The pitch is always the same. Connect the AI to your CRM, let it work your database, and it will nurture every lead automatically while your loan officers sleep. The demo is genuinely impressive. Messages go out, replies come in, and nobody touched a keyboard.
We built the opposite, on purpose, and it is the single most important decision in the product.
The regulatory problem is not theoretical
A message to a borrower about a loan is not marketing collateral. Depending on what it says, it can be an advertisement subject to Regulation Z, a communication governed by the TCPA, or an act of loan origination that requires an individual with an NMLS license. The person who holds that license is a human being with a number, and that number is attached to their livelihood.
When an autonomous system sends a message that quotes a rate, implies an approval, or solicits an application, the question a regulator asks is not whether the AI meant well. It is who originated the communication. There is no good answer to that question when the answer is software.
This is not a hypothetical risk that gets priced in as a cost of doing business. State regulators examine correspondence. Plaintiffs' firms build TCPA cases out of message logs. An autonomous outbound system creates a volume of communications that no compliance department reviewed, from a sender that holds no license, at a scale that makes the exposure interesting to sue over.
The quality problem nobody talks about
Set the regulation aside for a moment, because there is a simpler argument.
The AI does not know that the borrower's father died last month. It does not know that the referring agent is difficult and needs to be handled carefully. It does not know that this borrower has been told twice already that their debt-to-income will not work at the price they want, and that a third cheerful message about rates dropping is going to end the relationship.
Your loan officer knows all of that. It is in their head, not in your CRM, and it never will be. An autonomous system that sends without them is not just a compliance exposure. It is a system that will confidently say the wrong thing to the wrong person, and the loan officer will find out when the borrower calls someone else.
What we do instead
Lia prepares. A person decides.
Lia reads the pipeline, works out who needs attention and why, and writes the message. Then it stops. The draft sits and waits. The loan officer reads it, changes what they want to change, and sends it from their own number and their own inbox. The borrower receives a message from a licensed human being who is accountable for every word in it.
There is no setting to turn this off. It is not a conservative default we relax for enterprise customers who ask nicely. Outbound autonomy is not a feature we have chosen not to ship yet. It is a thing we do not build.
The part that surprised us
We expected to lose deals over this. We assumed we would be sitting across from an owner who wanted the fully automated version and would go find a vendor who would sell it to them.
That is not what happens. What happens is that the compliance conversation, which is usually the thing that kills an AI pilot in month three, ends in about ninety seconds. There is no data governance committee to convince that the robot will behave, because the robot does not send anything. The risk officer stops asking questions, and the conversation moves on to whether the drafts are any good.
Which is the right question. It is the only question. And it is a much better place to be competing.