
Key takeaways
- Generative AI governance depends on validating the data that trains, grounds, and feeds AI models and agents, not the models alone.
- A Data Bill of Materials maps every model and agent to the exact datasets, knowledge bases, and systems it actually depends on.
- Guardrail Gap Analysis correlates DBOM data types against active guardrail configuration to reveal exactly what's blocked, allowed, or unintentionally exposed to users.
- Effective AI governance solutions keep a live view of data dependencies and extend access analysis to non-human identities, including agents and service accounts, not only human users.
- AI governance must be continuous, since new data sources and types constantly change what a guardrail configured last month actually covers.
What is generative AI governance?
Generative AI governance is the set of policies, processes, and technical controls that oversee how generative AI models and agents are built, deployed, and monitored in production.
For systems built on Amazon Bedrock, that oversight is only as strong as the organization's visibility into the data feeding those systems, which is exactly where GenAI governance tends to break down in practice.
As generative AI systems move from experimentation into production, organizations are encountering a new class of security and governance challenges. These challenges aren't driven by the models themselves, but by the data those models depend on.
In a November 2025 AWS Builder Center article, Bedrock Data CTO and co-founder Pranava Adduri joined AWS Bedrock Solutions Architects to examine what it takes to ship GenAI systems securely on Amazon Bedrock. The central insight is simple but critical:
You cannot govern AI systems effectively unless you understand and continuously validate the data they rely on.
This post summarizes the key ideas from that article and explains why data-aware governance is becoming foundational for production AI.
Why does GenAI have a different dependency model than traditional software?
Traditional software depends on static code libraries. GenAI systems do not.
Instead, models and agents depend on:
- Training and fine-tuning datasets
- Vector stores used for Retrieval-Augmented Generation (RAG)
- Upstream enterprise systems that evolve over time
These data dependencies directly influence what a model knows, what it can access, and what it can generate. As a result, static inventories or one-time governance reviews are no longer sufficient.
This is the problem space Pranava has focused on for years: making data risk operational, measurable, and enforceable in AI systems.
Training and fine-tuning datasets
Training and fine-tuning datasets are what a model was originally built or adapted on: the foundation of what it "knows" before it ever answers a prompt. Governance has to account for this layer first, since it shapes every downstream output the model produces.
Vector stores for Retrieval-Augmented Generation (RAG)
Vector stores are the knowledge base a model or agent queries at inference time to ground its responses in current, organization-specific data. Unlike training data, this content can change daily, so what the model can retrieve, and potentially expose, shifts constantly.
Upstream enterprise systems
Upstream enterprise systems are the live business systems, including CRMs, HR platforms, and data warehouses, that agents connect to and retrieve from directly. These systems evolve independently of the model itself, so an agent's effective data access can change without anyone touching the model.
What is a Data Bill of Materials (DBOM)?
One of the core concepts explored in the AWS article is the Data Bill of Materials (DBOM).
A DBOM answers a question every security, compliance, and AI leader eventually asks:
What data actually feeds this model or agent, end to end?
In practice, a DBOM maps:
- Models to their training and fine-tuning datasets
- Knowledge Bases to their source systems
- Agents to the data they retrieve during inference
This mapping is not just about visibility. It creates the foundation for validating whether governance and safety controls are appropriate for the actual data in use.
Models to training and fine-tuning datasets
This traces which datasets shape what a given model knows and can produce, so teams can validate a model's outputs against exactly what it was trained or fine-tuned on.
Knowledge bases to source systems
This traces which upstream enterprise systems feed a RAG knowledge base, so a change in a source system's access controls or content can be flagged before it ever reaches the model.
Agents to retrieved data
This traces exactly what data an agent pulls in during inference, not just what it was trained on, since an agent's real-time queries can surface data no training review ever covered.
How does data-aware governance work on Amazon Bedrock?
The AWS walkthrough runs in four stages, plus a fifth that does not stop.
