Should you let an AI agent write your offer?

AI tools can now draft a purchase agreement or summarize a disclosure in seconds. Michigan law hasn't moved to match, and the software's judgment hasn't either — here's what these tools are actually good for and where a person still has to read the fine print.

Rennie Barton · September 14, 2026 · 6 min read

In this article

An AI tool can now produce a complete purchase agreement in about a minute — price, contingencies, a proposed closing date, even a paragraph explaining why the number makes sense. The output reads clean and confident, and that is exactly the difficulty: a draft that is wrong in two places reads precisely like one that is right. That's roughly where agentic AI sits in real estate today — genuinely useful, and not yet something you hand the pen to.

The tools people lump together under "AI agent" cover a wide range. Some are a chatbot wrapped around a template. Others, like the AI mode Zillow rolled into beta this spring, chain together search, cost estimates, and scheduling into something that behaves like a personal assistant. None of them are licensed to practice law. In Michigan, the line between what they can do and what they shouldn't has been on the books since before any of this software existed.

The real, working end of this is narrower than the marketing suggests. Conversational search tools that compare listings, estimate repair costs, and explain what you can afford are live and reasonably good. Document summarizers that condense a hundred-page disclosure packet or title commitment into a readable page are live too, and useful as a first pass. Drafting assistants that pre-fill a standard purchase agreement with the facts you give them — names, price, dates, the property address — do that mechanical part faster than a person typing it by hand.

  • Conversational search that compares homes and estimates costs, the way Zillow's new AI mode does in beta
  • Summarizers that condense a disclosure statement or title commitment into a shorter, plainer read
  • Drafting tools that fill the blanks on a standard, pre-approved form from facts you supply
  • Not yet real, whatever the pitch deck says: an AI that reliably negotiates on your behalf, or one you can trust to catch every legal issue in a title commitment on its own

Michigan's unauthorized-practice-of-law statute, MCL 600.916, has been in effect since 1963. In 2003 the Michigan Supreme Court carved out what's known as the scrivener's exception, ruling in Dressel v Ameribank that filling out a standard legal form isn't the practice of law unless legal knowledge or discretion goes into completing it. That's the exception real estate agents already work inside of every day — they can enter the price, the dates, the parties' names on a board-approved purchase agreement. What they can't do is rewrite a clause, add a term the form doesn't already contemplate, or tell you what a contingency legally means for your situation. Under Dressel, legal advice requires "legal discretion and profound legal knowledge," and only a licensed attorney is allowed to supply that — not your agent, not your lender, not a chatbot.

AI doesn't get a pass just because it's software instead of a person. The State Bar of Michigan's AI Workgroup said as much in its June 2025 report, and also admitted the honest complication: Michigan's statute is written around a "person," so whether an AI program can even be found to violate it is a question the Bar hasn't settled. That's not a loophole worth counting on. It means the legal exposure sits with whoever is using the tool and putting their name on the offer, while the boundary itself — fill in facts, don't supply legal judgment — hasn't moved at all.

Michigan sellers are required to give buyers a written disclosure statement before a binding purchase agreement exists, under the Seller Disclosure Act. An AI summarizer can compress that document, and a title commitment alongside it, into something you can read in five minutes instead of forty. What it's weighing when it summarizes, though, is usually salience in the text — which sentences look important by word frequency or position — not legal consequence. A throwaway line about a prior repair can matter more than three paragraphs about paint colors. On a title commitment, the exceptions are the whole ballgame: an old easement, a mineral reservation, a restrictive covenant. A model can list those exceptions accurately and still miss that the one buried on page eleven conflicts with the dock you're planning to build. Reading how an exception interacts with your actual plans for that actual lot is still a judgment call, and it's the kind AI hasn't earned trust for yet — a lesson the industry learned the hard way when the FTC filed a complaint in 2024 against a company that had marketed its chatbot as a "robot lawyer" without ever testing whether its output held up.

The workable version of this looks less like delegation and more like preparation. Run the disclosure and the title commitment through a summarizer before your own read, so you show up to the real document already knowing what to ask about. Let a drafting tool assemble the mechanical first pass of an offer, then have your agent and, where the stakes warrant it, an attorney go through it line by line. Use it to check your own list of contingencies against what similar deals typically include, not to decide which ones you need. The tool is a fast, tireless first reader. It is not, under Michigan law or under common sense, the last one. If you want a plain-language starting point for the property records a lawyer or title company would pull anyway, our Property Research Center is built for that — it's the county and state sources, not a substitute for reading the actual documents with someone qualified to judge them.