What an AI Letter Writer Can't Know About You

Every result on the first page for AI letter writer is a text box. I counted ten of them in July 2026, including an app listing on Google Play, and the closest thing to an article was a roundup of other text boxes.

Nothing wrong with that. It does mean the question underneath the search goes unanswered.

The short version, for anyone who wants it now. An AI letter writer is genuinely good at the part of a letter that's convention, which is the greeting, the shape, the closing, the tone, and it's structurally incapable of supplying the part that decides whether the letter works, which is the specific facts about your situation. For a handful of letter types that gap is trivial. For a few others there's a free template published by a government agency that's better shaped than anything a generator will hand you, and I'll name them below because none of the ranking pages do.

The Format Half Is Genuinely Solved

Let me be fair to the tools before I'm rude about them.

A business letter has a documented structure and it hasn't changed in decades. Purdue's writing lab lists eight parts, which are the sender's address, the date, the inside address, the salutation, the body, the closing, enclosures, and typist initials. It specifies "Times New Roman, size 12", body text "single spaced except for a double space between paragraphs", and block format where the "entire letter is left justified and single spaced". It also says to "use a colon (:) after the salutation (never a comma)", which is the single most commonly botched thing in letters I've been sent.

All of that is exactly the sort of convention a model reproduces perfectly. It's high-frequency and unambiguous, and it's been written down in a thousand places already. If your only worry is that you don't know how a letter is supposed to look, a generator solves your problem in four seconds and you can stop reading here.

What It Can't Reach

Now the other half. I put this table together because I couldn't find anything like it, and I want to be upfront that the boilerplate column is my own judgment rather than a measurement of anything.

The pattern in that middle column is that it's all particulars. Dates, names, reference numbers, the sentence someone said to you on the phone in March. A model has none of it and will cheerfully produce a letter with square brackets where the substance goes, which is the correct behaviour and also the reason the output feels thin.

Where it goes wrong is when the brackets get filled in badly, or worse, when the model guesses and the guess reads plausibly.

The Letters Where A Government Agency Wrote It Better

Here's the thing that surprised me while researching this, and it's absent from all ten results.

The US Consumer Financial Protection Bureau publishes five sample letters for dealing with debt collectors. Loaded 28 July 2026, they cover saying you do not owe the debt, asking for more information about the debt, asking the collector to stop contacting you, asking them to only contact you through your lawyer, and specifying how they can contact you. Five distinct scenarios, each with its own letter, free, downloadable.

Their guidance attaches two things that matter more than the wording. Timing, because "if you use any of these letters, it's important to do so as soon as possible after the debt collector first contacts you", and the page notes you only have 30 days after initial contact for certain requests. And documentation, because you should "keep copies of any letters you send and any other communications". The page also says plainly, "Note: These sample letters are not legal advice."

Now compare that to what a generator does with the prompt "write a letter to a debt collector". It will produce something polite and well formatted that expresses a general objection. It will not know which of the five moves you're making, because those are legally distinct and choosing between them is the actual work. It won't mention the 30-day window unless you already knew to ask about it.

I'm not claiming the generators are dangerous. I'm saying that on this specific category, a free official template is better shaped for the job, and the fact that no page ranking for AI letter writer mentions it tells you those pages exist to acquire users rather than to answer the question.

Resignation Letters And The Number Nobody Generates

A resignation letter is about 90% convention. There's one fact in it that matters and it isn't the tone.

Your last working day. That's a function of your notice period, which lives in your employment contract, and there's usually a statutory floor underneath it. In the UK, the government's guidance says "You must give at least a week's notice if you've been in your job for more than a month." Contracts commonly require more. That page also makes the practical point that you can resign verbally unless your contract says otherwise, and recommends putting it in writing if you might need to refer to it later.

Different country, different rules, and I only checked the one, so treat that number as an example of the shape of the problem rather than as advice for wherever you live.

The point stands regardless. A generator will happily write "my last day will be [DATE]" and the entire consequence of the letter, meaning whether you leave cleanly or in breach, sits inside that bracket. Getting the prose right is the easy 90%. The hard 10% is a document you have to go and read.

