Answer Hub · Software & workflow
What controls should a lender set before connecting an AI assistant to a vendor?
- Reviewed by
- Reviewed by RAZE
- Last reviewed
Short answer
Common controls are simple. Decide which assistants your team may connect. Give each user the RAZE role the job needs and no more. Keep confirmation on for anything that creates or changes a record. Limit who can add connectors. Have a reviewer check summaries against the report.
Five controls to settle.
This page is general education, not legal or compliance advice. Your institution's own policy and counsel decide what applies to you. These are the questions a lender's team commonly works through before an AI assistant touches a vendor's system.
- Which assistants may connect. Name the assistants your institution has approved, and say who approves a new one.
- What each person can do. An assistant should act as a named user with that user's permissions. Give each user the role the job needs and no more.
- What happens before a write. Decide whether the assistant must show a summary and wait for a yes before it creates or changes a record, and whether a user may turn that off.
- Who can add connectors. Many business plans let an administrator limit who adds a custom connector. Use that setting.
- Who checks the output. A summary is a starting point for review. Have a reviewer check it against the source document.
Two questions to ask every vendor.
Where does our data go? Data a team requests through a connector moves from the vendor to the AI assistant the institution connected. How the assistant stores or uses it is governed by your agreement with the assistant's provider and your plan's administrator settings. Include those terms in your vendor review.
What is logged, and where? Ask whether an action taken from an assistant leaves the same record as one taken in the vendor's own screens.
How the RAZE MCP fits these controls.
The Model Context Protocol (MCP) is an open specification, published at modelcontextprotocol.io, for how an AI assistant connects to outside software. The RAZE MCP is RAZE's server on it. This is what it does today:
- Access. Your team connects with the RAZE credentials issued under your agreement. Each person signs in with their own RAZE credentials, and the assistant can do only what that user's RAZE permissions allow.
- Confirmation. By default, the RAZE MCP shows a summary and waits for your confirmation before it creates or changes anything. A user can tell the assistant to skip that step, so set your team's rule: keep confirmation on.
- Write tools. The tools that change data, such as draw sheet updates and uploads, are managed in the RAZE MCP connector options.
- Logging. Actions taken through the RAZE MCP are logged in RAZE.
- Practice. The RAZE MCP works with the Stage sandbox, where orders reach no live inspector and nothing is billed, so a team can practice there first.
The RAZE MCP does not approve draws. Your reviewer accepts or rejects each submitted inspection, unless your institution has turned on the auto-approve workflow setting, and the funding decision stays with the lender.
The RAZE MCP FAQs go through each of these, and the setup page has the steps for each tested assistant. To begin, Get Started.
Sources
Primary sources only. Where a figure appears on this page, it comes from one of these.
See the workflow behind the answer.
RAZE standardizes how draw inspections are ordered, dispatched, captured and delivered, inside the loan administration workflow a lender already runs.
Get Started