Where AI agents actually stop

Where does the legal line actually sit?

The unauthorized practice of law is what happens when someone without a license applies legal rules to your facts and tells you what to do. That definition matters more than it sounds. The line runs right through the middle of my idea.

An agent can fill out a form for you. Preparing a document at your direction is typing, and typing is legal in every state. California even licenses a job for it, called a Legal Document Assistant, which requires county registration and a twenty-five thousand dollar bond but no law degree.

An agent cannot tell you which form to file. Choosing the form means applying the law to your facts and recommending a course of action, and that requires a license in every state.

One inch separates those two things. On one side you're a very good typist. On the other side you're practicing law in forty-some states at once. Every company in this business lives on that inch. That inch shapes everything they are allowed to say to a customer.

Fine, I thought. I can build inside an inch.

Why can't an agent submit the paperwork?

The Social Security Administration will not accept an application from anyone but the person named on it. The agency's own policy manual, section RM 10205.025, states that applicants over eighteen who are physically and mentally capable must file on their own behalf. Social Security knows how to build a delegation channel. The agency built one for benefit claims, on form SSA-1696. The agency looked at name changes and chose not to.

Federal REAL ID rules block the remote path separately. Section 37.25 of Title 6 bans remote processing whenever personally identifiable information materially changes. A name change is exactly that. Section 37.11 requires a mandatory facial image capture. Florida's guidance says it in plain English: name changes may only be completed in person.

The identity layer closes the last door. Login.gov's Rules of Use state that automated access is strictly prohibited, and the text names automated form submissions specifically. ID.me's terms ban what they literally call an internet agent, and then the service asks for a live selfie with liveness detection.

The robot cannot hold the phone.

Even the banks want paper. Capital One is the most remote-friendly major bank on name changes. Capital One still requires a hand-written signature on an IRS W-9. The instruction says not to sign digitally or use DocuSign. Bank of America requires a branch visit.

Then in March 2026, a court ruled in Amazon v. Perplexity that a user's permission does not cure a platform's revocation under the Computer Fraud and Abuse Act. That case involved an AI agent taking write actions with a user's credentials and explicit consent. The user saying yes was not enough.

So the agent can prepare everything and deliver nothing. The agent walks her to the door and stops.

One caution on all of that, and it's the same caution I give clients. Rules move. Every citation above was verified against a primary source in August 2026. If you are reading this later, check it again before you lean on it.

Why is this entire category worth about fifty dollars?

Here is the part that turned the whole thing around for me, and it's the reason I'm writing this up instead of deleting a document and moving on.

I kept treating those walls as my problem. A thing I had run into. Something to engineer around.

The walls are not my problem. The walls belong to the whole market, and they have since 1997.

Thirteen companies already sell name change paperwork in the United States. Name Change Kit has been selling it since 1997, which is twenty-nine years. MissNowMrs opened in 2007 and claims more than five hundred thousand customers across nineteen years. That works out to roughly one to two million dollars a year. The category leader, NewlyNamed, sits at an estimated seven and a half million in revenue. NewlyNamed sold to a lower-middle-market private equity firm in 2023.

In twenty-nine years, not one company in this category has raised venture capital. Not one.

Prices sit at thirty-nine, seventy-nine, and ninety-nine dollars across at least eight vendors. The premium tier buys you pre-stamped envelopes, and at one company, a tote bag.

Forbes Advisor reviewed the category and wrote that none of these services actually files paperwork with courts on the customer's behalf. GOV+, the best-funded company in adjacent government paperwork, says it on its own homepage: you send it, GOV+ never submits on your behalf.

GOV+ is not being modest there. GOV+ is a company that spent money learning something.

Thirteen companies did not fail to think of automation. Thirteen companies hit the same wall I hit, and priced themselves accordingly. The valuable half of this job was closed off before any of them opened their doors. So everybody built the cheap half. The cheap half is worth about ninety-five dollars.

The wall did not just block me. The wall set the price.

What do the numbers actually say?

The economics finish the argument, and the economics are not close.

HitchSwitch publishes an average order value of ninety-four dollars and ninety-nine cents on its own affiliate page. That number is what a real person really pays for exactly this service, which makes it far better evidence than any survey.

Finding one customer costs more than that. LocaliQ's 2026 legal advertising benchmarks put family law at a hundred and three dollars and fifty-four cents per lead. The best figure I could find anywhere, from a well-run family law account, was thirty-four dollars per lead. At a generous twenty percent conversion from lead to customer, acquiring one buyer costs about a hundred and seventy dollars.

Ninety-five dollars in. A hundred and seventy dollars out. Once per customer, with no repeat purchase to make up the difference.

I want to name a number I got wrong, too, because the mistake is instructive. I had quoted a survey finding that thirty-five percent of women would pay ten thousand dollars to never move again. That survey came from a moving company's marketing blog, it published no question wording and no margin of error, and it asked about a thing nobody sells. The same survey reported that twenty-nine percent would rather get divorced than move.

Asking someone what they would pay measures how much a thing hurts. Only a completed transaction measures price.

What happened to the people who tried this?

Avni Patel Thompson spent nine years building a family assistant called Milo. Y Combinator backed it in 2020. OpenAI invested. When she ran out of money she cold-emailed Sam Altman and it worked, which is a better hand than most founders will ever hold.

Patel Thompson shut Milo down on January 12, 2026, and wrote publicly about why. Her words: the technology is simply too early to be reliable in any useful way. She described eighteen months of treading water.

Eighteen days later, on January 30, 2026, Panasonic shut down Yohana, its own family concierge service. Yohana's Japanese entity had accumulated a deficit of roughly three point seven billion yen.

The company that made this shape of work pay did it by not selling to the person doing the paperwork. Empathy has raised more than a hundred and sixty million dollars for bereavement administration, and roughly ninety-nine percent of Empathy's revenue comes through employers and life insurers. The grieving family pays nothing. That was not Empathy's backup plan. That was the finding.

What should you ask before you buy an agent?

Everybody selling you an agent right now is showing you a demo, and a demo is a browser, a form, and a happy ending. Real life has a counter. Real life has a person behind the counter, and a rule that wants to see your actual face, and a bank that wants a pen.

Ask who is on the other side of the door. Every agent eventually arrives somewhere it does not control, and that somewhere has a written policy. Go read the policy before the check clears, not after.

Ask what the demo skipped. A demo shows the path where everything works. Go find the wet signature, the phone tree, the in-person requirement, the person who wants to see identification.

Ask what the market already pays. If a job is genuinely valuable and genuinely unsolved, look at what people hand over today for the nearest real version of it. That number is a fact, and what people say they would pay is a feeling.

I lost the idea in four days. Building it first would have cost me a year.

What I kept is better: a map of exactly where an agent walks up to real life and stops, with the rule numbers attached. My clients get that map now, and it is worth more to them than the app would have been to anybody.

Cracked open, not broken. Same as always.

So what is the door in front of the thing you are building right now, and has anybody actually gone and pushed on it?