Maria Scott, CEO of TAINA Technology, provider of a world-class tax technology platform.
Most conversations I have about AI in financial services still start with tasks. Can it read a document faster? Can it classify data more accurately?
These are useful questions, but in my experience they miss what happens when an entire process can respond to what it encounters along the way, rather than simply completing one predefined task.
I believe the distinction between automating a task and orchestrating a workflow is still too often overlooked, even by firms well into their AI evaluations. This distinction shapes where agentic AI can actually add value and where it simply adds speed.
Traditional Software Executes. It Doesn’t Decide.
Traditional software, even highly automated software, is fundamentally rule-bound. It takes predefined inputs and produces predefined outputs along paths designed in advance: extract the fields from a document, run a validation check on a form, send a notification when a trigger fires.
When conditions match what the system was built for, this works extremely well, and there is no reason to replace deterministic automation simply because a newer technology exists.
The challenge comes when something happens partway through a process that changes what should happen next.
Traditional systems can branch, but only along routes someone anticipated and coded beforehand. They cannot look at an unusual combination of facts, reason about the next step and change course. That has always required a person, which builds a human bottleneck into every process that involves judgment.
What An Agentic Workflow Does Differently
An agentic workflow is designed around an objective, not an isolated task. It breaks that objective into steps, works across the relevant systems and data, evaluates the outcome of each step and uses it to decide what happens next.
That does not mean giving AI unlimited authority. In regulated environments, some actions can be orchestrated by an agent, some decisions must remain governed by validated data and deterministic regulatory rules, and others require human judgment or approval.
The real sophistication is not in how much an agent can do, but in how clearly those boundaries are defined and held as the process runs, including the agent recognizing when it has reached a decision it should not make alone.
Tax operations, the world I work in every day, show this clearly.
Consider a payment to an intermediary, such as a bank receiving income on behalf of its own clients. It may be supported by a W-8IMY, the U.S. tax form intermediaries provide, along with a withholding statement and separate documentation for multiple underlying payees.
The operational challenge lies in determining whether the documentation and account information together support the next step: allocating the payment correctly across payees or withholding-rate pools, the groups of payees taxed at the same rate.
An agentic workflow can assemble the relevant documentation and account data, assess which parts of the structure are complete using approved rules and available evidence, and act within defined regulatory boundaries. A fully supported allocation continues through the normal process.
Where the operating rules permit it, missing documentation for one payee can trigger a targeted request without holding up payees who are already properly documented. Where information conflicts, an allocation cannot be supported or the facts fall outside an approved rule set, the case is routed to a specialist.
Crucially, in this case, the agent is not given authority to make the tax determination itself. Its role is to understand where the case sits in the wider process, use the result of each step to decide the next action and keep things moving until the objective is met or human judgment is required.
The same principle applies when a change in a client’s circumstances is detected, such as a new address or standing instruction. The change may have no impact on existing tax documentation, may trigger a defined cure or resolicitation process to obtain updated documentation, or may create an ambiguity that needs human review.
Instead of treating each event as an isolated task, the workflow uses the outcome to decide which governed path to follow and records the result.
Where Leaders Should Start
Much of the AI conversation I encounter still centers on speed and cost. While those benefits matter, workflow orchestration raises a different question: Can a process respond intelligently to what it encounters without a person manually rerouting every exception?
For leaders considering agentic AI, I would start there before focusing on the technology. Take one workflow where the next step currently depends on somebody making a judgment call, and then break that judgment apart. Which parts can be assessed consistently from the evidence available? Which are governed by rules that should not bend? Which require human judgment or sign-off?
Rather than giving an agent as much authority as possible, the goal is to define, deliberately and precisely, what it is allowed to decide, what must remain governed by deterministic rules and where its authority ends.
That, I believe, is where the move from task automation to agentic workflows really begins. It is also where the people doing this work every day start to trust it. When people can see exactly where the agent stops and their own expertise takes over, they are free to spend their time on the judgment only they can bring.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?