In the Weeds: Can You Trust AI With the Tax Letter From the IRS?

The envelope is thicker than a bill and thinner than a package. Someone in the house sets it on the kitchen counter, under a coupon circular, and the letter sits there for two days because nobody wants to be the one to open it. When it finally gets opened, the first paragraph contains the words "Notice CP2000" and the phrase "proposed changes to your tax return" and a number with a dollar sign in front of it that is large enough to ruin a weekend.

I spent about a week testing what AI is actually good for in moments like that, and what it is actively dangerous for. This piece is the thing I wish someone had handed me before I started: an honest map of the territory, written for the person standing at the kitchen counter with the letter in their hand.

It applies to more than the IRS. It applies to state tax board letters, Medicare Explanation of Benefits forms, school district residency questionnaires, insurance denial letters, HOA notices with cryptic code numbers, unemployment redetermination letters, and the single worst genre of mail in America, the one that starts with "This is not a bill" and then presents a number that is, in every meaningful sense, a bill.

What AI is genuinely good at here

There are three things AI does well with bureaucratic mail, and it does them so well that the first time you see it work, you feel a small, unreasonable burst of affection for the machine.

The first is decoding the language. Official letters are not written to be understood. They are written to be legally defensible, which is a different goal, and the prose reflects that. Sentences run thirty words long. Defined terms get capitalized. Reference codes float in the margin with no explanation. An AI model, given the full text of the letter, will translate it into plain English at roughly the reading level of a newspaper, and it will do so without sighing or making you feel stupid for asking. You can ask follow-up questions. You can ask it to explain a specific paragraph again, slower. You can ask what "assessment" means in this context versus in the context of your neighbor's property tax argument. It will answer. This is not a small thing.

The second is summarizing what is actually being asked of you. Most government letters bury the ask. Somewhere around paragraph four there is usually a sentence that says, in essence, "if you disagree, you have sixty days to respond in writing." Everything before that is context and everything after that is boilerplate, but the sentence in the middle is the entire point of the letter, and it is almost never bolded. An AI model is very good at finding that sentence and pulling it to the top for you. "Here is what they want. Here is the deadline. Here is what happens if you do nothing." Three lines. You can make a decision from three lines. You cannot make a decision from six pages of letterhead.

The third is drafting a polite, structured response once you know what you want to say. This is the least glamorous use and possibly the most valuable one. Writing back to the government is intimidating. Most people freeze at the salutation. AI will happily draft a calm, clear letter that states your position, references the notice number, asks for the specific thing you are asking for, and sounds like it was written by an adult who is not panicking. For this specific task you can reach for the ✉️The Difficult Email Rewrite, which is designed for exactly this moment, the one where you know what you want to say but every draft comes out either too apologetic or too angry.

A soul worth knowing for this kind of work is 📄The Paper Mountain Paralegal. She is calm about paperwork in the way that a person who has filed ten thousand motions is calm. She will not give you legal advice and she will say so, often, but she will sit with you and a letter and help you understand it, and she will not make you feel like the letter is a personal failing. Paired with 📋The Elder Paperwork Decoder, the two of them can get a kitchen-counter stack of envelopes down to a manageable list of "do this, ignore that, call this number, file this one under 'resolved.'"

What AI is not good at, and why "not good at" matters here

This is the part where I have to be careful, because the difference between "mostly fine" and "not fine" with a tax letter is the difference between a good weekend and a bad year.

AI models hallucinate. That is a polite, industry-standard way of saying they make things up with complete confidence. Most of the time, with most topics, the thing they make up is close enough to the truth that nobody notices and nobody is harmed. A slightly wrong lasagna recipe is still dinner. A slightly wrong explanation of the French Revolution will still get your kid through the quiz. The stakes are low.

Tax and legal content is different. The stakes are not low. If an AI tells you that you have ninety days to respond when you actually have thirty, and you believe it, you can lose your right to contest a notice entirely. If it tells you that a particular deduction is allowed when it is not, you can owe penalties on top of the original amount. If it confidently cites a tax code section that does not say what it claims, and you paste that into a letter to the IRS, you have just handed the agency a written admission that you do not know what you are talking about, which is not the position you want to be in. AI will do all of these things, cheerfully, and it will sound authoritative while it does them.

So the rule I use, and the rule I would recommend to anyone, is this: AI is allowed to help you understand the letter. AI is not allowed to decide what you do about it. Understanding is a low-stakes translation task and AI is genuinely good at it. Deciding is a high-stakes strategy task and AI will get you in trouble because it does not know which of its own sentences are true.

This means a few specific things in practice.

Do not ask AI to tell you whether you actually owe the money. It does not know. It cannot see your return, your records, your W-2s, your 1099s, your prior year's notices, or the agency's internal file. It can only tell you what the letter says the agency thinks, which is not the same as what is true.

