70% of AI Customer Service Agents Got Rolled Back. Tax Prep Is Much Harder Than Customer Service.

2026-07-273 min read

Last reviewed: July 24, 2026. Customer service AI failure data from a study of enterprise AI deployments. Tax preparation liability structure from IRS Circular 230 and EA representation authority.

The thing everyone got wrong about AI automation was the order. The assumption was: start with the easy stuff. Customer service chatbots. Data entry. Simple document processing. Work your way up to the hard stuff later.

The easy stuff isn't working.

A survey of thousands of companies that deployed AI customer service agents found that over 70 percent of them had to be rolled back or shut down. The reasons were predictable if you've ever talked to a chatbot: errors, miscommunications, answers that were confident and wrong.

Some companies got ahead of themselves. They saw the AI demos, calculated the cost savings, and laid off their human customer service teams. Then the AI went live. Then the complaints started. Then they had to hire everyone back. Frantically.

I keep thinking about this when I see headlines about AI replacing tax preparers.

Customer service is the easy target

Think about what a customer service interaction actually requires. Someone is unhappy about a product. They want a refund, a replacement, or an explanation. The company has clear policies. The conversation follows predictable scripts. The stakes are low. Worst case, the customer gets mad and posts on social media.

Tax preparation is not like this.

A tax return involves multiple forms that interact in non-obvious ways. The rules change every year. The client's situation is always specific. They forget things. They misremember things. They tell you one number and then find a different number a week later. The documents are incomplete. The deadlines are real. The penalties for getting it wrong are statutory.

And unlike customer service, when a tax preparer makes a mistake, the government notices. The IRS does not care that an AI made the error. The IRS cares that the return is wrong. The preparer whose name is on the return is responsible. Period.

What the EA credential actually protects

An enrolled agent has unlimited representation rights before the IRS. That means you can represent a client in an audit, an appeal, or a collection matter. You can sign powers of attorney. You can negotiate with revenue officers. These are not data-processing tasks. They are judgment tasks with legal consequences.

AI cannot do any of this. Not because the models aren't clever enough. Because the law does not recognize an AI as a representative. Circular 230 governs practice before the IRS. It applies to people. Real people with credentials and liability.

This is the structural moat that separates tax preparation from most other white-collar work. The credential is not just a qualification. It is a legal gate. AI cannot walk through it.

Why the 70 percent number matters

If AI cannot reliably handle "where is my refund" with a script and a policy document, it cannot handle a Schedule C with a client who thinks their personal vacation was a business expense.

If AI gets rolled back on password resets, it gets rolled back on qualified business income deductions.

If companies are rehiring the customer service reps they fired, tax firms are not going to fire preparers and replace them with models that confidently hallucinate tax code sections.

The technology will improve. It always does. But the gap between "impressive demo" and "production deployment where someone's legal liability is on the line" is not a gap that better training data closes. It's a gap that requires the model to be right in ways that matter, every time, with consequences for failure that the model cannot absorb.

Until AI can sign a return under penalty of perjury, the EA credential is not going anywhere.

Dojo Challenge · strike before you scroll

Retrieve it before the explanation retrieves it for you.

For court proceedings resulting from an examination, the IRS generally has the burden of proof for any factual issue, assuming the taxpayer has met certain requirements. Which of the following is not one of those requirements?

Enter a full practice round →Check readiness

Keep building this concept · 3 related guides

AI Isn't Killing Tax Preparation. It's Creating More Demand for Human Preparers.Everyone's worried AI will automate tax prep. The data says the opposite: AI displaces junior white-collar workers into gig work, and gig workers need human tax preparers. Two curves are bending toward the profession at once.US Citizens in Vietnam Need an EA. No Treaty Makes Every Return Harder.Vietnam is the only major US trading partner without an income tax treaty. No tie-breakers, no reduced withholding, no totalization agreement. Americans in Vietnam need credentialed tax help more than expats in most treaty countries.IRS Ramps Up Enforcement: Why Enrolled Agent Demand Is SurgingThe IRS processed 139 million returns in 2026, issued 90 million refunds, and is hiring thousands of new agents. Here's why EA demand is skyrocketing.

Find your weak topics →Open the 50-unit path →