If The Thing Is Going In An Envelope

Small section, mostly because it's the half that's completely absent from every tool page.

A generated letter is a body of text. A posted letter also has to be addressed in a way that machines can read. The USPS addressing standards say "Uppercase letters are preferred on all lines of the address block" and to "format all lines of the address with a uniform left margin", with the line order running recipient name, then delivery address, then city, state, and ZIP.

That's a mildly interesting mismatch. Purdue's letter format wants a conventionally capitalised inside address inside the document, and the postal standard prefers uppercase on the envelope. Both are right for their own surface. A generator gives you the document and nothing about the envelope, which is fine as long as you know that's where its job ended.

What "Free" Means Across These Tools

I read the claims on the two pages that hold the top spots.

Mailmeteor's tool says "No sign-up is required!" and describes itself as "completely free to use, with no hidden fees or limitations. You can create as many letters as you need, whenever you need them." Template.net's generator is also free with no signup, lists something north of 60 letter templates across professional, business, healthcare, and academic categories, offers exports to DOCX, PDF, and TXT, and has an Upgrade button in the corner, which usually means there's a ceiling somewhere that the page doesn't publish.

I didn't push either tool until it broke, so I can't tell you where the real limits sit. What I can say is that this whole category is unusually generous compared to the paid writing subscriptions, which makes sense, because letters are short and a short generation is cheap to serve. The wider question of what free tiers cost you in other ways is covered in what free AI writers actually give away. The short version for letters specifically, though, is that the free tier is usually the whole product, and the paid tier is buying tone controls and a longer memory you will not use on a two-paragraph complaint.

Why This Results Page Has No Articles In It

Slight detour into my own trade, because the shape of this search result is a decent worked example.

My pipeline reported roughly 2,900 monthly searches for the phrase in July 2026, a difficulty score of 3, and a cost per click of $4.86. A difficulty of 3 reads as one of the easiest numbers you'll ever see, and a beginner looking at that would reasonably conclude this is free traffic sitting on the table.

It isn't, and the reason is a column that difficulty scores don't have. I count how many results in a top ten are soft, meaning forums, Reddit threads, Medium posts, question sites, anything that's there by default rather than by strength. For this query that count is zero. Ten commercial results, no gaps, every one of them a page built by a company that wants the same visitor you do. A soft slot is a door. There isn't one here.

The one opening is that last row. Every incumbent is a tool, so a page that explains rather than generates isn't competing for the same job, which is roughly the bet this article is making. I'd put my confidence at moderate, and the honest version is that I'll know in six months.

I also wanted install counts and a rating for the Android app holding third place, since an app store listing outranking most of the writing vendors is genuinely odd. The listing wouldn't give me its numbers on the attempts I made, so there's no figure here rather than a guessed one.

The Way I'd Use One

Not a workflow so much as a habit, and it takes about the same time as writing badly from scratch.

Gather the particulars first. Dates, reference numbers, the exact wording of whatever you're responding to, the outcome you want stated in one sentence. Then hand all of it over and ask for the letter. The output improves out of all proportion to the effort, because you've moved the model from inventing to arranging, which is the mode it's actually reliable in. Same principle as everything else in how AI fits into a writing process.

Then read it once for the thing generators do worst, which is confident specificity you didn't supply. If a date, a figure, or a clause reference appears that you didn't provide, delete it. Don't correct it. It came from a pattern rather than from your situation, and a corrected invention is still an invention wearing better clothes.

And for anything with a deadline or a legal consequence attached, go find the authoritative version first. It exists more often than people expect.

The Uncomfortable Bit

I don't have a receipt for this one. I've not run a controlled test of generated letters against hand-written ones, nobody I know has, and I'd be suspicious of anyone claiming they did.

What I have is the observation that the letters which go wrong go wrong on facts rather than on phrasing, and that the entire product category is optimised for phrasing. That's not a flaw in the tools so much as a mismatch between what's easy to build and what's actually hard about the task. The same mismatch shows up across most of the writing stack, which the full breakdown of what these tools cost and cover gets into properly.

Good letters are mostly research wearing a polite voice. The voice is the free part now.