Do not ask AI for tax strategy. "Should I take the standard deduction or itemize," "should I file an amended return," "should I request an installment agreement or an offer in compromise," are questions for a human who will sign their name next to the answer. A CPA or an enrolled agent or a volunteer at a VITA clinic or a tax attorney, depending on the size of the number and the complexity of the situation. The AI will have opinions on these questions. The opinions are not advice. Treat them the way you would treat advice from a stranger at a bus stop who claims to be a lawyer.

Do not paste AI-generated citations into a response letter without verifying each one in the actual Internal Revenue Code or the agency's own publications. AI will invent section numbers. It will invent case names. It will invent IRS publication titles. These inventions look exactly like real citations because the model has been trained on millions of real citations. The difference is that the real ones are real.

Do not, under any circumstance, let AI negotiate for you. If you are on the phone with an agent, AI is not on the phone with you. If you are in an audit, AI is not in the room. The letter is the easy part. Anything after the letter is a human task.

The hallucination test I actually use

Here is a thing I started doing after watching an AI confidently misquote a notice I had in front of me. Before I trust any AI summary of an official letter, I run a small test. I ask the AI, "what does paragraph three of this letter say?" and I compare its answer to paragraph three of the letter I am holding. If the two match, the model is paying attention to the actual document and not to its training data's memory of similar documents. If they do not match, I stop the session, start a new one, and paste the letter in again, because once a model starts drifting it does not stop on its own.

This takes about fifteen seconds and it is the single most useful habit I have picked up from a year of doing this kind of work.

The other habit is running the same letter through a second model, from a different company, and comparing the summaries. If both models say the letter is asking for the same thing, I am reasonably confident that is what the letter is asking for. If they disagree, I read the letter myself, slowly, with a highlighter, because something is off and I do not yet know what.

The four-step workflow

Here is the specific thing to do, next time a letter lands on the counter. It is four steps. It takes about thirty minutes for a medium-bad letter.

Step one, stabilize. Open the envelope. Read the letter once, out loud, all the way through, even the parts you do not understand. Do not take notes yet. Do not look anything up. The point of this read is not comprehension, it is to lower the temperature in your body so that you can think. Letters like this trigger a small panic response, and a small panic response is exactly the state in which you will make the worst possible decision. Read it out loud, put it down, drink a glass of water, come back in ten minutes.

Step two, decode. Open a fresh AI session. Type or paste the entire letter, word for word, including the notice number and the date. Ask three specific questions, in this order. "What is this letter, in one sentence." "What is the agency actually asking me to do." "What is the deadline." Do not ask for advice. Do not ask for strategy. Just ask for translation. Then run the hallucination test I mentioned above, the one where you ask the AI to quote paragraph three back to you, and check that it matches.

If the letter is personal paperwork for an older parent, this is where 📋The Elder Paperwork Decoder earns its place. It is designed for exactly this use case, the stack of mail that arrives at a parent's house that a grown child is trying to help with from two states away.

Step three, decide. This is the step where you stop using AI and start using a human. Look at what the letter is asking. If it is asking for a small amount of money that you agree you owe, and you have the money, pay it and move on. If it is asking for a small amount that you do not agree with, the response is usually a short written letter explaining why, and an AI can help you draft that letter. If the number is large enough that losing it would hurt, or if the letter uses words like "audit," "lien," "levy," "summons," "appeal," or "criminal," stop. Call a real human. Your options, ranked by cost, are: a VITA clinic if you qualify, an enrolled agent, a CPA, or a tax attorney. The IRS also has a Taxpayer Advocate Service that is free and exists for exactly this reason. None of those options are AI and that is the point.

Step four, respond. If you are writing back, and it is a clear, simple response, draft it with AI help and then read it three times before sending. Use ✉️The Difficult Email Rewrite as the starting prompt. Ask the AI to keep the letter short, keep it polite, keep it factual, and include the notice number at the top. Then read the final version carefully and ask yourself: is every single factual claim in this letter something I know to be true? If the answer is no, delete that sentence. Do not send anything in writing to a government agency that you cannot personally defend.

Keep a copy of whatever you send. Keep it forever. "Forever" is not a figure of speech in tax matters.

The part where I admit what I do not know

I am not a tax professional. Neither is the AI. The AI is, in fact, extremely clear about not being a tax professional if you ask it, and then it will proceed to give you tax advice for forty-five minutes if you let it. Knowing the limit is on you.

What I do know, from a week of staring at model outputs with real letters in hand, is that the translation work is real and valuable, the strategy work is a trap, and the difference between the two is the most important thing to get right. Use a-gnt tools to understand the letter. Use humans, and their signatures, to decide what to do with it.

The letter is still on the kitchen counter. It is no longer frightening. It is a task. Tasks have four steps. Open the envelope.