- Discovery. ArgusAI assumes a pre-deployed role in the target AWS account and introspects Bedrock workloads through ListAgents, ListKnowledgeBases, GetAgent, and GetKnowledgeBase. It traces each agent through its knowledge bases to the origin data, whether that sits in an S3 bucket or a SharePoint site.
- Classification. Each source is scanned for the data types it actually contains. This runs on Bedrock Data's Metadata Lake, a graph of enterprise data and its attributes, so classification the organization already has is reused rather than redone for the agent.
- Gap analysis. The guardrail attached to the agent is retrieved and compared against the data types found, so every type resolves to blocked, explicitly allowed, or exposed.
- Remediation. ArgusAI returns the specific change that closes the gap, applied through the UpdateGuardrail API as a topic deny rule or a custom regex filter.
- Continuous re-evaluation. When a new source or data type appears, the DBOM updates and the gap analysis runs again against the current guardrail.
Why must guardrails be data-aware?
Amazon Bedrock provides powerful guardrails that allow teams to block or redact categories such as PII, PHI, and PCI. However, the AWS article highlights an important nuance: a guardrail is only effective if it covers the specific sensitive data an agent can access.
In the example discussed, a PTO Approval Agent retrieves HR data that includes not only PII but also compensation information. While common guardrails may block PII by default, they may not block salary data unless explicitly configured.
This creates a subtle but serious risk. The system appears protected, but sensitive information can still flow to the model and back to users.
What is guardrail gap analysis?
To address this problem, the article introduces the idea of Guardrail Gap Analysis.
By correlating:
- The data types identified in the DBOM
- The active Amazon Bedrock Guardrail configuration
Teams can determine whether sensitive data is:
- Properly blocked
- Explicitly allowed
- Or unintentionally exposed
In the PTO agent scenario, this analysis reveals a critical gap: compensation data is accessible even though it should not be returned to users. Updating the guardrail configuration closes the gap, but the key insight is that the risk only becomes visible when data lineage, classification, and policy coverage are evaluated together.
Properly blocked
The guardrail correctly prevents the sensitive data type from reaching the model or the user, exactly as configured.
Explicitly allowed
The data type is intentionally permitted through, reflecting a deliberate policy decision rather than an oversight.
Unintentionally exposed
This is the gap case: sensitive data flows through because the guardrail was never configured to catch it, as with the PTO agent's compensation data.
What should you look for in AI governance solutions?
Teams evaluating AI governance solutions for GenAI and agentic systems are often comparing tools against the wrong checklist. General AI governance platforms focus on policy documentation and model risk scoring. GenAI systems on Amazon Bedrock raise a narrower, more technical question: does the platform actually know what data a model or agent can reach, and does it prove that guardrails cover it? Three criteria separate platforms that answer that question from ones that don't.
Does it maintain a live Data Bill of Materials?
A one-time inventory goes stale the moment a new data source is connected. The solution should update the DBOM automatically as data sources and data types change, rather than requiring a manual re-audit every time an agent's access changes.
Does it validate guardrails against actual data exposure?
Configuring a guardrail isn't the same as confirming it covers the data an agent can actually reach. The solution should correlate what is classified as sensitive against what the active guardrail blocks, name the gap, and return the configuration change that closes it. Applying the change stays a decision a person makes.
Does it cover non-human identities?
As agents act autonomously across enterprise systems, governance needs to extend to service accounts and agent identities, not just human users. A platform that only maps human access will miss exactly the exposure GenAI systems introduce.
Why must AI governance be continuous, not a one-time exercise?
A final takeaway from the AWS article is that AI governance cannot be a one-time exercise.
Data sources change. New files appear. New data types are introduced. A guardrail that was sufficient last month may be incomplete tomorrow.
This is why data-aware governance must be operational:
- DBOMs need to update as data changes
- Guardrail coverage must be re-evaluated automatically
- Teams need early warnings before risks reach production
This approach allows organizations to move quickly without sacrificing security or compliance.
DBOMs must update as data changes
New files, new data types, and new sources appear constantly, so a static DBOM is out of date almost as soon as it's produced.
