Short answer
Agentic AI uses AI agents that plan multi-step work, call tools and read your data to complete tasks, instead of answering one prompt at a time. What decides whether agents succeed is not the model but the data and architecture around it: governed access to databases and knowledge bases, identity and permissions for every action, observability, and standard ways to connect tools. Start with a narrow use case and put governance in before scale.
For years, AI behaved like an isolated specialist: good at one task, with little awareness of the wider business. A model answered questions, another scored churn, another suggested code, and each lived apart from the data and tools your team uses. Agentic AI changes that pattern. Agents plan multi-step work, call tools and read your data to complete a task. This post explains what that means for your data and architecture. We rewrote it in November 2026. We removed statistics and product descriptions that had gone out of date, and every vendor claim below is tied to a current source.
Where the market stands
Gartner predicts that by 2027, 50% of business decisions will be augmented or automated by AI agents for decision intelligence. Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027. Both can be true: the opportunity is real, and many projects fail on cost, unclear value or weak controls. The projects that survive tend to start narrow and have the data foundation in place first.
Integration is a recurring obstacle. MuleSoft's 2026 Connectivity Benchmark Report found that 96% of respondents agree AI agent success depends on seamless integration, and Salesforce's summary of the report says that 27% of applications are connected together.
What agents need from your data
- Governed access to structured data. Agents that answer questions from your databases need clear definitions, quality standards and permissions. This is data governance work. See enterprise data management and data governance for associations.
- Access to documents and policies. Databases hold facts. Policies, procedures and manuals live in documents. Retrieval-augmented generation (RAG) improves answers by retrieving relevant information from your own data sources. Amazon Bedrock Knowledge Bases is AWS's managed way to integrate proprietary information into generative AI applications this way.
- A standard way to connect tools. The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external systems, including data sources, tools and workflows. It reduces the custom work of wiring each tool to each agent.
- Identity, memory and observability. Agents act on someone's behalf, so each action needs an identity and permission scope, and teams need traces of what the agent did.
An example architecture on AWS
AWS describes Amazon Bedrock AgentCore as a platform for building, deploying and operating agents securely at scale, using any framework and any foundation model. Its documented components include a runtime, memory, identity, a gateway for tools, a browser, a code interpreter and observability. Per AWS documentation, AgentCore Runtime runs each user session in a dedicated microVM, supports MCP and A2A protocols, and allows sessions of up to 8 hours on microVMs.
A typical pattern combines:
- Foundation models, hosted on Amazon Bedrock, including Anthropic's Claude models, or other providers.
- Knowledge Bases for curated internal content.
- Data services, such as Amazon Redshift, Amazon RDS and DynamoDB, for the facts an agent needs.
- Lambda and Step Functions to expose business logic as callable actions, with API Gateway and EventBridge for event-driven orchestration.
- IAM and logging for permissions and an audit trail.
AWS has published examples of this pattern, such as a post on structured knowledge bases with Amazon Bedrock and Amazon Redshift and healthcare posts on generative AI for patient outcomes and AI-powered patient profiles. Treat those as vendor examples, not benchmarks of results.
Choosing an agent framework
Frameworks differ in how much control they give you. AgentCore is framework-agnostic: AWS lists support for LangGraph, CrewAI, LlamaIndex, Google ADK, OpenAI Agents SDK and Strands Agents, among others. Check a framework's status before you adopt it. For example, the AutoGen repository states that AutoGen is in maintenance mode, limited to bug fixes, security patches and documentation, and names Microsoft Agent Framework as its successor.
Security and governance first
Agents widen the blast radius of a mistake, so governance has to come first:
- Scope permissions so an agent can read and act only on what it is authorized to. AWS provides IAM and AgentCore Identity for this.
- Filter inputs and outputs. Amazon Bedrock Guardrails offers content filters, including a prompt attack category, denied topics, sensitive information filters and contextual grounding checks to detect responses not grounded in source material.
- Log and trace. Keep an audit trail of what agents did and why.
- Keep humans in the loop for sensitive or irreversible actions.
- Plan for cost of failure. IBM's Cost of a Data Breach Report 2026 puts the global average cost of a breach at USD 4.99 million, a 12% increase over the prior year.
For healthcare organizations, any agent that touches protected health information needs a HIPAA business associate agreement and controls in place before it runs. Associations should apply the same discipline to member data.
How to start
- Pick one narrow, valuable use case with a clear owner and a measurable outcome.
- Check that the data it needs is accurate, documented and permissioned. Fix that first.
- Build with scoped identity, logging and guardrails from day one.
- Measure results against a baseline, and expand only when they hold.
Where Cruz Street helps
Cruz Street designs data foundations and AI-ready architectures on AWS. Our data engineering services and AI and data science services cover the data layer, integration and models, and we work with associations and standards organizations and healthcare organizations. If you need a strategy before a build, an on-demand chief data officer can set the roadmap. Contact us to map agentic AI onto your data.
Sources
- Gartner, Top Data & Analytics Predictions, June 17, 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-17-gartner-announces-top-data-and-analytics-predictions
- Gartner, Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, June 25, 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- MuleSoft, 2026 Connectivity Benchmark Report: https://www.mulesoft.com/lp/reports/connectivity-benchmark
- Salesforce, 2026 Connectivity Report announcement: https://www.salesforce.com/news/stories/connectivity-report-announcement-2026/
- Model Context Protocol, What is MCP: https://modelcontextprotocol.io/docs/2026-07-28/getting-started/intro
- AWS, What is Amazon Bedrock AgentCore: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html
- AWS, AgentCore Runtime: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agents-tools-runtime.html
- AWS, Amazon Bedrock Knowledge Bases: https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html
- AWS, Amazon Bedrock Guardrails: https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html
- Microsoft, AutoGen repository (maintenance mode, successor Microsoft Agent Framework): https://github.com/microsoft/autogen
- IBM, Cost of a Data Breach Report 2026: https://www.ibm.com/reports/data-breach
Frequently asked questions
What is agentic AI?+
Agentic AI refers to AI systems, called agents, that reason through multi-step tasks, use tools and data sources, and take actions, with human oversight set by policy. It differs from a single prompt-and-response chatbot because the agent decides which steps and tools to use to reach a goal.
What is the Model Context Protocol (MCP)?+
MCP is an open-source standard for connecting AI applications to external systems such as data sources, tools and workflows. It gives agents a common way to access context and invoke actions, so each tool does not need a custom connection to each AI application. Source: modelcontextprotocol.io.
How do AI agents use enterprise data safely?+
Through scoped identity and permissions, so an agent can only read and act on what it is authorized for, plus logging, guardrails and human approval for sensitive actions. AWS describes AgentCore as providing agent identity, memory, observability and secure runtime isolation for this purpose. Source: AWS, What is Amazon Bedrock AgentCore.
Will agents really make business decisions?+
Gartner predicts that by 2027, 50% of business decisions will be augmented or automated by AI agents for decision intelligence. Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027, which is why governance, measurement and a narrow first use case matter. Sources: Gartner press releases of June 17 and June 25, 2025.

