With teams of AI agents already deployed across PwC’s tax function, the firm’s US AI tax leader Dom Megna and his team are moving beyond pilots and into production and execution by using agents to automate everything from K-1 data mapping to tax diagnostics.
In this conversation, he explains how PwC trains its AI systems, governs risk, builds trust with skeptical CFOs and keeps its legacy tax employees engaged and upskilled during its agentic AI-infused transformation.
Dom Megna
US AI Tax Leader, PwC
Notable previous employers:
- Deloitte
This interview has been edited for brevity and clarity.
You’ve said AI agents can act on their own, even in unfamiliar tax situations. Can you walk me through a real-world case where an AI agent had to interpret data in a way that previously required a team?
One example that comes to mind is how we handle mapping data as it comes into our platforms. This happens across all our tax businesses. We get data from clients, sometimes from multiple systems, sometimes from third-party providers. These are often in different formats, different structures and different reports. Some of it isn’t clean. It’s not mapped correctly, or it’s missing metadata.
Another good example is in the asset and wealth management space. We often ingest K-1s and have to map items based on them. Historically, we’d build guides and train staff to interpret those K-1s, then map them to the correct treatments. Like, for example, a deductible expense. It involved a lot of human review, and then we layered in [optimal character recognition] tools to help with the extraction.
What the agents are doing is: they’ve been trained on how to treat different types of expenses and disclosures. So when they see something in a K-1 footnote, they assess it. “Is this deductible?” “Where do I map it?” That’s a decision a team of humans might have previously huddled to make, maybe around a cubicle or with a quick phone call, especially when the footnote was something we hadn’t seen before.
But the agent now says, “Here’s how I interpreted that. Here’s my confidence level.” And the user reviewing the output gets that flagged. Maybe it’s a 70% match, maybe it’s lower, but the human being can check it. If they override it, we ask them to tell the agent why, so that next time it sees that pattern, it can learn from that correction.
That’s the human-in-the-loop piece. It’s not a complete handoff. But instead of three people gathering to solve the issue, one person reviews and the agent improves from there.
How much feedback does an agent like that require? Is it a consistent need or something that will wane over time?
There’s a lot of training up front. You can think of these agents like staff members. Each one is using a set of tools. That’s where PwC’s Agent OS comes in. It’s a kind of operating system that manages which tools each agent uses, like data mapping or document extraction.
“It’s not a complete handoff. But instead of three people gathering to solve the issue, one person reviews and the agent improves from there.”
Dom Megna
US AI Tax Leader, PwC
As the agents use these tools, they learn from them. For example, “When I see this type of footnote or disclosure, I need to treat it this way.” We’ve trained them using the historical data we already have. But as new data comes in, the learning has to continue. Feedback is how we improve that confidence level — ideally to 100%, though that’s rare.
Just like training a human team, you want to reinforce what’s correct, guide the edge cases and help the agent interpret nuance. We constantly retrain them with real-world scenarios so they improve and apply that learning back into future use.
Given how new all of these things are, how did you know this technology was ready to present to clients?
We saw a huge need in the market for something like Agent OS, even inside our own environment. We were already deploying a lot of agents across the firm. PwC has over 800 custom GPTs and over 250 AI agents, including within tax. But we needed a better way to manage them. Organize, orchestrate, monitor, and those are all things Agent OS is built to do.
As we started working toward broader transformation [in FY26], we realized we weren’t alone. We were hearing from clients who were building agents into their enterprise IT teams, marketing, finance and even customer service. All of those needed oversight. You don’t want dozens of agents operating without a way to track performance or coordinate their logic.
Agent OS gives us a framework to integrate with internal data, like using APIs or [model context protocol] language, and gives agents access to tools like document extraction or mapping. It’s not just one function. It’s a system that connects tools, logic and workflows in one place.
Is there a line here where humans — and only humans — are being used in a certain area, or is every part of the tax function up for grabs?
The sky’s the limit. But yes, there are areas like policy interpretation or high-touch consulting work where the use of AI is still pretty limited. Not because it can’t help eventually, but because the kind of reasoning and client interaction involved is still very human.
That said, we have more than thousands of tax engagements a year. A majority of those follow well-understood processes, and that’s where we’re seeing a massive opportunity. The question becomes: Where do we focus first? Where do we get the biggest lift? That’s the same decision CFOs are making. Things like, “Where do I invest now? Where will I see the impact?”
We’ve been using AI in some form for over 10 years. Machine learning, modeling and the fundamentals of these things are not new to us. But what these new agents allow is a shift from passive data analysis to proactive action. It’s not just asking a model a question; it’s giving the agent a task and expecting it to complete it. That’s the big leap, and it’s why we’re making this a major priority for 2026.
Let’s talk about risk. When a CFO hears “AI” and “tax filing,” they’re going to think about audits. How are you preparing for the regulatory review of agentic systems? And what would you say to a client who asks, “Will this hold up under IRS scrutiny?”
It’s a fair question. Our assurance practice has been working on this closely, and we just launched an article and a service offering to provide assurance for AI systemss.
At a firm level, we’ve set up a governance committee that oversees how we build and use these systems, including in tax. So, as we create new processes, we’re asking: how do we monitor them, how do we document the reasoning and how do we show our work?
“It’s not just asking a model a question; it’s giving the agent a task and expecting it to complete it. That’s the big leap, and it’s why we’re making this a major priority for 2026.”
Dom Megna
US AI Tax Leader, PwC
The models now give us that transparency. One we use calculates effective tax rates. We’ve trained it to explain itself. What were the inputs? What’s the output? What assumptions did you make? That entire logic chain is now part of the deliverable.
In the past, we might have written a memo. Now it’s automated and traceable. We also control who can make changes. Once something is approved, it’s locked. People can suggest improvements, but we don’t want ad hoc adjustments that create risk.
One of the biggest concerns I hear from CFOs is, “I just don’t trust it yet.” If you’re pitching an AI-powered tax process to a skeptical CFO, how do you sell trust here?
I start with governance. We have a responsible AI toolkit that’s built around the principles of explainability, traceability and audibility. If you can’t explain it, you can’t trust it. If you can’t trace it, you can’t validate it. And if you can’t audit it, you probably shouldn’t use it.
We’re not just pushing a button and accepting the result. Our agents still live within a human-reviewed, multi-eye process. Especially for complex outputs, like tax calculations, the oversight is strong. For simpler diagnostics, you might not need three layers of review, but you still have that human in the loop.
We also talk a lot about risk tolerance. Early on, your guardrails should be tight. That’s smart. But as you see the same result over and over, trust builds. Then you start to open up and think, “Where else can I apply this?” That’s when the innovation starts to take off.
This is a technological shift, but it’s also a people shift. How are you making sure your teams, especially the experienced Big Four tax professionals, stay engaged and don’t feel left behind by this new technology?
Here, we’re lucky to have a really strong workforce transformation team that’s embedded within tax. We’ve rolled out a few different initiatives to support people through this.
First, we’ve identified over 3,200 AI champions across the firm, and many of them are inside tax. These aren’t just senior leaders; they’re spread across every business unit. Their job is to help teams adopt tools, answer questions, and bring AI into live client engagements in a way that’s helpful and not overwhelming.
Second, we’re investing heavily in upskilling. Through our My AI program, we’ve clocked over 350,000 training hours in the U.S. firm. About 80% of our people have taken some form of AI training, whether it’s basic awareness or more advanced tool use.
We also encourage experimentation. There isn’t just one tool that solves everything. So we want teams trying new tools, giving feedback and applying them thoughtfully. And we protect time for training, because we know people are busy. It can’t be an afterthought.
