Eleven Questions to Take to Compliance Before You Draft With AI

Most articles about AI in mortgage will confidently tell you what the rules are. This one will not, and the reason is worth stating up front.

Mortgage communication sits on top of federal lending law, federal and state advertising requirements, state law that varies by jurisdiction, your investor’s overlays and your own company’s policies. Those interact differently for every lender, and they change. A general article cannot account for all five layers as they apply to a particular lender.

What a general article can usefully do is tell you which questions to ask, so that the answers come from your compliance team, your counsel or your broker-owner rather than from a blog. Here are eleven. The harder part is often recognizing that the questions need to be asked at all. A question nobody asks looks exactly like a question with no problem behind it.

Before you use the tool at all

1. Which AI tools, if any, am I approved to use for borrower communication?

This is the first question and it is often the shortest conversation. Some lenders have an approved tool. Some prohibit all of them. Some have not decided, which is not the same as yes.

2. What is the vendor allowed to do with what I type into it?

Retention, training use, subprocessors, geography. Your compliance team may already have reviewed this for the approved tool. If you are using something they have not reviewed, they cannot have.

3. What borrower information am I permitted to enter, and what must be removed first?

Get this as a list rather than a principle. And ask the follow-up that people skip: at what point does a combination of details that contains no name still identify a borrower?

Before you send anything it drafted

4. What is our review requirement for AI-assisted communication, and is it different from ordinary correspondence?

Some lenders treat an AI draft exactly like anything else you write. Some require a second reader. Ask rather than assume.

5. Which of our communications require approved wording, an approved form or specific disclosures?

If you only get one of these answered, make it this one. Knowing which messages are free-form service copy and which are not is the difference between a tool that saves you time and a tool that rewrites something it should not have touched.

6. What terminology am I required to use for loan status, and which words are prohibited?

Prequalified against preapproved, conditional against clear to close, what may be called an estimate. AI drafts drift toward warmer, more confident language, and that drift lands exactly on these words.

7. What am I allowed to say about why underwriting requested something?

Including whether there is an approved neutral sentence you can use when you do not have an explanation you are permitted to share.

Advertising and marketing

8. What counts as advertising here, and what triggers additional requirements?

Many officers do not realize that a social post with a payment example, a co-branded flyer or a newsletter mentioning a rate can be treated as advertising. Ask where the line is at your company.

9. What is the approved source for any rate, payment, fee or savings figure, and what review does a piece containing one require?

The rule that matters practically: no figure produced by a model goes into anything public. Confirm the process for the figures that do.

10. What do our required identifiers and disclosures look like on each channel?

And ask directly whether including an NMLS number and an Equal Housing mark is sufficient, because a lot of people assume it is.

Adverse action

11. What is my role, and what is explicitly not my role, when the answer is no?

Denials and counteroffers are the last place you want a fluent draft filling a gap. Ask what your process is, who issues what, and what if anything you are permitted to write alongside it. Ask it before you need the answer rather than on the afternoon you do.

Why the questions matter more than the answers

Every one of these has a different answer at a different lender, and several have different answers in different states. That is the point.

The failure mode with AI in a regulated business is rarely someone deliberately cutting a corner. It is someone using a tool competently, in good faith, without ever having established which parts of their job the tool was allowed near. The draft is fluent, it looks professional, and nobody notices that a piece of approved wording quietly got improved.

Getting these eleven answered turns that risk into a known boundary. It can also make it easier for compliance to distinguish permitted drafting uses from restricted ones.

What is safe to assume in the meantime

Two things, both boring.

Anything the tool produces gets read against the source before it goes anywhere. Not skimmed for tone. Compared, line by line, against the verified facts you supplied and the approved language you were given.

And prompting reduces the chance of an unsupported statement rather than removing it. Instructions can be missed, partially applied, or outweighed by something later in the conversation. Treating a well-written prompt as a compliance control is a common mistake in this subject, and it is one that feels safe while you are making it.

The Claude Workflow System for Mortgage Loan Officers is built around that assumption rather than against it. Every workflow that touches a figure, a status or a reason has a bracket where you paste your own approved language, and every prompt panel carries the same reminder to verify facts, status, figures and disclosures before sending. It is a drafting system, not a compliance system, and it says so.