The Future of Agentic AI: Beyond Traditional Automation

Agentic AI replaces rule-based automation with software that reasons, plans and adapts. Traditional RPA follows a fixed script and breaks when a screen changes. Agents understand the goal instead. In production at UnitedHealth Group that shift delivered 60% productivity gains, multi-million dollar annual savings, and Snowboard running 3,200 bots at 99.7% uptime.

Play0:52

Agents put the intelligence in the software, not the script

Robotic process automation records a path of clicks and keystrokes and replays it. Agentic AI hands the steps to a model that understands the goal. The difference is where the intelligence lives: RPA puts it in the script, agents put it in the agent, which is why the same agent absorbs a change that would stop a bot dead.

Rule-based bots fail because they encode the interface, not the intent. A button moves, a field is renamed, a consent dialog appears, and the bot either fails loudly or quietly does the wrong thing. Every breakage is a maintenance ticket. Across a few thousand automations, the running cost of the estate starts to eat the savings that justified it. Tools like UiPath excelled at mimicking clicks. They could not understand context or handle the unexpected.

The architecture: a conductor and specialist agents

At UnitedHealth Group, where I lead enterprise AI, we built the Orchestra framework, part of CRYO, to coordinate specialized agents. The orchestrator owns sequencing, hand-offs and failure. Each agent owns one narrow job.

The workflow runs in three moves. An analysis agent turns raw case data into a structured picture. A decision agent applies the governing rules and policies and chooses what happens next. An action agent carries the decision out in the systems of record and reports back. The orchestrator passes work between them and decides what to do when any agent fails or returns something it does not trust. That separation is what makes the system explainable to an auditor and safe to change: each agent can be replaced, tested or constrained on its own.

What it delivered

In production we are seeing 60%productivity gains and Multi-milliondollar annual savings, and healthcare workflows that once took 10 minutes now complete in under 1 minute. The automation estate changed shape too. Snowboard, our in-house agentic platform, retired a Eight-figureRPA contract and now runs 3,200bots at 99.7% uptime.

What breaks first

Autonomy fails at the edges, not in the happy path: an agent that can act can also act wrongly, at machine speed. Four conditions apply before any agentic system goes near production: boundaries and validation on what an agent may do, observability into what it is doing, human review on critical decisions, and security treated as a first-class requirement.

Guardrails get literal once agents write and run code. That is what CAGE is for: a Rust sandbox built on defense in depth, independent layers of control so that no single failure reaches the infrastructure underneath.

The other discipline is restraint. An agent is not always the right tool. MNM, our 6.75M-parameter numeric transformer, decides prior authorizations far better than a general-purpose model would. Reach for the smallest thing that does the job.

Where this goes next

The systems that earn wider autonomy will be the ones that earn it incrementally. Cross-domain reasoning and adaptive learning are only useful inside a governance frame that a regulator, a clinician and an engineer can all read. Build the boundaries first, then widen them as the evidence comes in.