What KPMG Says Replaces RPA

The speaker opened with a story every team building AI agents will recognize: the early excitement of standing up a custom agent, wiring in a skill file, getting a demo working, and thinking this changes everything. Then reality sets in. Guardrails multiply. Behavior gets harder to predict. Models get updated or deprecated out from under you, and suddenly the thing you built three months ago doesn't scale to case two, let alone case two hundred.

The Problem With Describing an Entire Process in a Skill File

The core issue KPMG identified: teams are trying to cram an entire business process, every rule, every exception, every piece of required evidence, into a single skill file or prompt, and asking an agent to just figure it out from there. That approach has a ceiling, and most teams hit it faster than they expect. A skill file is a description. It's not a governed, centrally updatable definition of how a process is actually supposed to run.

The speaker's framing of the shift needed: "It's kind of like the map consists of the letter of the law, the agent consists of the spirit of the interpreter." A well-defined process gives an agent the letter of the law to work from. Without that, the agent is left interpreting spirit alone, on every single run, with no consistent reference point.

The Fix: Formalize the Process Once, Let Agents Execute Within It

KPMG's proposed answer is a new primitive sitting between raw agent flexibility and rigid, brittle automation: a formal process definition, covering the process itself, its rules, its exceptions, and what evidence needs to be captured. That definition gets exposed once through MCP, an API, or a CLI, so it isn't locked into any one tool or vendor.

From there, either a custom-built agent (drawing on its own data and skills) or an existing harness, the deck named Antigravity, Codex, and Claude Code specifically, can consume that same process definition. Every execution runs against defined guardrails, producing one of three outcomes: allowed, blocked, or routed to human review.

Five specific benefits came out of that structure:

• Define the process once, expose it everywhere it's needed, instead of redefining it inside every new agent or tool.
• Give custom agents and off-the-shelf harnesses a shared operating context, so behavior stays consistent no matter which one is running the task.
• Cut the ambiguity that comes from trying to describe an entire process inside a single skill file.
• Let agents choose how to execute the work, but only within clearly defined guardrails, not as an open-ended judgment call.
• Update the process centrally when policies or workflows change, instead of hunting down every place that process logic got copied.

Why RPA Is Crossed Out on the Slide

Traditional RPA (robotic process automation) appears on the slide with a line straight through it, and the reasoning is direct: RPA's failure mode was rigid, brittle scripts that broke the moment anything unexpected happened. The primitive KPMG is proposing isn't a smarter version of that same rigid approach. It's built specifically so an agent can use its judgment to find a path through the process, guided by defined rules rather than a fixed, unbending script.

The speaker's own description of what this is really for: something that acts as "that exoskeleton of the information worker." The goal isn't replacing judgment entirely. It's supercharging the people already doing the work, giving them a structure that's process-centric and organization-centric rather than yet another disconnected point tool.

The Real Lesson: Individual Agent Wins Don't Scale on Their Own

The most honest moment in the talk was the arc the speaker described: build one agent, get a great demo, feel like it changes everything, then discover that a process taking 10 to 12 weeks to build properly doesn't scale across every future use case the same way. Models get updated. Models get deprecated. Multi-agent behavior shifts in ways nobody planned for. Without a shared, governed process definition sitting underneath all of it, every new use case starts back at square one.

What This Means for Any Company Scaling Past Its First Few Agents

If your organization has one or two successful agent pilots and is now trying to figure out how to get to ten or twenty without rebuilding governance, guardrails, and process logic from scratch each time, this is the exact gap KPMG is describing. The fix isn't a smarter agent. It's a formal, centrally managed process definition that any agent, custom-built or off-the-shelf, can plug into and execute against consistently.

Scaling past your first few AI agent pilots and starting to feel the strain? Winsome helps companies build the shared structure that makes the tenth agent easier than the first. Talk to Winsome about your AI process strategy.