From FDC3 to MCP: How to Make Desktop Workflows AI-Ready

A practical guide to making enterprise workflows AI-ready using MCP and FDC3 — covering composite intents, better metadata, and how to turn interoperability standards into reliable AI actions.

At the Open Source in Finance Forum (OSSF) 2025, Robert Myers (Head of io.Intelligence) and Kalin Kostov (Lead Engineer) shared valuable lessons from six months of tackling the challenge of deploying AI copilots in enterprise environments.

Their session, “FDC3 and AI: Using FDC3 Intents as MCP Tools”, revealed how FDC3 intents and MCP — if mapped correctly — have the potential be the bridge that beings AI into practical, reliable enterprise workflows.

FDC3 Intents as the Foundation for AI

Their central premise challenges the way most organizations think about bringing AI into enterprise systems. Rather than building new AI-specific integrations from scratch, firms that have already implemented FDC3 already have the foundation they need.

FDC3 intents — the standardized, action-oriented definitions of what applications can do — can be mapped to AI-driven actions, but not blindly.

Every FDC3-enabled application represents a potential action an AI system can take, but making this work requires intentional design and an understanding of the limitations of current LLM tool handling.

The Core Challenge: From Intents to Reliable AI Actions

The biggest hurdle is bridging the gap between an LLM’s nondeterminism and the enterprise’s need for predictable outcomes. For decades, business software has been built around determinism: same input, same output.

AI is different. Models make autonomous choices about which action to invoke, in what order, and with what data. Users also phrase requests differently, which increases variability. Even so, enterprise workflows require deterministic outcomes.

FDC3 provides structure, but only if the intents are designed and presented to AI correctly.

The Golden Rule: Composite Intents Over Atomic Ones

One of the most counterintuitive findings in the presentation contradicts traditional software engineering norms. Architects typically favor small, atomic components that each do one thing well. But exposing dozens of tiny FDC3 intents directly to AI creates problems.

The model becomes confused and may pick intents randomly. It also may skip required steps. The solution is simple: design composite workflow-level capabilities, not dozens of tiny ones.

Instead of exposing these individually:

ViewInstrument
GetPortfolio
GetPositions
ViewChart
GetAnalysis

Create one composite workflow definition, such as:

AnalyzeClientPortfolio

This approach dramatically reduces nondeterministic behavior while lowering cost and token usage.

Enriching FDC3 with AI-Ready Metadata

FDC3 application definitions are essentially the instruction manual for AI systems. Their quality directly determines whether AI behaves intelligently or gets confused. High-quality FDC3 definitions should include:

detailed application descriptions
complete intent definitions with clear purpose and expected behavior
rich context types using standard FDC3 schemas (or well-documented custom ones)
explicit result types
examples of typical usage
publicly accessible schemas

Vague intent names, generic descriptions, or missing schemas make intents “close to useless” for AI integration, since they put the model in a position where it has to guess. Investing in richer metadata creates safer, more reliable AI behavior.

Mapping FDC3 Intents to MCP Tools

The Model Context Protocol (MCP) provides the structure for presenting FDC3 capabilities to AI. But mapping is not automatic. There is a lot of noise out there about how these two standards can magically work together. Yes, you can make it look flawless in a demo by just giving a couple of tools and exact instructions to the agent. But in practice — where an enterprise has hundreds or even thousands of intents — selecting the right one turns into a “magic eight ball” where you don’t know what you’re going to get.

You cannot simply expose every FDC3 intent as an MCP tool and expect reliable behavior.

Successful deployments instead:

identify user workflows
map workflows to supporting FDC3 intents
design composite MCP tools around those workflows
add descriptions detailing when and how each tool should be used, which intents it touches, and what it accomplishes

When organizations expose too many similar-sounding intents without curation (for example, ViewInstrument, ViewChart, ViewContact), the AI becomes confused and inconsistent.

The Demo Results: Theory Meets Practice

The OSSF presentation puts this theory to the test in a demo showing the different AI results to the different approaches.

Given the prompt: Review John Smith’s portfolio, identify his largest holding, and show recent price movement.

With workflow-level MCP tools, the system produced consistent results and predictable steps with lower token costs.

However, with dozens of atomic intents exposed, the system produced:

different results every time
random tool selection
unnecessary or missing steps
higher cost

One approach reflected real workflows; the other treated intents as an undifferentiated list of actions. You can see the AI was not able to complete the task successfully with dozens of FDC3 intents.

The Bottom Line

For organizations just beginning to deploy AI across their workflows, the advice is clear: commit to standards like FDC3 that are metadata-rich and designed for interoperability (learn more about interop & AI). Rather than proliferating copilots across individual applications and vendors, establish a single application fabric that allows governed, secure action across your enterprise.

For firms already using FDC3, you are significantly ahead — but must design the AI layer intentionally and strategically.

Organizations that treat FDC3 intents as first-class inputs to their AI strategy are already seeing faster deployments, lower costs and better user experiences. Metadata enrichment and thoughtful tooling pay dividends in safety, consistency, and user satisfaction.

FDC3 & MCP: The Right Questions to Ask

As you evaluate your approach, consider:

Which FDC3 intents should be exposed as MCP tools?
Which should not?
What metadata enrichment is required for AI readiness?
How do you test and validate deterministic AI behavior across your workflows?
How should workflow-level tools be designed to keep the AI predictable and cost-efficient?

Organizations that address these questions early will be best positioned for success.