an email someone may forget to send.
  • Logs are designed up front, so you can answer "what did it do, and why" months later.

How we run it

The same five steps work for one chatbot or fifty agents. Only the depth changes.

How we govern AI systems and agentsInventory to audit
Step 1Inventory

Every AI system and agent, and its data

Step 2Assess

Personal data, lawful basis, risk

Step 3Guardrails

Permissions, filters and limits

Step 4Human oversight

A person approves high-impact actions

Step 5Monitor and audit

Evaluation and an audit trail

Now:
  1. Inventory. We record each model, copilot, built-in AI feature of your SaaS tools and agent your teams run, with who owns it, what it is for, which data and providers it involves and what it is able to do.
  2. Assess. Each entry is rated on the data it touches, the decisions it influences and the damage a wrong action could do. Where personal data is involved, the lawful basis goes to your counsel. The riskiest uses are tackled first.
  3. Guardrails. Each agent gets its own identity, the smallest set of tools and data scopes its task needs, limits on value and volume, input and output checks, and handling that treats content it reads as untrusted.
  4. Human oversight. Together we pick the actions that pause for a human, then build a review queue where the approver sees the proposed step, the reasoning and the data behind it. Each decision is stored.
  5. Monitor and audit. An evaluation suite runs before release and on every change. Every action is logged with the agent version, trigger, data accessed, tool called, approver and outcome.

Implementing agents: deployment, approvals and evaluation

When we build an agent with you, we keep the first one narrow and useful, such as triaging support tickets, preparing refund decisions or reconciling invoices. The pattern then carries to the next one.

  • Tools as APIs. The agent acts through well-defined APIs that check permissions themselves. If the agent calls your own services, our API security work makes sure those checks hold.
  • Staged rollout. Shadow mode first, where the agent proposes and people act. Then action with approval. Then action within limits, with sampling review.
  • Evaluation harness. Realistic tasks with known outcomes, scored for success, policy adherence, unsafe actions, cost and time.
  • Stop and roll back. A switch that halts the agent at once, and a way to undo or compensate for its recent actions where the systems allow.

Before go-live we recommend an adversarial test by people trying to steer the agent off course; see AI red teaming. If the agent answers from your documents, data readiness for AI covers the retrieval layer.

Frameworks we use for structure

India has no AI-specific statute. On 5 November 2025 MeitY published the India AI Governance Guidelines, saying it would rely on existing laws wherever possible. Where an AI system handles digital personal data, that includes the DPDP Act, which has no AI carve-out. For programme structure we draw on the NIST AI Risk Management Framework, a voluntary framework built around four functions (Govern, Map, Measure, Manage), and ISO/IEC 42001:2023, the international standard for an AI management system. We use them as scaffolding, not as a substitute for controls in your systems.

What you get

DeliverableWhat it isWho uses it
AI inventory and risk ratingsEvery AI system and agent with owner, data, providers, actions and ratingTechnology, risk and privacy leads
Working agentA first agent for an agreed workflow, deployed in stages in your environmentThe business team it serves
Guardrail configurationAgent identities, tool scopes, limits and input and output checksEngineering and security
Approval matrix and queueThe actions that pause for review, the approver for each, the context shown and the record keptOperations and business owners
Evaluation suiteTask set, scoring and a baseline report, ready to re-run on every changeEngineering and QA
Audit trail and governance packLog design, dashboards, review cadence and the policy mapped to controlsInternal audit, leadership, customers' reviewers

How to start

Bring one workflow you would like an agent to take on and a rough list of the AI tools your teams already use. We will map both, agree the limits and approvals, and plan a staged first release. To set that up, talk to us.

The DPDP side of this

Our DPDP practice, DPDP Ready, has a detailed guide on how this work maps to India's data protection law. Read it on DPDP Ready →

Frequently asked questions

What is agentic AI, and how is it different from a chatbot?

A chatbot answers questions. An agent takes steps towards a goal: it reads data, decides what to do next and calls tools such as email, a CRM, a ticketing system or a payments API. Because it acts, an agent needs the controls you would give a new staff member, such as limited access, approval for big decisions and a record of what it did.

How do you deploy an AI agent safely in production?

Give it a narrow task, its own credentials and only the tools that task needs. Start in a mode where it proposes actions and a person carries them out, then move to acting with approval, and only then to acting alone within value and volume limits. Keep a switch that stops it at once, and log every action.

What should an AI governance policy include?

At minimum: who may approve a new AI use, how risk is rated, which data and actions need extra controls, who reviews high-impact decisions, how systems are tested and monitored, and how incidents are handled. The policy is far more useful when each of these maps to a control in your systems that can be checked.

How do you evaluate an AI agent before go-live?

Run it against a fixed set of realistic tasks with known correct outcomes, and score task success, policy adherence, wrong or unsafe actions, cost and time. Include tasks designed to tempt it outside its scope. Repeat the run whenever the model, prompt or tools change, and compare against the last result.

Who should own AI governance in a company?

A named senior owner, usually from technology, risk or operations, with a small group that includes security, privacy, legal and the business teams using AI. Each AI system also needs its own accountable owner. Without named owners, inventories go stale and approvals get skipped.

Scope it with us

A 30-minute call to understand your systems and what good looks like. We aim to send a written, fixed-scope proposal within three working days.

Talk to us

Related services

This page describes engineering and security services. Where it mentions laws, rules or standards, it is general information, not legal advice.

Talk to us