A couple days ago one competitor said that because we were trying to make AI Agents "follow a workflow," we somehow lacked foresight. Of course,  it got me thinking. 

On the surface, it may seem like a no-brainer. If AI can reason, plan steps, call tools, ask questions, and correct itself, why constrain it with predefined workflows? Why not just give an agent a goal and let it figure out how to get there? 

But the more time we spend deploying AI inside real insurance operations, the more obvious it becomes that the workflows vs agents framing is the wrong debate altogether. 

Because both workflows and agents have their merits, and it’s when their powers are combined that we get to build something outstanding.

So instead of arguing if AI agents are better than AI workflows, maybe we should agree on the power of agentic workflows instead.

What each side gets right

Workflows give guarantees: every step is defined, every output is checkable, and every run is repeatable. The cost, of course,  is upfront design effort and brittleness, when inputs vary in ways you didn't anticipate.

Agents, on the other hand, handle variability beautifully. An orchestrator picks the right tool, loops back when something feels off, adapts to inputs you never imagined. But they’re a lot less predictable and a lot more challenging to audit. 

So, in a way, both AI workflow and AI agent proponents are right about the benefits of their approach. But, at the same time, they’re both wrong about what should win. 

‍In regulated industries like insurance, you can't pick agents and shrug off predictability, because decisions carry financial, legal, and governance weight. A wrong underwriting call ripples across a portfolio, a balance sheet, a regulator's report. So the cost of a bad decision is not just operational; it's reputational and legal too.

But you also can't pick pure workflows and shrug off variability. Because every now and then loss runs arrive in formats no one designed for or submission reference exposures that a runbook doesn’t cover.

If a workflow is too rigid, it breaks the moment reality kicks in.

‍

‍

The architecture: Workflow first, orchestrator on top

This is why our approach is to run the workflow as the default execution path (deterministic, fast, and predictable) and layer an orchestrator agent on top (not as the executor, but as the reviewer). 

The orchestrator audits the workflow's output and decides one of three things:

↳ Ship it ↳ Make a small correction ↳ Escalate to a human

This is the core architecture. The workflow does the work, and the orchestrator decides whether the work is good enough.

Three workers: Junior, manager, boss

The clearest way to think about agentic workflows is to imagine you're orchestrating three workers.

The workflow is a junior IC. Hand them a runbook and they execute it fast, consistently, exactly as written. They don't deviate, and predictability is what makes them trustworthy.

‍The orchestrator is a manager. They don't redo the junior's work by default but review it. If something looks off, they decide whether to make a small correction themselves or escalate it to the boss.

‍The human is the boss. They handle the genuinely ambiguous calls: the decisions that require judgment, context, or accountability that no system should take on alone.

This maps directly onto how real operations teams already work: SOPs, exception handling, escalation paths. It is the operating model regulated industries already run on, mapped into the AI architecture.

The compounding loop

Good managers don't just review work, they notice patterns. And if they notice their junior workers tripping on the same thing over and over again, they update the runbook. This is how an exception becomes a part of a standard process. 

That's the real unlock in agentic workflows. 

Every orchestrator intervention becomes a signal. At a batch level, you can see what the orchestrator has been doing across thousands of runs and feed those patterns back into the workflow. As a result, the workflow gets smarter, the orchestrator escalates less and less, and the system compounds with usage.

The workflow on day one is not the asset, but the loop that turns every orchestrator intervention into a workflow improvement is.

Why this matters for regulated industries

For regulated industries, agentic workflows are the best way to go. 

Deterministic execution gives you the auditability and repeatability that compliance demands: every step is traceable; every decision has an owner, and every output can be replayed.

The agentic reviewer catches the variability that pure workflows miss. And when inputs deviate from the runbook, the orchestrator catches it before it ships, not after. And with the escalation path built in, the human safety net handles the genuinely ambiguous calls.

What’s critical is that the system improves itself without compromising the guarantees: the workflow stays deterministic, the orchestrator stays in a review role, while the human stays in control. And because every exception is a signal, the runbook gets better every week. 

Pure agents can't offer this. The execution path changes every run, which makes auditability hard and improvement diffuse. And pure workflows can't offer this either. They don't know what they don't know, so exceptions may slip through without anyone noticing.

Agentic workflows are the architecture that gives you both: the rigor of a workflow and the adaptability of an agent, in a system designed to learn.

‍

DISCLAIMER 

This article is for general informational purposes only and does not constitute legal, regulatory, compliance, underwriting, or other professional advice. The content reflects information available as of the date of publication, and FurtherAI undertakes no obligation to update it as laws, regulations, or AI technologies evolve.

‍