AI answers a loan officer can check.

The 60-second version.

The lender rules and figures in these screens are sample data.

Does this loan fit this lender?

A loan officer's question is usually about one borrower and one lender: does this loan fit this lender's rules? That means the agency guideline (Fannie Mae, Freddie Mac, FHA) plus the lender's own overlays on top of it.

Guideline Buddy answers from the guidelines loaded on the account, and from documents the loan officer attaches, such as a credit report or DU findings.

Many guidelines, one chat box.

The hardest part was the chat screen itself. It had to hold every guideline a loan officer might pick and still feel as simple as typing a question.

In the earlier design, which was live before the redesign, you picked a module first (Guideline, Income, Compliance, Marketing or Eligibility Buddy). Then you made your picks from dropdowns, such as a lender partner and a guideline, and only then typed.

In the redesign, choosing guidelines is part of writing the question:

• The composer has two buttons, Documents and Guidelines. Whatever you pick sits in the composer as a chip you can remove, so the scope of the next question is visible where you type.
• You can pick one guideline, several, or all of them. A question asked across all of them says "Searched all 24 of your guidelines".
• Every answer opens with what it used, such as "Based on: DU Findings - Credit Report - Fannie Mae - UWM".
• With nothing selected, Guideline Buddy still answers. It says the answer is general and offers to add guidelines so it can be checked for a specific program.

Guideline BuddySample data

01 — Before: pick a module, then a lender partner and a guideline, then ask.Earlier design

Guidelines - picker openSample data

02 — Guidelines are picked in the composer.Chat redesign

Answers that show the rule.

• Requirement by requirement. An eligibility answer opens with the verdict, then lists each requirement against the lender's limit, such as "712 against a 680 minimum", marked "Met".
• Sources quote the rule. Each source shows the guideline, the section, its date and the rule itself in quotes, with a link to open the guideline. I decided on this because a loan officer can't act on an AI's paraphrase. With the exact wording on screen, they can check the answer themselves and point to the rule when an underwriter asks.
• Asking again. If someone adds a guideline after an answer, the answer says it wasn't based on that guideline and offers to ask again. The team decided this; I designed it.
• Changes in the thread. When guidelines are added or removed mid-conversation, the thread says so: "Pennymac added to this conversation - earlier answers did not include it".

Grounded answer with sourcesSample data

03 — Each requirement against the lender's limit.Chat redesign

Sources expandedSample data

04 — Sources quote the rule, with its section and date.Chat redesign

Context persists and changesSample data

05 — The thread notes when the guidelines in play change.Chat redesign

Beyond one lender.

• Comparisons. "Compare gift rules between Fannie, Freddie and FHA" comes back as a table showing who can give and the key difference.
• Matches and near misses. For a borrower's scenario, Guideline Buddy lists three things:
  • the lenders who can take the loan as it's structured
  • the lenders who are one requirement away, such as "FICO short by 20"
  • the changes that would bring more lenders in, such as "Reduce LTV to 70%", marked "+1 lender"

  The team decided on matches, near misses and suggested changes; I designed them.

• Document review. Attach a credit report and DU findings and ask what's wrong. Each issue names the document it came from, down to the page for the credit report ("Credit Report, p.2").

Comparison answerSample data

06 — Gift rules for Fannie Mae, Freddie Mac and FHA in one table.Chat redesign

Qualification - matches and near missesSample data

07 — Lenders who fit, lenders one requirement away, and the changes that bring more in.Chat redesign

Documents + guidelinesSample data

08 — Guidelines and documents sit in the composer as chips. The answer says what it used, and each issue names its source document.Chat redesign

The first chat.

A ChatGPT user already knows what a chat model does. A loan officer has no way to guess that Guideline Buddy can compare programs, scan every lender or review a document.

So I decided a new account's empty chat shows three suggestions: "Compare Fannie vs Freddie", "Which lenders allow 12-mo bank statements?" and "Review my DU findings". That teaching job is temporary. The suggestions retire once the account has a handful of conversations, and from then on an empty chat shows only the greeting and the composer.

Empty chat - first runSample data

09 — A new account's empty chat: three suggestions that show what it can do.Chat redesign

When an answer is wrong.

Escalation tickets were my call. Feedback told us an answer was bad, not what was wrong with it, and in mortgage a wrong answer can become a compliance problem.

• A user files a report by picking the chat, then the exact question and answer, and the ticket carries the guidelines that answer used.
• They set the severity and the business impact, and write what the answer should have been.
• The ticket shows the disputed answer next to the expected one, and the Guideline Buddy team replies on the ticket. Each ticket moves from submitted to under review to resolved.

Tying each report to the exact question and answer, with what it should have said, gives the team something specific to fix and gives the user a reply.

Escalation ticket workflowSample data

10 — A report starts from the exact question and answer.Support

Escalation ticket workflowSample data

11 — The disputed answer next to the expected one, with the team's reply.Support

The admin side.

• Bots and knowledge bases. Admins build bots for their team, starting from templates such as Mortgage Guidelines, Compliance Checker or Rate Sheet Analyzer, and add their own documents. Every lender and broker works from its own overlays and documents, not only the agency guides everyone shares. Each bot shows its status (active or draft), the documents read next to the documents added, and when it was last updated. The read count tells an admin the bot has taken in what they uploaded.
• Question history and feedback. For any user, an admin can see every question with the guideline it used and the date. They can open a question to read the answer and the user's feedback, and review all feedback in one place. This shows admins what their team actually asks and which answers people flagged. It tells them what to add to their bots and where their team needs training.

BotsSample data

12 — Bots start from templates, and each shows documents added and read.Admin

Prompt detailSample data

13 — Question history, with one question opened: the guideline it used, the answer and the user's feedback.Admin