# The UMOS™ Advantage *Unified Mortgage Orchestration System* > UMOS runs the whole mortgage: sixteen long-horizon AI agents on one live loan model, reasoning against one version-controlled rulebook, writing into the LOS you already run. A constellation of long-horizon AI agents that execute every workflow in a loan’s life. They share one live model of the loan, reason against one version-controlled rulebook, and run continuously — asynchronously, for days, until the goal they hold is reached. ## What UMOS is ### A point-of-sale is part of UMOS, not a thing you bolt onto it The borrower’s front door is an agent inside the system. Because it sits on the same loan model as underwriting, it asks for what this file will actually need. Not a fixed form that collects the same nine documents from everyone and discovers the tenth three days later. ### It does not replace your system of record Your LOS stays your LOS. UMOS writes into it. Nothing is migrated, and nothing is re-keyed, and all audit logs and decisions pushed to the same LOS. ### It runs the manual middle Conditions. Income. TRID timing and cures. Pre-close QC. Boarding validation. Escrow. Loss mitigation packages. Claims. The steps that currently sit in someone’s queue. ## What changes | From | To | | --- | --- | | **Forms** — The POS asks the borrower what they think they need. | **Anticipation** — The system knows the document set from loan type, investor and borrower profile, validates real-time and walks the borrower through their options, in a customised way that speaks your brand. | | **Queues** — Work waits for a human to pick it up. | **Continuous execution** — An agent holds the goal and drives it, overnight and over weekends. | | **Tools** — You buy features and assemble a workflow. | **Desks** — You staff a role, and your team oversees the agents' progress. | ## What it is not - **Not a system of record** — No migration, no re-platforming. - **Not a copilot** — Nobody has to prompt it. It is not waiting to be asked. - **Not an RPA script** — It re-plans when the file changes, which a recorded macro cannot do. - **Not a services team behind a login** — Sei runs it for you, and what runs is machines. Every decision is machine-made, confidence-scored and cited, and automation rate is a number you can hold us to. ## How it orchestrates Unified loan state. Every agent reads and writes the same live model of the loan, enabling shared context for all the AI agents to make the right decisions. ``` ┌──────────────────────────┐ events ─────────────▶│ THE ORCHESTRATOR │ (doc arrives, │ goals · dependencies · │ payment misses, │ scheduling · escalation │ fee changes, └────────────┬─────────────┘ clock starts) │ assigns goals to │ ┌────────────────────────────┴────────────────────────────┐ ▼ ▼ ┌───────────────────────────┐ ┌───────────────────────────┐ │ ORIGINATION AGENTS │ │ SERVICING AGENTS │ │ LOA · POS · Document │ │ Boarding · Escrow · │ │ Concierge · Income Desk │ │ Collections · Default │ │ Conditions · TRID & │ │ Desk · Loss Mit · │ │ Cure · Pre-Close QC · │ │ Claims Analyst · │ │ Compliance │ │ Payoffs · Compliance │ └─────────────┬─────────────┘ └─────────────┬─────────────┘ └──────────────────┬───────────────────────┘ │ read / write the same state ▼ ┌──────────────────────────────────────────────────────────────┐ │ THE LOAN GRAPH │ │ borrower · income · assets · property · investor overlays · │ │ conditions · disclosures · escrow · clocks · documents │ └──────────────────────────────────────────────────────────────┘ │ │ reasons against writes through ▼ ▼ ┌────────────────────┐ ┌────────────────────┐ │ THE RULEBOOK │ │ TRIPLE CHECK │ │ agency guides · │ │ three layers, │ │ regs · overlays · │ │ each with a "no" │ │ your SOPs │ └─────────┬──────────┘ └────────────────────┘ ▼ ┌────────────────────┐ │ THE EVIDENCE TRAIL│ │ cited · replayable│ └─────────┬──────────┘ ▼ your LOS, unchanged ``` ### 1. The Loan Graph One live model of the loan: borrower, income sources, assets, property, investor overlays, every condition and its state, every disclosure and its clock, escrow position, payment history. See: The Loan Graph — /platform/loan-graph ### 2. The Rulebook The rule set agents reason against: the Fannie Mae Selling Guide, Freddie Mac, FHA Handbook 4000.1, investor overlays and your own SOPs. Plus TILA, RESPA, TRID, ECOA, Fair Housing, HMDA, Reg X, Reg F, FDCPA, TCPA and UDAAP. Version-controlled, with an effective date. An agent never acts on a rule it cannot cite. See: The Rulebook — /how-it-works#rulebook ### 3. The Orchestrator Holds the loan-level goal, decomposes it into desk-level goals, resolves dependencies — income cannot finish before the VOE lands, a loss mitigation evaluation cannot run on an incomplete package — runs what can run now, parks what is blocked and wakes it when the block clears, and escalates to a human when a desk hits its confidence floor. See: How a decision is made — /how-it-works ### 4. The Agents Sixteen desks custom built, each holding a goal and working it until it is reached. Eight on the origination