Polly

The business would stand whole without its AI. The pricing and eligibility engine would still carry hundreds of investors' products, still run automated lock desk workflows and still map into the lender's origination system. The loan trading exchange and analytics products would be untouched. That engine is rules and eligibility logic evaluated against loan attributes, which is deterministic work, and it is what customers buy.

The AI line is real and sits inside the operation rather than beside it. The loan officer agent reasons over the engine's own ineligibility logic to surface the closest qualifying product, which directly changes what a borrower is quoted. But it arrived in 2024 on top of a platform that already worked.

No control is described at any point, and the material asserts an outcome instead. The agent examines near miss eligibility, interprets ineligibility logic and proactively articulates specific recommended actions, and the only assurance offered about its operation is a customer's observation that it enhances rather than replaces the human connection, which is a description of intent rather than a mechanism.

Nothing states a confidence threshold below which the agent declines to recommend, what happens when it misreads an eligibility rule, whether a recommended price is verified against the engine before an officer repeats it to a borrower, or whether any category of recommendation requires review. The stakes make the silence conspicuous: the output is a price and a product recommendation that reaches a consumer through a loan officer.

Nothing measurable is published about how the models behave. The material describes the technology in category terms, citing AI, machine learning and natural language processing alongside proprietary technology and data. It offers no accuracy rate for eligibility interpretation, no error rate on near miss identification and no evaluation results. Nor does it offer a retraining cadence, ISO 42001 certification or validation documentation.

Competitors in loan level mortgage analytics compete substantially on documentation, published back testing and validation resources, and this vendor publishes none of it while making strong AI claims.

Named institutions appear beside specific measured results. ResiCentral, a wholesale mortgage company, is reported to have tripled year on year volume in 2024 without adding secondary, capital markets or lock desk staff, which is an outcome stated in a form that would be visible to the customer and awkward to overstate.

At New American Funding, a large national lender, more than 1,250 employees are stated to be actively engaged with the platform and it is credited with providing clarity on more than 7,200 ineligible products, and a named manager at the company is quoted describing the change in how officers quote. A third lender, American Financial Resources, is named as adopting the pricing engine in 2025.

Venture backing adds parties with money at stake, with a $37 million Series B and a further $25 million raised. The residual weakness is that every figure reaches the reader through the vendor's own announcements rather than a channel the customer is accountable for.

Safety practice, output screening, red teaming and a data boundary all go undescribed. The multi client question is pointed because of what the platform holds: pricing strategy, margin settings and lock behavior for lenders who compete directly with one another, alongside product and pricing data from hundreds of investors.

Nothing states whether one lender's pricing behavior informs recommendations produced for another, whether a lender's configured strategy is isolated from the models, or what happens to that data when a lender leaves. At least one competitor answers precisely this question by stating its agent is trained only on its own materials, so the question is answerable and Polly has not answered it.

Polly publishes no privacy policy detail, processing agreement, subprocessor list or retention schedule. The engine evaluates loan level borrower attributes to determine eligibility and price, which places credit, income and property characteristics inside the system for consumers who never chose this vendor and will never know it was involved.

Nothing states how long pricing scenarios and lock records are retained, what a lender can require to be deleted, or whether borrower attributes submitted for a quotation that never becomes a loan are kept at all.

There is no public certification, attestation, trust center or security page. The data position makes this a live question rather than a formality. The platform holds borrower eligibility attributes and lock positions for national lenders, and separately holds the pricing and margin behavior of institutions that compete with one another. Its buyers include banks and credit unions whose own third party risk processes require documented evidence before onboarding, so this is a gap a purchasing committee would raise immediately.

The company is a technology provider to licensed mortgage lenders and holds no license, registration or supervisory standing of its own, the usual posture for a technology vendor. No regulator engagement, examination or industry accreditation is published. Its position sits adjacent to supervised activity, since pricing, eligibility determination and lock desk practice are all subject to examination at the lender, and the vendor supplies the mechanism without being examined itself.

The exposure is direct, and no fairness evaluation, differential outcome analysis or governance disclosure appears. The agent recommends which product a borrower should be offered and at what price, and near miss reasoning specifically identifies borrowers who fall just outside a product's criteria and suggests actions to bring them inside it. Which near misses the system surfaces, and which it does not, is a determination about who gets offered a better product.

Fair lending is the most closely supervised area of United States mortgage. Nothing published examines whether recommendations, near miss identification or suggested actions distribute evenly across borrower characteristics, geographies or channels.

No liability position, error rate, remediation commitment or correction route is published. The consequence path is short and concrete: the agent generates a price and product recommendation, a loan officer relays it to a borrower, and a wrong eligibility interpretation means a consumer is quoted a product they cannot have or is not shown one they could.

Nothing states who bears that, whether the lender or the vendor stands behind a quoted price the engine produced, or by what route an error is identified and corrected once a borrower has acted on it. The borrower has no relationship with the vendor and no way to know it was involved.

The model behind the agent is unnamed, with no provider, family, version or hosting arrangement disclosed. The description consistently pairs AI and machine learning with proprietary technology and data. That fits a common market pattern, where vendors that license their model capability tend to name suppliers and vendors that call it proprietary usually name nobody.

The natural language and voice interaction the agent offers is exactly the kind of capability a company of this size would ordinarily license. A lender cannot tell from published material whether borrower scenarios typed or spoken into the agent pass through a third party.

The right system of record is named as a category and the specific products are not. The engine maps both standard and custom fields directly into the lender's loan origination system (LOS) through published interfaces. That is the correct integration target for pricing and lock workflows and more than a generic connectivity claim, and a separate partner platform extends the ecosystem.

On the supply side the engine ingests and maintains product and pricing data from hundreds of investors, a substantial integration burden in its own right and a quantified one. No origination system, servicing platform or investor system is identified by name. There is no integration count, certified partner directory or public API documentation.

Cloud native is the only deployment statement, repeated throughout as a differentiator against older on premises pricing engines, and it describes an architecture rather than a location. No hosting region, residency commitment, tenancy model or single tenant option is published, and nothing states whether lenders competing in the same markets share infrastructure. Availability on web and mobile is described as a user convenience with no accompanying statement about how data is handled on a device.

Price, tiers, minimums and billing basis are all unpublished. The platform is described as offering near infinite configurability and unlimited flexibility through customizable options, which signals a configured enterprise sale. A buyer cannot tell whether charging is per user, per lock, per loan, by volume or by subscription. Those imply materially different economics for a lender whose volume swings with rates, and the implementation effort implied by that configurability is never sized either.

The buyer set is varied within its market and the market is narrow. Banks, credit unions, independent mortgage lenders, wholesale operations and the brokers they serve all appear. The named references span a large national retail lender, a wholesale company and a further national lender, which is three distinct operating models rather than one described three ways. Hundreds of investors have products carried on the engine, which is a distinct form of reach on the supply side.

Coverage is bounded to a single asset class and a single country, with nothing outside United States residential mortgage. No client count, volume figure or market share is published to size the adoption.