The Three Levels of Self-Service AI: From Document Extraction to Multi-Agent Systems 

Continued from the Databricks World Tour, Washington, DC – December 11, 2025

At the Databricks World Tour on December 11, government agencies and international organizations described how their business users are building production AI systems in weeks. In the previous post, I covered the old IT bottleneck model and the breakthrough that let business users build AI themselves. Now let’s look at the three levels of complexity organizations are working at.

 

Three Levels of Self-Service AI  

Throughout the conference, a pattern emerged in the case studies and demonstrations. Organizations are building AI at three different levels of complexity, all with business users in the lead. 

• Level 1: Information extraction—reading documents at scale to pull out specific data. Jonathan, a Databricks solutions engineer, demonstrated this with a fictional fraud detection scenario. The agent reads claims documents and extracts key fields automatically. It’s ”gloriously boring,” as one presenter called it, but Centers for Medicare & Medicaid Services (CMS) used exactly this approach to flag $106 million in improper payments.
• Level 2: Knowledge assistant—making documents searchable in natural language. IDB Invest deployed this for their investment officers. The bank has decades of economist reports and field analyses sitting in PDF files across SharePoint sites. Now investment officers can ask questions and get instant answers pulling from that institutional knowledge. “No one truly had the opportunity to access that information before”, Igor said.
• Level 3: Multi-agent systems—combining multiple capabilities into one AI that can handle complex workflows. This is where CMS and IDB Invest are seeing the biggest impact.

At CMS, 83% of employees now use AI daily. Patrick Newbold showed how an executive can ask a multi-agent system for a briefing on the fraud detection program. The system queries their fraud database for the latest numbers, searches policy documents for relevant procedures, checks external sources for emerging fraud trends, then synthesizes everything into a briefing. “Something that would have taken days before from my team now takes seconds,” he said.

The agents respect governance automatically. A junior analyst sees masked personally identifiable information. A senior analyst sees the full data. Same agent, different permissions based on who’s asking.

 

How This Actually Works

The technical details matter because they explain why this is possible now when it wasn’t two years ago.

Databricks provides pre-built agent patterns. Instead of starting from scratch, organizations pick a pattern—information extraction, knowledge assistant, multi-agent supervisor—and customize it for their needs. The agents integrate with Genie for querying databases, can read PDFs and documents, and can connect to external data sources. Business users configure the agents through a graphical interface and provide examples of what good output looks like. Then domain experts review the results and give feedback, and in turn the agents learn and improve automatically based on that feedback. All of this runs on the same data platform, with the same governance built in. The agents can only see what the user building them is allowed to see. There’s no separate security review for each agent because they automatically inherit the policies that are already set up in Unity Catalog.

“We don’t have big buckets (of money),” Igor said, explaining why IDB needed cost controls built in. The platform manages that automatically, scaling resources up when needed and shutting them down when they’re not in use.

For government users, FedRAMP and IL5 compliance are already built into the platform. So, when CMS builds an agent, it’s automatically running in an authorized environment; no extra steps are required.

 

The Numbers Behind the Change

IDB Invest’s three-year transformation produced measurable results:

• 85% increase in people using their analytical systems
• 60% faster delivery on data and AI projects
• Seven-figure cost savings from automation
• Changed their hiring practices—job descriptions now require Databricks experience

At CMS with 160 million beneficiaries:

• 83% of employees using AI daily
• Average of 6 hours saved per week per employee
• $106 million in improper payments identified
• Executive decisions that took days now take seconds

These are not pilot projects, they’re production systems processing real healthcare claims and managing real investment portfolios.

 

What IT Actually Does Now

This might sound like IT departments are being cut out. The opposite is true; their roles hasn’t disappeared, but it has fundamentally changed. IT at IDB Invest provides platform infrastructure, sets up the lakehouse architecture, configures the Unity Catalog with appropriate governance policies, manage the cloud environment and ensure compliance, and train business users on how to build agents safely.

What they don’t do anymore is build every single report, dashboard, and analysis that the business requests. That work has moved to the business units themselves. “IT was the enabler for us,” Kapil said. “I have never seen that in the last 20 years of my career.”  The transformation required IT to let go of building everything while taking on the harder job of building platforms that others can safely use. Done right, IT becomes more valuable, not less, because they’re multiplying the capability of the entire organization.

 

Coming Next

The obvious question remains, how do you maintain security and quality when business users are building AI systems? In the next post, I’ll examine how organizations handle governance, quality control, and the cultural shift required to make this transformation work.

In the meantime, checkout Analytica’s Insights page for more blogs and case studies!