What are data agents? The bridge between agentic AI and enterprise data.

Data agents bridge agentic AI with trusted enterprise data, empowering systems to answer, decide, and act with enterprise-grade precision.

1. Data agents are specialized AI agents that connect agentic AI systems to enterprise data, retrieving and updating data while enforcing governance. 

2. Most agentic AI projects fail in production not because of LLM capabilities, but because enterprise data for AI agents is inaccessible. 

3. Minimum Viable Data (MVD) is a core principle where data agents provide only the smallest, most relevant dataset needed for each task – reducing AI costs, by limiting token usage, and increasing productivity, by delivering greater accuracy. 

4. Entity-centric data products organize information around business entities (customers, orders, devices) rather than applications, providing the unified foundation data agents need.

5. Data products implement agentic AI data protection by automatically applying access controls, data masking, audit logging, and compliance rules. 

6. Organizations can deploy pre-built data agents for common scenarios or build custom agents with commercially available tools to address specific needs. 

7. Data agents provide business context and autonomously determine what data to retrieve from which systems, through chain-of-thought reasoning. 

8. Data agents support multiple use cases including conversational AI chatbots, back-office workflow automation, synthetic data generation, risk monitoring, and compliance auditing.

Artificial intelligence is changing how businesses work. Companies are building AI agents that can reason, make decisions, and take actions on their own. These AI agents promise to transform customer service, speed up operations, and help businesses make better decisions.

But there's a problem that many companies face when they move from testing to real production. Their AI agents can’t easily access and update the data they need from enterprise systems. The agents may be smart, but they just can't reach the right information at the right time.

This is where data agents come in. Data agents act as the bridge between AI agents and your company's operational data by governing the information for completeness, compliance, and context – in short, data readiness. 

Think of it this way: An AI agent is like a skilled doctor who diagnoses illnesses and recommends treatments. But without access to a patient's medical records, test results, and current medications, even the best doctor can't help. Data agents are like medical assistants who gather all the right information and present it clearly, so the doctor can diagnose and treat with greater efficacy. 

Unlike chatbots designed to answer simple questions, AI agents can plan multiple steps, work with different tools, and complete complex tasks without a human in the loop. 

Consider AI customer service. To help a customer with a billing question, the agent may need to know about the user's account status, recent payments, active services, and billing history. This information sits in different systems: CRM platforms, billing databases, payment processors, and customer data warehouses. Traditional data architectures make it hard for AI agents to quickly form a complete picture.

What are the 4 main agentic AI challenges?

When  AI agents can't access the data they need, 4 major problems emerge: 

1. Accuracy issues 

   When an agent receives incomplete, outdated, or conflicting information, it tends to make mistakes. For example, the customer service chatbot relying on that agent might reply that a payment was made when it wasn't, or suggest a service upgrade that the user already has. 

2. Slow responses 

   It takes too long for an agent to call multiple systems and unify data. With conversational AI, delays of even a few seconds feel awkward and frustrating. 

3. Rising costs 

   Large Language Model (LLM) usage fees are based on the amount of data processed. When LLM agents receive more information than they need, costs multiply quickly, especially when handling thousands of requests per day.

4. Security and compliance risks 

   Giving AI agents access to entire databases needlessly exposes sensitive data – and makes compliance with CPRA, HIPAA, GDPR, and DORA European regulations so difficult.

Data agents are specialized AI agents that focus on one critical job: connecting other AI agents to enterprise data. While an AI agent handles reasoning and decision-making, a data agent handles data access and operational actions.

Data agents work inside your company's agentic AI framework. On one hand, they interact with AI agents, understanding what data is needed for each task. On the other hand, they connect to data products and enterprise systems, retrieving information and executing actions. 

How are data agents different from traditional tools?

