Infrrd

The models are the entire product. Deep learning, computer vision and language processing with a proprietary platform and machine learning character recognition, sold specifically on template free extraction, which is the capability that cannot exist without inference. This is the extraction led shape where the residue test resolves cleanly in the vendor's favour: strip the models out and the platform has nothing to operate on, unlike a workflow platform that keeps working.

Membership settled on the horizontal test, and comfortably: financial services is not an industry page here but a set of separately named, separately purchasable products with domain rules encoded, alongside a general purpose business elsewhere.

Two halves pulling in opposite directions and both belong on the record. The strong half is a control that cites the rule it enforces, which is the bar this index sets for a regulated workflow: tolerance tests under the federal disclosure rule and comparisons between the closing disclosure and the loan estimate are run at the point of upload, with tamper resistant logs and version control kept for regulators and investors. The weak half is the disposal path.

Exceptions are routed to a person and, in the vendor's own words, everything else flows straight to delivery, with no confidence threshold, review rate or escalation criterion published. For an extraction product feeding an underwriting file, the question is not whether exceptions get seen but what decides that a field is not an exception, and that is exactly what is missing.

One genuinely uncommon mechanism: a public self serve demo that invites a prospect to upload a difficult document and see what the system reads, which lets a buyer generate their own accuracy evidence rather than take a figure on trust. Almost nothing in this index offers that. Alongside it, validation against business rules, audit trails, version control and tamper resistant logs.

Held off A on the standing extraction accuracy question, and this vendor produces a sharp new claim shape while failing it. Its content marketing establishes the human baseline, citing manual processing error rates of one to two percent, and its product page then claims better than human accuracy without ever publishing its own rate.

Naming the comparison point and withholding your own side of it is a more sophisticated version of the unquantified accuracy claim, because it borrows the credibility of a number the vendor never has to stand behind.

One named customer with a described scope and a quote, an insurance programme business whose intake spans forms from 2,100 separate carriers, plus a quoted partner on the loan quality control launch. Two independent analyst evaluations name it a Leader, from a field of eighteen vendors in one case.

Held at B rather than A because analyst placement evidences market recognition rather than measured performance, and the quantified claims that do exist (approvals sixty percent faster, loan review twice as fast) are attached to no institution. Banked check: a customer stories library exists on the site and was not opened in this pass, and that library has moved evidence grades four times in this sweep.

Nothing published on whether customer documents train models. The question has real force here because the platform explicitly allows an organisation to create and train new models, and the corpus is loan files and claim files containing borrower financial detail and medical reports, yet no statement isolates one customer's training data from another's or describes retention after extraction. Redaction is offered as a customer controlled feature, which is a control over output rather than a commitment about the vendor's own use.

More specific than the usual gesture, and the discriminator is worth stating so this is not relitigated: the vendor names the principles it operates under, ENUMERATES the individual rights it supports (access, erasure, portability, objection), names standard contractual clauses as its transfer mechanism, and ships redaction as a control. That is materially more than built with privacy in mind, which this index credits nowhere.

Held off A on a specific and important gap: there is no position at all on the United States financial privacy regime, even though the mortgage product extracts borrower income, assets and liabilities, which is the textbook definition of nonpublic personal information.

A service organisation control report of the second type, with the three trust service criteria it covers named explicitly rather than left implied, supported by continuous control monitoring and annual audits, and a standing security page describing encryption in transit and at rest, fine grained access management, isolation and redaction. That is a genuine posture rather than a badge row.

Held off A: the report is available on request rather than through a trust portal, no bridge letter covers the gap between audit periods, no international information security certification is held, and no penetration testing is disclosed.

Unlicensed and unsupervised, the norm for a document processing vendor. It operates close to named regulatory machinery, encoding federal mortgage disclosure tolerance tests and building files to be examinable by regulators and investors, but the party answerable to the supervisor remains the lender.

No fairness position, no governance disclosure. The exposure is not abstract: extracted income, asset and liability values feed underwriting decisions on individual borrowers, and a separate capability flags fraud patterns and discrepancies on loan files. An automated fraud pattern flag is a consequential inference about a named person, and nothing describes how it is calibrated or who reviews it.

No recourse route and no allocation of responsibility. The specific chain matters here: a misread income figure passes through automated validation, is not classified as an exception, and flows to underwriting and then to investor delivery. The audit trail is thorough and records what was extracted rather than whether it was correct, so the very completeness of the log can make a wrong value look verified.

Proprietary and patented, with the model classes named at the broadest level and no model, provider, architecture or version identified anywhere. The patent position is used as the differentiator in place of disclosure, which is a variant worth noting: a patent proves originality to a lawyer and tells a model risk function nothing about what is running.

Integration is described as the real work rather than glossed, and unusually the vendor says so itself, writing that mapping extracted fields into a loan origination system, a core banking platform or a compliance database is consistently underestimated and frequently becomes the primary implementation cost. Exception queues, investor delivery and a partnership with a mortgage compliance specialist are all named. Held off A because no specific system, connector or protocol is named, so the depth is asserted by category rather than demonstrated by counterparty.

Residency is addressed rather than ignored, which is what separates this from the majority of the index: regional hosting options are offered, standard contractual clauses are named for transfers, and the security material describes tenant isolation, hardened cloud infrastructure and customer controlled policies. Held off A because no region is named, no single tenant or customer hosted option is described, and the hosting options are offered rather than specified.

No price, tier or unit published. One unusual and creditable commitment sits next to the absence: the service level is presented as a customer choice on a stated range from fifteen minutes to twenty four hours, which is a real operational term published openly. A buyer can learn the turnaround they will get and nothing about what it costs.

Within financial services: mortgage lenders, servicers, investors, quality control and compliance teams, and insurance carriers, across origination, quality control, post close, servicing, claims and policy administration. Volume evidence is real (more than a thousand document types, twenty two languages, more than a hundred billion pages).

Held off A because the institution types are narrower than the page count suggests, with no banking, capital markets or wealth buyer addressed, and a material share of the business sits outside financial services entirely in invoice processing, construction and manufacturing.