Agentic AI Sounds Exciting, But Workflow AI Actually Gets the Job Done

Talk to enough people building AI products, and the word "agent" starts showing up everywhere. Everyone wants to build agents. Investors ask about agents. Slide decks promise agentic intelligence that "acts autonomously."

But here’s a basic question. Is that even what we want in high-stakes, highly regulated work like anti-money laundering (AML) and know-your-customer (KYC) operations?

The reality is that most AI being used in this space is not agentic. It is workflow-based. And that is not a failure. It is a better fit. Because the work analysts are doing is not agentic either.

What Even Is an Agent?

Anthropic, the AI research company behind the Claude models, published a post in December 2024 that laid out a clear distinction between workflows and agents. It is still one of the best explanations out there.

In their view:

• A workflow is when an AI follows a predefined series of steps. Everything is structured ahead of time.
• An agent decides for itself how to proceed. It plans, selects tools, evaluates outcomes, and adjusts course based on what it learns.

Agents are more flexible. They can be more powerful in the right context. But they are also harder to control, more expensive to run, and more prone to unpredictable outcomes.

Human Analysts Follow Steps

Before asking AI to be agentic, we should ask whether the humans doing the work are.

In practice, AML and KYC analysts are not given a lot of autonomy. They do not get to decide how to run investigations from scratch. Their work is shaped by policy, regulation, and oversight. Their decisions are structured around checklists, playbooks, and escalation rules.

It is not that there is no judgment involved. There is. But the space in which judgment is exercised is narrow and heavily constrained by compliance procedures.

Workflow AI Actually Matches What Analysts Do

The most useful AI in this space handles the same kinds of structured, repeatable tasks that analysts do every day. A typical analyst might:

• Review flagged transactions
• Conduct internal database checks
• Perform internet research
• Document findings using predefined formats

While there are decisions made along the way, they happen within a controlled framework. Analysts are not improvising. They are following detailed procedures that prioritize consistency, auditability, and defensibility.

This is where workflow-based AI shines. You can chain together LLM calls, add validation steps, and build systems that support human analysts without stepping outside the bounds of what regulators require.

There is limited need for creative planning or open-ended reasoning. What matters most is speed, accuracy, and clear records of what was done. Workflow AI delivers on that.

So Why Do So Many Companies Claim They’re Building Agents?

Because it plays well to investors and in marketing.

"Agent" suggests intelligence, autonomy, and innovation. It sounds more advanced than saying your product follows a script with some API calls and a couple of programmatic checks. If you are raising money or trying to differentiate in a crowded market, "agentic AI" is a stronger pitch.

You could say,"don’t hate the player, hate the game." But honestly, I don’t hate either.

Companies are responding to what the ecosystem rewards. Investors want to back what sounds like the next big leap. Marketers need language that stands out. Calling a system an agent is often a branding move more than a technical truth.

And that is fine. As long as the thing works, it does not matter what you call it. In practice, most of what is labeled agentic AI today is a well-structured workflow underneath. In AML and KYC, that is not a compromise. It is exactly what these systems need to be.

What Would Real Agentic AI Look Like in AML Compliance?

Maybe one day, agents will be able to run an entire investigation from end to end. They will decide what data to pull, plan their own research, and write up suspicious activity reports that are ready for submission.

But even then, full autonomy will be hard to justify. Compliance demands structure, traceability, and human review. You need to be able to explain every decision. You need to be able to say why something was flagged or cleared.

Autonomous agents may someday play a role, but only with guardrails, oversight, and constant evaluation. For now, and likely for a long time, structure is not a limitation. It is the core of how this work gets done.

Conclusion: Build for the Work, Not the Hype

There is nothing wrong with workflow AI. In fact, it is the right match for how AML and KYC teams operate.

The goal is not autonomy for its own sake. The goal is systems that help analysts work faster, reduce error, and make defensible decisions. If that means workflows instead of agents, that is not a compromise. It is a good design decision.