Valon

Ditto does servicing work itself: it researches escrow and tax questions before drafting a reply to a homeowner, investigates credit reporting disputes and re allocates payments, and servicers can build new agent processes on their own workflows. The ledgers, workflows and money movement would still run without it, which keeps the AI an engine of core capabilities rather than the whole product.

The agent proposes and code decides: deterministic tools do the arithmetic, validators reject illegal actions, an output check rejects malformed results, and each action is either committed or routed through an operator who reviews it on the way out. Discretionary operations such as fee waivers and status changes are not in the agent's menu and go to a person, and after a bounded number of failed attempts the agent escalates with the gap named. The threshold that trades a wrong action against an unnecessary escalation is tuned by Valon rather than set out in figures.

Changes are measured against a golden dataset built from production traces before they ship, and the agent is monitored in production for clusters where it gets stuck. Accuracy rates for Ditto are not published, and the validation work is not tied to a supervisory framework.

Every run keeps the context the model received, the code it wrote, the validation results and the outcome, rerunning a trace gives the same result, and Valon reports that its validator rejects about one proposal in five, most of which escalate.

Figure ran an A/B test in which partner loans serviced on Valon's technology showed a 40 percent improvement in delinquency outcomes against the control group. That result measures servicing on the platform rather than the agents, and Valon gives no test period or loan count for it. Newrez, Carrington Mortgage Services and ServiceMac are named customers, with Rithm's chief executive quoted on ValonOS.

Valon says it signed more than $200 million in annual recurring revenue in its first six months on the market. Its operating figures, from 60 percent plus margins to more than 3,000 loans per employee against an industry average of 750, come from the servicer Valon ran itself.

Agents are described as working within each servicer's workflows, data and controls, an assurance that does not say whether one servicer's data or agent processes improve the models another servicer uses. Model hosting and retention of agent inputs go unstated.

Retention, sharing and subprocessor terms for the borrower data ValonOS holds are not set out in its product material, for a system carrying millions of homeowners' records. Data integrity is the theme instead: one ledger, reconciled records and a full audit history.

Neither the auditor nor a way to obtain the reports is given. SOC 2 Type 1 and Type 2 examinations are listed as completed beside the platform, in a badge row that also carries agency approvals and state licensing.

ValonOS, as software sold to servicers, states no license or supervisory admission of its own. The Fannie Mae, Freddie Mac and Ginnie Mae approvals and licensing in all 50 states shown beside the platform are standing earned by the servicer Valon ran and has since sold.

Fairness or bias testing of agent work is not published, though agents draft homeowner communications and work disputes and payment questions. Governance is asserted rather than shown: a former Fannie Mae chief executive describes ValonOS as having governance built in so servicers can deploy AI safely.

A guarantee, indemnity or correction commitment behind an agent's work is absent. What stands in its place is traceability: every material action keeps its actor, time, approval and record, so a servicer can reconstruct and contest what happened on a loan.

The models behind Ditto and the agent building platform go unnamed, as do hosting providers and subprocessors. Valon describes the architecture instead: agents and people work from the same source of truth under the same guardrails inside ValonOS.

ValonOS is itself the servicing core, replacing legacy mainframe platforms at named servicers including Newrez and Carrington, which adopted it as its core servicing platform. Connections to investors, custodians, vendors and partners are offered out of the box, though none is named.

Where ValonOS runs, in which regions and under what residency terms is not described, an open question for servicers whose investors and regulators ask where borrower data sits.

Price, tiers and the unit of charge for ValonOS are unpublished, and the only buying route is a conversation with Valon. Its contracted revenue is public, which sizes the business rather than a servicer's bill.

Mortgage servicers are the buyers, from the largest United States servicers to government loan specialists such as Carrington, and the material is built around servicing work: payments, escrow, defaults, reporting and audits. A long list of other servicing businesses, from student loans to insurance claims, is named as where the platform is designed to extend, without material for any of them yet.