Unlike traditional data integration tools or APIs, data agents use AI to:

• Understand a natural language request and figure out what data is needed 

• Determine which systems or data sources contain that information 

• Generate queries or API calls to retrieve the data 

• Apply AI data governance and security rules automatically 

• Transform and assemble the data into a useful format 

• Execute updates or actions on systems of record when needed 

For example, when a customer service AI agent asks, "Why is this customer's bill higher this month?", the data agent would: 

1. Identify the customer being discussed 
2. Comprehend it needs billing information (current and previous bills, usage data, plan changes) 
3. Retrieve this information from the relevant data products 
4. Mask sensitive data on the fly 
5. Format the response so the AI agent can understand it clearly 
6. Deliver only the essential information needed to answer the question 

This all happens in seconds, without requiring someone to write custom code or create new API endpoints.

A key principle that guides data agents is called Minimum Viable Data (MVD), or providing the smallest, most relevant, dataset needed to perform a specific task.

You might think that giving AI agents more data would lead to better results. Actually, the opposite is true. Too much data creates noise, slows down processing, increases costs, and raises security risks.

MVD isn't about providing less data. It's about providing exactly the right kind and amount of data, in terms of: 

1. Task relevance

   Only information directly needed for the current task is accessed and used. 

2. Context richness

   Metadata and semantic structure help the agent understand what the data means.

3. Data minimization

   Exposure to Persoanlly Identifiable Informatikomn (PII) and other sensitive data should be severly restricted to reduce privacy and compliance risks.

For instance, if a customer is calling about an Internet speed upgrade, the AI agent doesn't need a 5-year history of support tickets. It needs to know the current plan, available upgrade options, contract terms, and billing information. That's the minimum viable data required for this task.

Why context matters as much as content 

Minimal viable data without context lacks meaning. Rich metadata adds context by providing:

• Meaning to each data field 

• Understanding about the relationships between the different data elements  

• Valid values and data types 

• Business rules and constraints 

• Data quality indicators 

MVD calls for minimum size but maximum understanding.  The data agent ensures that even a small dataset carries all the meaning needed to reason accurately.

For data agents to work well, they need access to data organized in a specific way. This is where the entity-centric data product comes in, another crucial element of agentic AI best practices.

Traditional data systems organize information around applications or databases. An entity-centric approach organizes data around business entities such as a customer, order, invoice, or device.

What is an entity-centric data product? 

A data product is a self-contained unit that manages all the data for a specific business entity. For example, a customer data product brings together everything about a customer: their profile information, account details, purchase history, support interactions, preferences, and more.

Each data product: 

• Retrieves data from all relevant source systems 

• Syncs information in real time 

• Includes semantic metadata and business rules 

• Enforces governance and security policies 

• Provides a unified view of that entity 

This synergy creates a live, reliable operational view that data agents can depend on. 

Why are traditional data architectures inadequate? 

Many companies try to support AI agents using data lakes or APIs. Both have significant limitations: 

Data lakes store huge amounts of information, which is great for analytics. But for agentic AI, they fall short due to: 

• Stale data updated only in batches

• Scattered information about each entity

• Too much noise and irrelevant data

• Lost meanings and relationships

• Security risks posed by accessing entire data lakes

APIs are essential for system integration, but they can't serve as the primary data foundation for AI agents because: 

• They're designed around how systems store data, not how businesses think about entities. 

• Getting a complete picture requires calling many different APIs, which is slow and tedious.

• You can’t build enough API endpoints to answer every possible question.

• APIs expose fields without their semantic meanings or relationships.

• It’s impractical to create a new API for each new agent question. 

Entity-centric data products solve these problems by providing unified, real-time, semantically-rich views of business entities.

Let's walk through a real example to see how data agents operate. Imagine a telco with an AI chatbot for customer service. 

A customer calls and says, "My bill is $50 higher than usual this month. Why is that?"

The customer service AI agent fielding this question needs to provide the right answer quickly. The step-by-step workflow might look something like this: 

1. Understand the request 

   The AI agent understands this is a billing question that requires comparing the current month's charges with previous months. 

2. Invoke the data agent 

   The AI agent asks the data agent for the necessary billing information for this specific customer. 

