Contact usGet started

AI product guidance

Amazon Bedrock governance guidance

How to govern Amazon Bedrock and the applications and agents teams build on it, with guidance for executives, risk teams and IT on guardrails and change.

How should an organisation govern Amazon Bedrock?

Amazon Bedrock pays off when what is built on it is governed as an ongoing practice rather than approved once at launch. Three things matter whatever the role. The first is knowing every application and agent in use, what each one can reach and who is accountable for it. The second is setting shared rules at the platform level, covering guardrails, approved AI, data handling and identity, so every team inherits them and changes are checked before they go live. The third is being able to show, at any time, how decisions were made and that controls held. Organisations that get these right can let teams build faster, with confidence rather than caution.

Reviewed 3 October 2026. Aicura is not affiliated with Amazon Web Services.

About Amazon Bedrock

Amazon Bedrock gives an organisation’s own teams access to AI from many providers inside its AWS accounts, along with the building blocks to ground that AI in company data, apply guardrails and build agents that call tools and act on the organisation’s behalf. Its reach grows with every team that builds on it. A single internal assistant can become dozens of applications and agents across many accounts, each choosing its own AI, data and tools. Good governance lets the organisation keep that pace of building while knowing what is running, what it can reach and who answers for it.

Where governance makes the difference

  • Building outpacing oversight. Any team with access to an AWS account can stand up a new application or agent in days. Without a register and a clear approval path, the organisation can end up with AI in production that nobody assessed and nobody owns.
  • Agents acting beyond their purpose. An agent given broad permissions or tools and reading content from outside the organisation can be steered into reaching data or taking actions nobody intended to allow.
  • Data travelling further than decided. Choices about which AI to use, where requests are processed and what is logged or retained each move data. Left to individual teams, those choices may not match what the organisation has agreed.

For the executive team

Amazon Bedrock is a platform, so govern what is built on it

Amazon Bedrock is less a single tool than a foundation that many teams build on. The risk and the value sit in the applications and agents those teams create. Oversight needs to cover each of them, with the platform itself treated as shared infrastructure that sets the rules they all inherit.

Every application and agent needs an owner

Each system built on Amazon Bedrock needs a named person accountable for it, a clear purpose and agreed limits on the information it uses and the actions it can take. Agents that act on the organisation’s behalf need this most.

Choice of AI is a business decision

Amazon Bedrock offers AI from many providers, and those options change as providers release new versions and retire old ones. Which providers the organisation is comfortable relying on, and what happens when a chosen AI is retired, are decisions for management rather than for each project team.

Oversight needs to keep pace with change

New agents, new data sources and new AI arrive continually. The estate approved last quarter will not be the one running today, so oversight works best as an ongoing practice rather than a one-off sign-off.

Questions worth putting to management

  • Is there a complete picture of every application and agent built on Amazon Bedrock and who owns each one?
  • Which of them act on the organisation’s behalf, and what limits are in place on what they can do?
  • Who decides which AI providers teams may use, and how is a provider retiring an AI handled?
  • How would the organisation know if an agent went beyond its purpose or information went somewhere it should not?
  • What is being spent on Amazon Bedrock and is it in line with the value delivered?
  • If a regulator, customer or insurer asked how this AI is governed, what would be shown to them?

What good reporting looks like

Expect regular reporting on the applications and agents in use and their owners, the main risks and how they are changing, spend against budget, any incidents or near misses and the significant decisions taken, with who made them and why.

How Aicura can help

  • One register of every application and agent built on Amazon Bedrock, each with an owner, an accountable executive and a technical contact.
  • Spend shown three ways, as contracted, as billed and as estimated from observed use, alongside budgets for each system.
  • Board reporting drafted from the governance record, covering use, risks, incidents and the decisions taken.

For risk, compliance and assurance

Set where Amazon Bedrock sits in the risk appetite

Decide which classes of information systems built on Amazon Bedrock may work with and which decisions their output may inform. Systems that inform decisions about people or customers, or that act without a person reviewing each step, warrant a higher level of scrutiny than internal drafting and search tools.

Assess each system, not just the platform

Assess the platform once for the protections and settings shared across the organisation. Then assess each application and agent on its own terms, covering the AI it uses, the data it draws on, the tools and actions it has, whether its output affects people and who reviews it.

Reassess when anything changes

A new AI, a new data source, a new tool or a change in what an agent may do moves a system away from what was assessed. A provider retiring an AI forces a change that needs the same treatment. Each needs to trigger a fresh assessment rather than wait for the annual cycle.

Make the data decisions explicitly

Decide where AI requests may be processed, since some options route requests across regions within a geography or worldwide. Decide whether full prompts and responses are logged, how long those records are kept, who may access them and for what purpose. Some AI on the platform keeps inputs for a period for abuse detection, so check each AI’s terms before approving it and record the outcome.

Confirm which protections apply

Check which of the vendor’s data commitments apply to each feature and each AI the organisation uses, since agent services and individual AI providers can carry different terms. Confirm that teams build only in accounts that sit under the organisation’s agreed settings, and bring any experimentation in personal or unmanaged accounts into that setup.

Set limits and name who acts

Agree the limits that matter, such as guardrail interventions rising, an agent calling tools it was not approved for, use of an unapproved AI or spend moving sharply, and name the person who acts when one is crossed.

