AI Tools

Agentic AI for mortgage is software that carries out multi-step operational work on a loan file on its own — locating documents, extracting data, cross-referencing a system of record, checking results against a requirement, and recording what it did — rather than answering questions about the file or extracting data for a person to act on. The distinction is whether the software completes the task or prepares it.
Areal Copilot Agent passed 10 million such tasks on production mortgage loans this week. It passed 4 million in June. Every one ran inside a live lender operation on a real file moving toward funding.
A number like that is worth exactly as much as its definition, so the rest of this post is the definition.
A task is one discrete, completed unit of operational work on one loan file. Classify this page. Confirm this signature is present and legible. Match this fee against the one in the LOS. Verify this post-closing package against what the receiving investor actually requires.
Several agents can work the same file at once without waiting on each other. So 10 million tasks is not 10 million loans, and it is not 10 million agent runs — it is the individual unit of work, counted cumulatively since the platform went live.
That matters because volume numbers in this category are quoted in at least three incompatible units, and the differences are large. Areal's 10 million is counted in tasks, which is the same unit as the 4 million figure published in June. The two are directly comparable, which is the point of using the same unit twice.
Take one: confirm the notary stamp on the deed of trust.
The agent locates the deed of trust among several hundred pages in the closing package — which is not trivial, because it does not arrive in the same position, under the same filename, or with the same cover sheet twice. It finds the notarial block. It checks that a stamp is present, that it is legible, that the commission has not expired, and that the name on it matches the signer. It records the result, the document it read, and the rule it applied, on an execution bench a person can open afterwards.
If everything holds, nothing reaches a human. If it does not, the exception says which check failed and which document it was reading — so the person who picks it up starts from a question rather than from the beginning.
That last property is the whole design. An exception that arrives with its reasoning attached is a two-minute task. An exception that arrives as a red flag is a re-review, and automation that produces re-reviews has moved the work rather than removed it. An agent that completes 10 million tasks unreliably has not completed 10 million tasks. It has created 10 million things to check.
Five questions separate the category.
Areal Copilot Agent ships with out-of-the-box agents for borrower onboarding, funding review, post-closing review, insurance review, appraisal review and title review, among others, and supports agents a lender writes themselves. It launched at MBA Annual in October 2025 as the industry's first agentic AI for mortgage origination platform, and runs at 99% accuracy on critical fields. If the underlying concept is new to you, the agentic AI primer covers it from first principles.
Not the number of people. What reaches them.
Quality rose with volume. As execution grew from 4 million to more than 10 million tasks, review outcomes improved alongside it. At peak, 94 of every 100 loans in post-closing review needed no second opinion before the file was ready for investor shipment. Whole loans, not tasks, and a peak rather than a daily average. When a loan does stop, the agent names the check it could not clear.
The cost that disappears is not really the minute each task took. It is the switching tax: re-orienting and re-concentrating dozens of times per file, loan after loan, which is where small things slip and where month-end turns into overtime. When the routine work completes on its own, the month-end delivery queue stops scaling with loan volume and starts scaling with exceptions. That is a materially different month.
The savings compound in one direction only — seconds per task into hours per loan, hours per loan into files cleared, files cleared into a flatter month-end. Areal's published figure for the full platform is 5 to 7 hours recovered per mortgage. Areal CD Balancer, the second flagship product, covers CD balancing and fee reconciliation at the closing table and carries its own separate measurement.
Ten million is the volume. Quality rising with it is the part that matters.
Software that completes multi-step operational work on a loan file on its own — locating documents, extracting data, cross-referencing the LOS, checking against a requirement and recording the result — rather than extracting data for a person to act on. The test is whether it completes the task or prepares it.
A task is one completed check. An execution is one agent run, which contains many tasks. A loan is a whole file. Per loan there are many tasks, a handful of executions, and one file — so a volume number is meaningless until you know which unit it uses.
Areal Copilot Agent, launched at MBA Annual in October 2025.
No. Areal runs as an AI layer on existing infrastructure with native integrations into ICE Encompass, MeridianLink and Byte LOS. No migration is involved.
What unit the volume number uses; whether the platform completes work or prepares it; what an exception tells the person who receives it; whether the reasoning is readable afterwards for audit; and whether it requires an LOS migration.
We'll show you which workflows to automate first — and the exact ROI for your volume.
Book a Demo →