Agent governance cannot stop at model testing
Traditional software usually follows instructions written by developers. AI agents can interpret requests, retrieve information, call tools and take actions. That makes governance more complicated.
In a November 17, 2026 article, AWS sets out a three-phase roadmap for platform and application teams building agentic AI systems. The phases are governance foundations, pilot deployment, and evaluation and hardening.
The core message is that the platform team must establish guardrails while application teams build agents within those boundaries.
Phase 1: Build the governance foundation
Before deployment, organisations should inventory agents, models, tools and data sources. Each agent needs an owner, a business purpose and a risk classification.
The platform team should approve models, define access policies and establish a central registry for agents and shared tools. If a resource is not registered, it should not become an unmanaged route into company systems.
This is also the time to define regulatory obligations, data handling rules and evidence requirements.
Phase 2: Pilot with controls
The first agent should solve a measurable business problem. It should not receive broad access simply because broad access makes a demo easier.
AWS highlights safety, accuracy, controllability, security and privacy. Practical controls include personal data redaction, knowledge grounding, permission boundaries, escalation rules and human approval for high-risk actions.
For example, a customer service agent may retrieve an order and draft a refund recommendation. It should not automatically issue a large refund unless the business has explicitly authorised that action and the control framework permits it.
Phase 3: Evaluate and harden
A successful pilot is not the end of governance. Teams need continuous evaluation, fairness testing, security exercises, audit trails and release gates.
The system should be tested for prompt injection, unauthorised access, data leakage and unexpected tool use. Evaluation results should be checked before and after deployment, not only during initial development.
High-risk actions should have clear escalation and appeal paths.
The proposed 12-week implementation
AWS describes a customisable 12-week maturity timeline. It is a planning framework, not a guarantee that every organisation can become compliant in 12 weeks.
A small team could use the first phase to inventory and classify agents, the second to pilot one bounded workflow, and the third to run evaluations and security testing before expanding access.
A simple operating checklist
Before an agent goes live, ask:
- Who owns it and what business outcome does it serve?
- Which model and data sources can it use?
- Which tools can it call, and with what permissions?
- Which actions require human approval?
- Can the organisation reconstruct what happened?
- How will quality, bias, security and cost be evaluated?
- Who can stop the agent if it behaves unexpectedly?
What Indian enterprises should take away
The easiest way to create AI risk is to give an agent access to systems before deciding who is accountable for its actions.
Start with a narrow workflow, least-privilege permissions and measurable outcomes. Expand only after the agent passes quality, security and operational checks.
AWS’s framework is useful because it treats responsible AI as an operating discipline shared between platform and application teams. It is not a substitute for legal advice or industry-specific compliance review.
Source note: The three-phase framework and 12-week timeline are AWS’s proposed approach. Each organisation must adapt controls to its own risks and regulatory obligations.