SitusAMC

Graded on the Clearwater precedent for a long established platform carrying a genuine embedded artificial intelligence line, and on the services hybrid rule, which judges a services firm on the product it also ships. Acuity qualifies as that product and the company describes it in model terms: machine learned optical character recognition performing document classification and data extraction, running across loan files at the front of the diligence pipeline.

What keeps this at the lower grade is what survives its removal. The company leads its own description with outsourcing, advisory and talent, and the diligence itself is performed by reviewers; strip Acuity and roughly 870,000 review scopes still complete, more slowly.

The surrounding technology portfolio is deterministic by design, since a compliance rules engine, a document custody system, a warehouse lending system of record and a pricing engine contain no models and are not supposed to. Revisit if the extraction and classification layer starts performing the review rather than preparing it.

Human review is structural here rather than promised, since the business performs hundreds of thousands of review scopes with reviewers and Acuity prepares files for them by indexing and stacking documents so the most material items appear first.

One published control is genuinely well designed and worth recording: in the government agency engagement, files that Acuity flagged as missing critical documents were placed on hold and deliberately not reviewed until the documents arrived, while complete files were prioritised. That is a named gate with a stated position in the workflow.

Held off the top grade because it is described once in a case study rather than published as a product property, and nothing states what confidence threshold governs a classification, what happens when the model records a document as present that is not, or how a misclassification is caught before the exception report goes out.

Nothing on validation, accuracy, error rate, monitoring or governance of the machine learned classification and extraction was located. The consequence chain makes the silence more material than usual. Acuity determines which documents a loan file is recorded as containing, that record feeds the exception report, the exception report informs a rating agency's view of the collateral, and the rating informs what investors pay.

A classification error at the front of that chain is not visible anywhere further along it. A published accuracy figure for document classification would be an unremarkable disclosure for a vendor operating at this scale and none exists.

Production scale is published as a dated figure rather than a claim: roughly 870,000 unique review scopes across roughly 460,000 unique residential loans in 2025, alongside support for more than 20 billion dollars of commercial loan originations annually and a technology organisation of more than 650 staff.

A published case study describes running upward of 50,000 loan files for a government agency through Acuity in the first three to five weeks to identify which loans were missing critical documents. The strongest evidence property is independent rather than self reported: SitusAMC holds rating agency approval as a residential loan level diligence provider, which is a formal admission process operated by parties whose own published ratings depend on the quality of that diligence, and its Clarity compliance system is stated to be relied on by state and federal regulators as well as lenders. Individual client names are not published, which is the norm in diligence work and is the one thing keeping this short of the strongest possible showing.

SitusAMC sits on both sides of the same market and says so: Acuity is presented as useful for purchasers assessing a pool and equally for sellers preparing one, and the firm advises, underwrites, values and reviews for counterparties who transact with each other. Combined with a share of United States secondary market diligence large enough to be measured in hundreds of thousands of loans a year, that creates a data position no participant could assemble independently.

Nothing states whether loan level data or analytics derived from one client's pool inform work performed for a counterparty, or what separation exists between engagements on opposite sides of a trade. This is the same structural shape the index recorded where a vendor sold to both supervised institutions and their supervisors.

This is the widest consumer privacy exposure of any vendor assessed in this session and the public disclosure is the thinnest. A residential mortgage loan file is the most complete financial dossier that exists on a household, carrying income, employment, assets, liabilities, credit history, identity documents and property detail, and SitusAMC processed roughly 460,000 of them in a single year on behalf of third parties who are neither the originator nor the borrower.

The federal financial privacy regime applies squarely to that processing. No data processing agreement, subprocessor list, retention schedule or borrower facing privacy statement was located publicly. The borrower has no relationship with SitusAMC, did not consent to its involvement and in most cases will never learn the review occurred.

No certification, attestation, trust centre, penetration testing statement or subprocessor list was located in this pass. That is a disclosure finding rather than an assessment of the underlying controls: a firm handling mortgage loan files for banks, government agencies and rating agency reviewed securitisations would be very unlikely to operate without a service organisation control report, and its absence from public material is more probably a decision not to publish than an absence of the report. Worth confirming on a later pass, since the difference between holding a certification and publishing it is precisely what this axis measures.