Guardrail coverage must be re-evaluated automatically
A guardrail sufficient last month may miss a new sensitive data type introduced this month, so coverage needs to be checked continuously, not on a fixed review cycle.
Teams need early warnings before risks reach production
Catching a gap before deployment is far cheaper, and far less damaging, than catching it after sensitive data has already reached a model or a user.
How does static governance compare to data-aware, continuous governance?
The difference between the two approaches is concrete enough to put side by side:
| Dimension | Static/one-time governance review | Data-aware, continuous governance |
|---|---|---|
| Data mapping | Manual inventory at a point in time | Live DBOM mapping models/agents to actual data dependencies |
| Guardrail validation | Configured once based on assumed data exposure | Guardrail Gap Analysis correlates DBOM findings against active guardrail configuration |
| Detecting new risk | Missed until the next scheduled review | Automated re-evaluation as new data sources or types appear |
| Example blind spot | PTO agent guardrail blocks PII but not compensation data, unnoticed until audited | Gap surfaced automatically by correlating HR data types against guardrail scope |
See how this plays out across a full GenAI and LLM data control program. Not sure where your own GenAI or RAG pipelines stand? Explore Bedrock Data's approach to GenAI and RAG risk management before your next deployment.
How can organizations ship AI with confidence?
The broader message from the AWS article, and from Bedrock Data's work more generally, is not about slowing AI adoption. It's about making it safe to move fast.
By combining:
- Amazon Bedrock for scalable GenAI capabilities
- Data-aware governance grounded in DBOMs and guardrail validation
Organizations can deploy AI systems that are both powerful and trustworthy.
That intersection, where AI capability meets operational governance, is exactly where Pranava Adduri and the Bedrock Data team are focused.
The walkthrough was published in November 2025. ArgusAI by Bedrock Data has since extended the same DBOM to MCP server discovery and Snowflake Cortex, and exposes its governance intelligence to AI workflows through an MCP server of its own. The Amazon Bedrock workflow described above is unchanged.
Read the full AWS article
This post is a summary of ideas originally presented in the AWS Builder Center article:
Securely shipping GenAI systems in Amazon Bedrock with data-aware governance
By Pranava Adduri, Ishara Premadasa (AWS), and Jean Malha (AWS)
See your own DBOM and guardrail gaps on Amazon Bedrock
If you're running GenAI or agentic workloads on Amazon Bedrock, the fastest way to know whether your guardrails actually cover your data is to look at your own environment. Request a live demo to see your Data Bill of Materials and Guardrail Gap Analysis mapped against your actual Bedrock deployment.
FAQs
What is generative AI governance?
Generative AI governance is the set of policies, processes, and technical controls that oversee how generative AI models and agents are built, deployed, and monitored. On Amazon Bedrock, that oversight depends on knowing exactly what data a model or agent can access, not just reviewing the model in isolation.
What is a Data Bill of Materials (DBOM)?
A Data Bill of Materials (DBOM) maps a model or agent to every dataset it depends on: training and fine-tuning data, the source systems behind a knowledge base, and the data an agent retrieves during inference. It answers the question every security and compliance leader eventually asks: what data actually feeds this system, end to end?
What is Guardrail Gap Analysis?
Guardrail Gap Analysis correlates the data types identified in a DBOM against a model's active guardrail configuration to show whether sensitive data is properly blocked, explicitly allowed, or unintentionally exposed. It turns guardrail coverage from an assumption into something a team can verify.
What should I look for in AI governance solutions for GenAI systems?
Look for a solution that maintains a live DBOM as data changes, validates guardrails against actual data exposure rather than static rules, and extends coverage to non-human identities like agents and service accounts, since GenAI systems increasingly act without direct human input.
Why isn't a one-time AI governance review enough?
Data sources and data types change constantly, so a guardrail or DBOM that was accurate last month can miss new sensitive data introduced this month. Continuous, automated re-evaluation catches those gaps before they reach production, rather than at the next scheduled review.