Be ready to show governance to others

  • Policy. Maintain an AI policy that covers building on Amazon Bedrock specifically, including which AI and data may be used, who may build and release agents and how output affecting people is reviewed. Tie each commitment to the control that carries it out.
  • Accountability. Record committee and management decisions with their reasons when they are made. Have system owners confirm periodically that their applications and agents still match what was approved.
  • Operating evidence. Keep records that controls actually ran, such as approvals before release, guardrail versions in force, test results and interventions.
  • External scrutiny. Prepare for internal and external audit, certification against standards such as ISO/IEC 42001, regulator enquiries and insurer questions about AI controls.
  • Customers and partners. Where customers interact with or are affected by a system built on Amazon Bedrock, expect due diligence questions about which AI is used, where data is processed and how output is checked. Have the answers ready and supported by evidence.
  • Incidents. Be able to reconstruct which version of an agent, which AI, which data, which guardrails and which permissions were in place at the time.

How Aicura can help

  • Risk and impact assessments against recognised frameworks for each application and agent, run again on any material change and feeding a managed set of risks and controls that can mirror to the GRC platform.
  • The organisation’s own position on each AI it uses held alongside the provider’s, with notice to every affected system when a provider changes or retires an AI and the response recorded.
  • Limits on activity with responses the organisation chooses, from notifying the owner and re-running an assessment to re-scoping or quarantining an agent.
  • Approvals, guardrail releases and control actions recorded as they happen and verifiable by a third party without relying on the organisation’s word.

For technology and security

Discover every system built on Amazon Bedrock

Inventory every application, agent, knowledge base and guardrail across all AWS accounts and regions, including experimental accounts. Record for each one the AI it calls, the data sources it draws on, the tools it can use and the identity it runs under. Usage records and logs show which identities are calling which AI, which helps surface systems nobody has registered.

Set the guardrails at the organisation level

Amazon Bedrock guardrails can filter harmful content, block denied topics and mask sensitive information, and they can be enforced across accounts and organisational units rather than left to each application. Set a baseline guardrail centrally for every account, let teams add stricter ones for their own systems and keep guardrails versioned so the version in force is always known.

Put applications and agents under change control

Amazon Bedrock resources such as agents, guardrails, knowledge bases and prompts can be defined as infrastructure as code and changed through APIs. Keep these definitions in source control and run a preflight check in the deployment pipeline, so a new agent, tool, data source, AI or guardrail change is checked against agreed rules before it goes live.

Control who can build, call and grant access

Use AWS identity policies and organisation-wide policies to limit which teams can enable AI, which AI they can call and which regions requests may go to. Give each application and agent its own identity with only the permissions it needs, and review what an agent can reach before it is released. Where agents act on behalf of users, confirm they carry the user’s permissions rather than broader ones of their own.

Govern model calls in the request path

Calls from the organisation’s own applications and agent code to the Amazon Bedrock runtime use a documented API, and the AWS SDKs allow a custom endpoint, so this traffic can pass through an AI gateway on its way to the vendor’s usual endpoint. Use it to redact personal information before it leaves, screen inbound content for injected instructions, restrict each system to approved AI and tools, cap spend and stop an agent when needed. Private network connections to Amazon Bedrock can keep this traffic off the internet.

Rely on platform controls where the gateway cannot reach

Where Amazon Bedrock orchestrates the AI calls itself, such as managed agents or knowledge base answers, those calls happen inside the service. Govern them through enforced guardrails, tightly scoped identities, network restrictions and the service’s own logs, and place the gateway on the call into the service.

Feed everything Amazon Bedrock exposes into monitoring

API activity is recorded in AWS CloudTrail, with some agent, knowledge base and guardrail activity captured only once additional logging is switched on. Full prompt and response logging is off by default and can be delivered to the organisation’s own storage. Agent platforms on Amazon Bedrock produce OpenTelemetry traces. Bring all of these into monitoring, apply the limits risk has agreed and set responses from alerting the owner to tightening an agent’s permissions or stopping it.

Watch for AI being retired

Each AI on Amazon Bedrock carries a lifecycle state, and a retired AI stops responding after its end date. Track the state of every AI in use, plan migrations as soon as notice is given and treat each switch as a change that goes through assessment and testing.

Test against the live deployment

Red-team and evaluate applications and agents before release and on a schedule afterwards, against the deployed system with its real AI, guardrails, tools and data. Watch for changes in behaviour between tests, especially after an AI or guardrail changes.

How Aicura can help

  • One inventory of systems built on Amazon Bedrock, with the AI, tools and identities behind each, and discovery of unregistered AI from AWS cloud inventories and code repositories.
  • Preflight checks in the deployment pipeline before an agent, tool, AI or guardrail change goes live, with every answer recorded.
  • A gateway in the request path for calls to the Amazon Bedrock runtime, with guardrails derived from the agreed controls covering redaction, injection screening, tool and AI limits, budgets and the ability to stop an agent.
  • A collector inside the network for activity records and agent telemetry that flags undeclared tool calls, plus scheduled red-teaming and evals against the live deployment.

Getting lasting value from Amazon Bedrock

Amazon Bedrock pays off when what is built on it is governed as an ongoing practice rather than approved once at launch. Three things matter whatever the role. The first is knowing every application and agent in use, what each one can reach and who is accountable for it. The second is setting shared rules at the platform level, covering guardrails, approved AI, data handling and identity, so every team inherits them and changes are checked before they go live. The third is being able to show, at any time, how decisions were made and that controls held. Organisations that get these right can let teams build faster, with confidence rather than caution.

Sources

More AI product guidance

Related guides

In the platform

Register Amazon Bedrock with the rest of your AI

Aicura's AI register and discovery are free to start. Register Amazon Bedrock from a catalogue of known AI products, give it an owner and find the AI nobody has registered yet.

Sign upContact us