TCWGlobal is a contingent workforce management company that helps organizations manage and support workers across more than 150 countries. With a global footprint and thousands of active worker assignments at any given time, its operations and client-support teams depend on fast, accurate access to data to do their jobs.
The Apollo team, TCWGlobal's internal technology and data group, set out to build a suite of internal agents that staff could ask directly, in plain language, instead of waiting on a ticket. “Mini” was the first, and the most data-intensive. Halle Davis led the product and operations side; Jill Arldt handled the data infrastructure. The layer connecting the agents to TCWGlobal's systems runs on CData Connect AI.
Giving an AI agent live data access without exposing PII
One of the first challenges the team tackled was enabling self-service, and that required rethinking the technology setup. The current one wasn't built for it: Every time someone needed to see data in a new way, they sent a ticket. The answer typically came back in 48 hours, on a good week. When it didn't, a client was left waiting on a call, or a recruiter was blocked from doing their job. The answer was to build an internal agent. Getting the data layer right turned out to be the hardest part.
The Apollo team evaluated building the integrations in-house. TCWGlobal has the technical capability; writing direct API connections to each platform was possible. But an AI agent that stays useful means keeping its data layer current: tracking API changes, managing authentication, adding fields as needs evolve. The ongoing engineering cost of maintaining those connections was more than a two-person team could absorb.
The PII constraint sharpened the architecture question further. TCWGlobal handles significant amounts of personally identifiable information across its worker population. Any design that moved or replicated that data to feed an AI model created compliance exposure the team wasn't willing to accept. The agent needed to query live data without storing it—and it needed to surface only what each user was permitted to see.
"Our customer service agents had to search through three or four different systems just to find basic information about a worker or a client. We needed a faster, more direct line between our teams and the data."
— Halle Davis, Director of Apollo, TCWGlobal
A governed data layer across multiple systems for the AI agent
TCWGlobal found what they needed in CData Connect AI.
Pre-built connectors across all four operational systems, and a managed MCP server platform that exposes them to any AI agent in a standardized way, no custom integration work required on the AI side.
Derived views that let Jill pre-join and pre-filter data from each source, so the agent accesses a clean, controlled representation of the data rather than hitting raw production tables directly.
Role-based access controls that scope what each user can see at the connection level, enforcing the same permissions that already exist in the underlying systems without having to rebuild that logic.
The most time-consuming part of the build was developing the agent instructions, a markdown file covering TCWGlobal's data structure, naming conventions, and business-specific definitions. The agent needed to know what counts as an active employee, what direct staffed payroll means, and that 'companies' and 'customers' refer to the same thing. The same file also contained behavioral guardrails: limits on tool calls per query, record count caps, and rules on how to link across tables to avoid inflated results.
The resulting AI agent, named Mini, launched in stages. Mini runs on Claude, accessing data through Connect AI's MCP server and caching frequent results to keep response times fast. The full architecture, multiple source systems, a managed MCP server, a caching layer, and a natural-language interface, was built
“We wanted to use Connect AI for the connectors — we didn't want to spend our time maintaining changing API endpoints. But it also gave us a safe way to expose data to the agent: we could cache what we needed, build derived views to control exactly what it saw, and manage who had access to what. That let us focus on the logic of the agent instead of the plumbing underneath it.”
— Jill Arldt, Database Administrator, TCWGlobal
AI agent replaces multi-system data hunting
Mini can answer both data questions and process questions and the staff only has to ask in plain language and doesn’t have to issue a ticket or wait. Staff ask things like
"Who are our top 25 clients and what are their industries?" Mini pulls billing data and fills in information where the field is unpopulated, flagging the gap.
"Here's an email address. Can you find this person in our systems and tell me who they're on assignment with?" Mini searches across all four connected systems and returns a unified record.
"I have a worker moving from Canada to the US. How should I handle this?" Mini walks through the prerequisite steps and links to the relevant standard operating procedure (SOP) with its last review date.
The impact showed up quickly. A recruiter needed to process an urgent worker termination from their phone while out of office; no laptop, no VPN access. They asked Mini for the worker's details and engagement information and got them instantly. That kind of immediacy is now the norm. New hires get answers on day one without interrupting a senior teammate. Every question and answer is logged, giving the team visibility into what's being asked and a way to keep improving the instructions over time.
“People are really excited. I was looking at the questions being asked yesterday and I was blown away by the range of things people wanted to know—and the agent was answering them.”
— Halle Davis, Director of Apollo, TCWGlobal
Mini is one agent in a growing suite. The Apollo team is building agents across different business functions. The next one on the roadmap is an external-facing version that will give TCWGlobal's clients direct access to their own workforce data, without needing to contact the team for answers.
CData Connect AI: The data layer that makes AI agents work in production
Building an AI agent that works in production starts with the data layer underneath it; live, governed, and scoped to what each user should see. CData Connect AI makes that possible without replicating sensitive data or building custom integrations from scratch.
