In a modern cockpit, the autopilot flies most of the route. And yet, no airline would ever call its pilots passengers. The crew works in two defined roles: one pilot flies, the other monitors. Monitoring isn’t the lesser job. It is trained, checked, and taken just as seriously, because aviation learned long ago that automation doesn’t remove risk. It moves it.
Tax is heading right into this configuration. And I believe our profession is better prepared for it than its reputation suggests.
Tax has always been in the residual risk business
Tax has a reputation for being cautious with new technology. That reputation is getting harder to defend. Across the profession, AI is moving from an occasional experiment to an everyday tool, faster than many of us expected. And this spring, Anthropic’s CFO mentioned on a podcast that the company’s number one internal user of Claude is its Head of Tax. Not an engineer, not a researcher. At one of the world’s leading AI companies, the heaviest user sits in tax.
It suggests that people who deal with residual risk every day are quick to see where these tools fit. That shouldn’t surprise us. Look at what our discipline is built on: documentation duties, the four-eyes principle, tax compliance management systems, materiality judgments, positions that must still hold up in an audit years later. Tax functions have never promised zero errors. What they do is manage residual risk systematically, and they can explain how. That is exactly the capability workflow agents demand.
So I don’t think tax should wait to see how agents perform elsewhere in finance. Tax should be setting the standard for how they are governed.
For CFOs, this matters beyond tax. Managing residual risk, insisting on a traceable trail, focusing review where risk is highest: that’s a blueprint for governing agents across close, reporting, and planning. Tax just happens to have practiced it the longest.
Why do we expect perfection from agents?
Let me be clear: the four-eyes principle works. It is one of the reasons tax functions deliver the quality they do. But it has never ruled out errors, and nobody expects it to. Time pressure, high volumes, complex fact patterns: every experienced tax professional knows where human work reaches its limits. We act on and accept that residual risk because we understand it.
With agents, many of us expect something different: zero errors. I think the reason is control. We simply feel more confident following a colleague’s line of reasoning than an agent’s, at least for now. When a person gets something wrong, we can usually retrace and discuss their thinking. With an agent, we don’t yet have that same confidence in how and where it might go wrong, so we compensate by demanding perfection.
That’s understandable, but it’s the wrong conclusion. The answer to missing intuition isn’t an impossible standard – it’s traceability. Trust in professional work has never come from someone telling you it’s right. It comes from being able to check it yourself. That’s how we review a colleague’s work, and it’s exactly what we should demand from an agent: which data it used, which rule it applied, how confident it was. If we can’t yet sense how an agent reasons, we need to be able to see it.
Then the right question becomes answerable. Not whether the agent is perfect, but whether the combination of agent and reviewer controls residual risk better than today’s setup. It often will, even for entirely rule-based steps. A deterministic step can be prepared by an agent and confirmed by a person just as well as the other way round.
Same goal, better visibility
In tax, the goal is and remains full correctness. Liability is real, and nobody signs a return on the assumption that it is probably right. What I expect to change is not the goal, but how we get there.
Today, reaching that goal relies largely on implicit safeguards: care, experience, and a thorough review. They work well, but they leave little trace of where the residual risk actually sat. In the future, I believe tax will manage that risk far more explicitly: knowing where uncertainty arises, making it visible, and controlling it deliberately.
This is where intelligence can genuinely support the goal of full correctness. A simple example: mapping P&L and balance sheet items to tax return positions. It is rule-driven, high in volume, and easy to get subtly wrong. In a mapping agent we are currently developing at Lucanet, every proposal comes with its basis and a confidence level. That gives the reviewer something a finished number never could: visibility into where the system was certain, where it wasn’t, and why. And the outliers often reveal where the underlying data, or our own definitions, are less precise than we thought.
That visibility is not a reason to look less closely. A confidence score is a claim, and it needs to be challenged like any other. But it turns review from checking a result into understanding how the result came about. For me, that is the bar any agent in tax must clear: calculations based on deterministic logic, and every step traceable. It’s also the thinking behind the Intelligence Core, the architecture all of Lucanet’s agents are built on. An agent that only delivers an answer can’t be reviewed. It can only be believed. And belief is not control.
You can delegate work, not accountability
For tax leaders, this changes the design question. Tax organizations have always been structured around the work: who prepares what, who reviews what, who signs? Hybrid teams need to be structured around accountability instead. Where does judgment sit? Who owns the standard an agent is measured against? Who notices when its performance drifts, or when a regulatory change makes yesterday’s correct answer wrong?
These aren’t IT questions. They are leadership questions. Only the tax function can answer them, because only the tax function knows what a defensible position looks like.
Learning to review as preparation shifts
Aviation calls it automation complacency: the more reliable the system, the harder it is to stay alert. That’s why pilots still fly by hand.
Tax faces a similar challenge. Our judgment has always been built through preparation: you learn when a number looks wrong by producing enough of them yourself. As agents take on more of that preparation over time, the path to building this instinct will shift. And reviewing won’t get easier. To judge a figure you didn’t produce, you have to understand every step that led to it.
I believe this will change how tax professionals learn, both at university and on the job, more fundamentally than we might expect. Learning will no longer happen only before or alongside the technology, but with it. Working through an agent’s reasoning, questioning its outliers, understanding why it was wrong: that becomes part of how judgment is built. Reviewing will need to be taught as a discipline in its own right, not assumed as the next career step. For those earlier in their careers, the work isn’t disappearing. It’s becoming more demanding, and more interesting.
Three considerations for every tax leader
- Hold agents to the same standard as your team: aim for perfection, measure against reality, and know your own baseline before you compare.
- Manage residual risk explicitly, not implicitly: know where uncertainty arises, and expect agents to make their reasoning and confidence visible.
- Design how your next generation learns judgment, while the work that used to teach it is shifting.
Workflow agents won’t take tax out of the cockpit. They will change who flies, who monitors, and what it takes to do both well. Tax, of all functions, knows how to manage that kind of risk. It’s time we lead with it.