3. Decide which data is most relevant 

   The data agent applies chain-of-thought reasoning to determine what information is required, for example: 

   – Current month's bill with line-item details

   – Previous month's bill for comparison 

   – Any recent plan changes or add-ons 

   – Usage information if relevant 

   – Active services and features 

4. Retrieve the data 

   The data agent accesses the customer's billing data product, which unifies information from multiple systems, such as the billing platform, usage databases, plan management, and payment systems. It then generates the appropriate queries and retrieves the data. 

5. Apply governance 

   The data agent automatically applies any necessary data masking. For instance, it might hide payment card details while keeping transaction amounts visible, ensuring proper agentic AI data protection. 

6. Assemble the response 

   The data agent packages the information in a structured format that includes: 

   – Previous month: $85 total 

   – Current month: $135 total 

   – Difference: $50 increase 

   – Cause: Roaming charges ($45) + activation fee ($5) 

   – Context: Customer traveled abroad with active roaming

7. AI agent response 

   The AI agent receives this clear, contextual information and can now explain to the customer: "I see your bill increased by $50 this month. That’s because you used international roaming while traveling, which added $45, plus a one-time activation fee of $5." 

The bottom line: The customer receives a clear, accurate, and satisfying answer in a matter of seconds.

Security and compliance are critical concerns when AI agents access enterprise data. Data agents play a key role in maintaining security, making them essential for agentic AI data protection. 

Built-in governance 

Data agents enforce governance rules automatically. They apply: 

• Access controls based on who is requesting the data 

• Data masking to hide sensitive fields 

• Audit logging to track what data was accessed and why

• Privacy rules to ensure CPRA, HIPAA, GDPR, or DORA compliance. 

For example, if a junior customer service representative asks about a particular customer's account, the data agent might mask certain financial details that only senior staff can see. All of this happens transparently, without requiring the AI agent or the user to understand the complexity behind the security rules. 

Limiting data exposure 

By implementing minimum viable data principles, data agents significantly reduce risk. Instead of giving AI agents access to entire databases, data agents provide only what's needed for each specific task.

Leading consulting firms emphasize treating agents like new employees, giving them access only to what they need. Data agents make this practical by automatically determining and enforcing data minimization.

Monitoring and auditing 

Data agents create clear audit trails showing: 

• What data was requested? 

• Why was that data requested?

• Which systems provided it?

• How was the data transformed?

• Who accessed it?

• When was it accessed?

This visibility is essential for compliance and for investigating any issues that may arise.

Organizations can take different approaches to implementing an LLM agent architecture, depending on their needs and technical capabilities. 

Pre-built data agents 

For common use cases, pre-built data agents offer the fastest path to value. These agents handle standard scenarios like: 

• Answering operational questions 
• Generating synthetic test data 
• Auditing for compliance 
• Retrieving and transforming data 

Pre-built agents can be configured to work with your specific data products and systems without custom coding. 

Custom data agents 

For specialized needs, organizations can build custom data agents. Modern platforms provide low-code or no-code builders that make this accessible to business-oriented teams, not just developers. 

Creating a custom data agent typically involves: 

1. Defining the purpose 

   What business problem will this agent solve? What questions should it answer or actions should it perform? 

2. Identifying data sources 

   Which data products or systems contain the relevant information? 

3. Configuring the logic 

   How should the agent reason out data requests? What are the rules for data access and transformation? 

4. Setting governance rules 

   What security and compliance policies must the agent enforce? 

5. Testing and refining 

   Which scenarios should the agent be run through to ensure it works correctly and securely? 

Integration with agentic frameworks 

Data agents operate inside your organization's broader agentic AI framework. This might be built on platforms like Microsoft Copilot Studio, Salesforce Agentforce, AWS Bedrock, or open-source frameworks.

The key is that data agents work alongside other agents, including LLM agents, each specializing in different aspects of the workflow. While some agents handle conversation, others make decisions, and data agents ensure everyone has access to the right information.