Three distinct pieces of standing, more than almost anything else in this index carries. SitusAMC is approved by rating agencies to perform residential loan level due diligence, which is a formal admission process rather than a self assessment, and the securitisation market treats that approval as a precondition. Its Clarity compliance system is stated to be trusted by state and federal regulators as well as by lenders, so a supervisory body is a user rather than merely an observer.

The valuation business is separately described as independent and licensed. Held off the top grade under the standing index ruling: none of this is a supervised regulatory test of the artificial intelligence product itself. The approvals attach to the diligence services and the compliance system, and Acuity has passed no equivalent process.

The loans under review are consumer mortgage credit, which places this work inside fair lending territory even though the duty falls on the originator rather than the reviewer. Clarity is an automated loan level compliance system that determines whether loans meet regulatory requirements and is relied on by lenders and regulators, so its rule coverage is itself a fair lending detection mechanism.

The company's own commentary acknowledges increasing scrutiny of automated decisioning and data governance in the origination lifecycle. Against that, nothing public describes how its own automated systems are tested for differential effects, what the compliance system covers and omits, or how a systematic gap in automated review would be detected.

No error rate, remediation commitment, liability position or correction path is published for the classification and extraction layer. The recourse question here is unusual because the party ultimately affected is furthest from the vendor: a misclassified document produces a wrong exception, which informs a securitisation review, which informs a rating, which informs the price investors pay.

The borrower whose file was misread has no visibility and no standing, the investor relying on the rating has no line of sight to the extraction step, and nothing states who bears the consequence when the classifier is wrong.

The extraction capability is described as machine learned optical character recognition, which names the technique and no supplier. No model provider, family, version, training data description or country of processing is disclosed.

The technique named is an older and narrower one than the large language model stack most of this pocket now runs on, which suggests an in house or long standing licensed component rather than a recent foundation model integration, but nothing states it either way and the inference should not be treated as disclosure.

This vendor does not integrate with core systems so much as supply several of them. The portfolio includes what the company describes as the most widely used warehouse lending system in the industry, the most trusted document custody system in the industry, a comprehensive system of record for the management and accounting of loan related assets, loan pricing, underwriting and delivery software, a conduit counterparty and pipeline management solution for aggregators, a pricing engine for reverse mortgage aggregators, and an automated loan level compliance system. Acuity sits across those as the document classification and extraction layer. Owning the system of record at several points in the loan lifecycle is the deepest form of what this axis measures.

No region list, residency commitment, hosting description, single tenant option or self hosted path was located. Operations span the United States and Europe with a named European leadership team, so processing plainly occurs in more than one jurisdiction, and the material handled includes consumer mortgage files subject to differing national requirements.

Much of the work is delivered as a managed service rather than as software the client runs, which shifts the residency question from where the buyer deploys to where the provider operates, and that is the version nothing addresses.

No pricing, tier structure, rate card or indicative range is published for either the technology products or the diligence services, and the route to a number is a contact form or a demo request. This is the index norm and is measured against Sumsub, which publishes per verification rates on a public page.

The pattern is more entrenched in outsourced diligence than in software, since engagements are scoped and priced per transaction, but a per loan review rate is exactly the kind of figure a buyer comparing providers would want and no provider in this category publishes one.

The broadest institutional footprint of any vendor assessed in this pocket. Buyers span banks, alternative lenders, commercial mortgage securitisation issuers, insurance companies, government agencies and the government sponsored enterprises, servicers, aggregators, warehouse lenders, document custodians and investors.

Both halves of real estate finance are covered rather than one, commercial and residential, and the commercial side runs across debt, equity, balance sheet, securitised and multifamily agency business, with asset types including commercial, multifamily, affordable housing under the low income tax credit programme, senior housing and hospitality.

Coverage extends across the whole capital stack from the senior note to the equity position, and across the lifecycle from origination through secondary market execution to servicing and asset management. Operations run in the United States and Europe.