The skill that matters isn't using AI - it's knowing when it's wrong

The skill that matters most for tax professionals working with AI is evaluating its output, not operating the tool. That means reviewing an AI-prepared return with the same professional skepticism you'd apply to a junior preparer's work: checking it against the source documents, knowing the common failure modes, deciding when to override, and giving feedback that improves the system for the whole firm.

Here's something we've learned at Filed from working inside tax practices every day:

There's a large gap between knowing how to use an AI tax automation platform and knowing how to review what it produces. Most of the training available to tax professionals right now addresses the first. Almost none of it addresses the second.

Using an AI tool means knowing which buttons to press. Evaluating AI output means knowing that a Schedule D looks right on its face but the cost basis was pulled from the wrong lot. It means recognizing that the system handled the W-2 correctly but missed the fact that the client changed jobs mid-year and the state allocation is off. It means applying the same professional skepticism you'd apply to a junior preparer's work, except the "preparer" processes returns in minutes and never flags its own uncertainty.

That's the actual skill. And almost nobody is teaching it.

What AI skills gap are tax firms missing?

Most firms train people to use AI tools, but few train them to review AI output, and reviewing is the skill that decides whether AI saves time or creates rework. The tooling side is covered in AI tax software: what actually works.

Apollo's chief economist Torsten Slok recently argued that AI will grow the accounting industry, not shrink it. The Jevons paradox: cheaper services, more demand. Fortune pushed back, citing Dallas Fed data showing a 13% employment decline for 22-to-25-year-olds in AI-exposed occupations since 2022.

Both are probably right. The industry grows. The skills that get you hired change.

The practical question for anyone running a firm: what does "qualified" look like in 18 months?

A tax professional who can review an AI-prepared return with real rigor, catching failure modes, identifying edge cases, knowing when to override and when to escalate, is worth significantly more than one who can't. A tax professional who can also give structured feedback that improves the AI system over time, who can help define the guardrails for how their specific firm uses these tools, is a different category entirely. That second person doesn't just use AI. They make the AI better for everyone at the firm. The gap between firms that have people like that and firms that don't will compound every single tax season.

How should you review an AI-prepared tax return?

There are three distinct things a good AI reviewer does that most "how to use AI" training ignores entirely.

Catching output failure modes. AI systems have predictable patterns of failure. They pull from the wrong source document, handle mid-year life changes poorly, miscategorize income on unusual K-1 structures. A skilled reviewer knows the failure modes before they start, not after.

Knowing when to override. The hardest call isn't "is this number right." It's "is this structure right." An AI can produce a technically accurate return that still misses the strategy. The reviewer has to catch that.

Improving the system over time. This is the one most people don't think about. Good feedback given to an AI system improves it for every other user at the firm. That feedback loop is what separates practices that get compounding value from AI from ones that plateau.

How can tax professionals learn to evaluate AI output?

Through structured training that treats AI review as its own skill, which is why Filed's Rachel DiFuccia built a course for it.

Our colleague Rachel DiFuccia came to Filed as a tax professional. She spent two years on the product side, working alongside engineers and ML teams. The thing that kept coming up: the hardest part of building AI for tax isn't the technology. It's the translation layer. Someone has to understand the tax work deeply enough to know what "right" looks like, and understand the AI system well enough to know why it got something wrong and how to fix it.

That role doesn't have a formal name yet. But it's the role the best firms are already building around.

Rachel built the AI Tax Specialist and Engineer course with Miles Masterclass to make that skillset teachable. Seven modules:

1. Evaluating AI output with professional skepticism. Does this return reflect what the source documents actually say? Where are the common failure modes? What does AI miss that a human wouldn't?
2. Operating inside AI-enabled workflows. When to intervene. When to let the system run. Where the risks live in each handoff.
3. Making the system smarter over time. Structured feedback, firm-specific rules and guardrails, working effectively with the teams building these tools.
4. Keeping it compliant and defensible. Documentation standards, auditability, risk considerations for when a client or regulator asks how the return was prepared.

The full course is seven modules, ends with an applied project through a real AI-assisted tax workflow from intake to delivery, and takes 12-15 hours. It's CPE eligible and ends with a LinkedIn credential.

Who should take this: If you run a firm, this is the fastest way to move your team from "aware of AI" to "capable with AI." Send your reviewers and the person you're hoping becomes your internal AI lead. If you're earlier in your career, this is the credential that will matter in 18 months, not two years. "I can use AI tools" is already table stakes. "I can evaluate AI output and identify failure modes" is not.

Frequently asked questions

What are the most common failure modes in AI-prepared tax returns?

Pulling figures from the wrong source document or lot, missing mid-year life changes such as a job change that shifts the state allocation, and miscategorizing income on unusual K-1 structures. A reviewer who knows these patterns checks them first.

Who is responsible for an AI-prepared return?

The tax professional who signs it. The IRS holds practitioners to the due diligence standards in Circular 230 no matter what tool produced the first draft, which is why review skills matter more as AI does more of the preparation.

How does Filed help reviewers check AI output?

Filed Reviewer compares the source documents against the draft return line by line and produces a cited review list, so every flagged item traces back to the document it came from. Filed scored 94% line-by-line accuracy on TaxCalcBench. It works with CCH Axcess, UltraTax, Lacerte, Drake, and ProConnect across 1040, 1065 (with K-1), 1120, 1120-S, 1041, and 990 returns. Firms using Filed report 64% shorter review cycles, and a named reviewer at the firm still makes the call.

Is the AI Tax Specialist and Engineer course CPE eligible?

Yes. The course takes 12-15 hours, is CPE eligible and ends with a LinkedIn credential.

Sign up for the AI Tax Specialist and Engineer course