side, eight on the servicing. See: Origination desks — /for-lenders · Servicing desks — /for-servicers ### 5. Triple Check Three independent layers of verification, each able to say no. Nothing reaches your system of record on one model’s opinion. See: Triple Check — /how-it-works#triple-check ### 6. The Evidence Trail Every action, every citation, every reasoning step, kept and replayable. Examination-ready. See: The Evidence Trail — /how-it-works#evidence-trail ## A day in the life ### Origination A file in underwriting, seven conditions open, and nobody at work. > **11:42pm** A borrower uploads two of the four bank statements the Document Concierge asked for on Tuesday. The pages are classified, read and validated against the Loan Graph in under a minute. One is the wrong account — so the Concierge sends a single message naming the account it still needs and why, and does not ask again for the three it already has. > > The two valid statements unblock the Income Desk, parked since Tuesday waiting on asset seasoning. It runs, calculates qualifying income across a W-2 and a K-1, cites each figure to the page it came from, and posts a rep-and-warrant-eligible result. > > That clears two of the seven open conditions. The Conditions Manager closes them with citations and re-plans the remaining five. > > **8:30am** The processor opens the file. Two conditions cleared overnight, one document still outstanding with the borrower already chased, and a note naming the one item that needs a human. That item is a large deposit needing sourcing, flagged below confidence, with the three documents that would settle it. **Agents that work 24/7 till goal is achieved. Every hour saved is an hour shaved off in your clear-to-close timeline.** ### Servicing A transfer of 1,900 loans boards on a Friday night. > **Friday, 9pm** The Loan Boarding Agent validates every loan against the prior servicer’s data and the Loan Graph it inherits. On 41 of them the escrow cushion has mapped to the wrong field — a single mapping defect, so it is not random, it is every loan the defect touches. > > It does not silently correct them. It quarantines the 41 and files one exception naming the field, the source value, the mapped value and the rule it violates — before the first statement goes out. > > **Monday, 8am** The boarding lead has one ticket instead of forty-one surprises, and 1,859 loans that boarded clean. **Airtight rules ensure that a statement that should never have gone out is blocked for human escalation.** ## Long-horizon agents > An agent that runs for 300 milliseconds is a feature. An agent that runs for 30 days is a colleague. Most AI in this industry is request–response. You give it a document, it gives you fields. That loop is useful, but Sei’s agents act on a goal that stays open across weeks while the world changes underneath it. ### 1. It holds a goal, not a prompt The unit of work is “clear this file to close” or “cure this delinquency”, not “extract this PDF”. The agent decides what to do next; nobody queues it. ### 2. Its memory is durable The agent that picks up a file on day 19 knows what it asked for on day 3 and what came back — which is why it never asks a borrower for the same document twice. ### 3. It is woken by events, not by users A document arriving, milestone changing, a payment missing, an appraisal landing, a fee changing, a Reg X clock starting. The agent sleeps at no cost and wakes when the loan moves. ### 4. It re-plans A new condition, a changed investor overlay, a borrower who switches from salaried to self-employed mid-file, a workout that falls out — the plan changes. A recorded macro breaks here. ### 5. Its autonomy is bounded It acts inside the Rulebook. Below its confidence floor it stops and escalates to a named human with the exact question and the documents that would answer it. The floor is a number you set. ### 6. It is accountable Every step is in the Evidence Trail, cited and replayable. All traces and decision logs are stored along with the loan data. A copilot’s ceiling is the number of hours your staff spend asking it things. A long-horizon agent’s ceiling is the number of hours the work takes — including the nights and weekends when your staff are not there. Cycle time falls because the loan stops waiting, not because anyone types faster. ## The constellation Every desk runs on the same three capabilities — document intelligence, voice and call monitoring — against the same Rulebook, writing into the systems you already run. Origination conduct and servicing conduct are different rulebooks. Each desk names its own. | Origination | Servicing | | --- | --- | | Loan Officer Assistant | Loan Boarding Agent | | POS | Escrow Manager | | Document Concierge | Collections Agent | | Income Desk | Default Management Desk | | Conditions Manager | Loss Mitigation Agent | | TRID & Cure Management | Claims Analyst | | Pre-Close QC Agent | Payoffs Manager | | Compliance Agent | Compliance Manager | --- _Source: [https://seiright.com/platform/umos](https://seiright.com/platform/umos) · Sei AI_