TidalWave

The platform is described as the first generative application built for mortgage seekers and lenders, and the model is trained on structured mortgage data including standardised loan application files and bank statement transaction records rather than adapted from a general purpose system. Agents automate a reported seventy percent of daily origination tasks, read bank statements for risk indicators and eligible assets, and detect document errors without human review. Apply the removal test and what remains is a point of sale form with connectors, which is the incumbent product category this positions against.

The automation claims are extensive and the controls are not described. Agents automate up to seventy percent of daily tasks including document collection, compliance checks and income verification, and document error detection is explicitly stated to occur without human review, while the platform delivers immediate pre approvals through direct connections to both automated underwriting systems.

Nothing public sets out which determinations require a licensed loan officer's sign off, what threshold routes a file to a human, or how an agent's compliance check is itself checked. In a process where a licensed originator carries personal accountability, that gap is the notable one.

The first top grade on this axis earned through published independent evaluation. TidalWave commissioned a benchmark with a university research laboratory measuring its model against a leading general purpose system on 90 questions across 10 borrower scenarios, built from complete application files and bank statement data, with a mortgage subject matter expert designing questions from real usage patterns and edge cases deliberately included.

The results were released publicly with methodology, an academic named on the record, and the company's own weaker score on one category disclosed alongside the explanation for it. Publishing a category where you lose is what separates evaluation from marketing. Model documentation and a validation package are still absent.

Named deployments span the market's structural tiers: the country's largest mortgage brokerage rolling the platform across more than 3,200 loan officers in 48 states and Puerto Rico, described as the first enterprise scale agentic deployment in the sector, alongside two named mortgage banks.

A 22 million dollar round was led by an investment firm with participation from the largest homebuilder in the country, which is a strategic rather than purely financial signal, and the founder previously built and sold an advertising technology business and served as chief technology officer at a listed digital mortgage lender. Automation is quantified at up to seventy percent of daily tasks. Market share ambition is stated as a target rather than an achievement.

The benchmark work is the clearest evidence of stewardship, since it was constructed entirely on synthetic borrower data built from synthetic account data specifically to protect privacy, and the identifier stripping practice is documented rather than asserted. Domain training on structured mortgage data narrows the model's scope in a way that reduces the failure surface. Two things hold this back.

Hallucination free is used repeatedly as a marketing claim, and it is an absolute that no system supports. And nothing states whether lender or borrower data informs model improvement across customers.

One disclosed engineering decision is stronger than most published privacy policies in this index: the company strips personally identifiable information from its artificial intelligence interactions, and it disclosed this in the specific context of explaining why its own benchmark score on account verification was lower than it might otherwise have been. Accepting a worse public result rather than sending borrower identifiers to a model is a costly signal that the control is real. Around that, the platform handles bank statements, payroll data and complete application files. No published privacy framework, retention schedule or subprocessor list was located.

No trust centre, enumerated certification list, attestation scope or audit period was located in this pass. A brokerage deploying across 3,200 loan officers and integrations with both government sponsored enterprises' systems would have required security assessment, and the enterprise partners impose their own connectivity requirements, so assurance exists privately. The grade records what an outside buyer can verify.

TidalWave supplies technology while its customers hold the licences, and its regulatory grounding runs through direct integration with the government sponsored enterprises' automated underwriting systems, which are the gatekeepers determining whether a loan is saleable and which impose their own approval requirements on connected systems.

Compliance checks and disclosure handling sit inside the origination workflow, and the founder identifies a specific supervisory concern in publishing that off the shelf models answer compliance questions poorly. No formal admission programme with published criteria is evidenced.

Fairness is the company's own stated motivation, with published material arguing that human bias has always been present in lending and that approvals have never been transparent to the consumer, positioning the product as a route to equity and transparency. That is a fair lending claim, and mortgage lending sits under equal credit opportunity rules and disparate impact analysis.

No fair lending testing, demographic outcome analysis, adverse action documentation or independent audit was located to support it. The benchmark measured accuracy, not equity, and the same reasoning applied elsewhere in this index holds: a stated fairness ambition without evidence invites reliance it cannot carry.

Publishing an independent benchmark with methodology creates an accountability of a different kind, since a buyer can hold the vendor to a measured result rather than a claim, and that is more than most vendors here offer. It is not a commitment. No accuracy guarantee, no remediation term and no published obligation where an agent's compliance check misses a defect or an automated verification produces a wrong pre approval.

For the borrower, the transparency the company says it wants to create is not yet described as a mechanism: nothing sets out what an applicant is told about automated processing or how they contest it.

More of the chain is visible here than for most vendors, largely because the benchmark work required naming what the model was compared against and how it differs. The company states its model is trained on structured mortgage data rather than general purpose text, names the university laboratory that co conducted the evaluation, identifies three verification data providers, and discloses that identifiers are stripped before model interaction. What is not published is whose infrastructure runs the models, whether any external provider is called in production, and no subprocessor list exists.

The integrations are the ones that decide whether a mortgage product is usable. Direct connections to both government sponsored enterprises' automated underwriting systems allow instant risk assessment at the point of application, integration with the dominant loan origination system runs through its published partner interface, and income, employment and asset verification connect through three named data providers.

A partnership with an independent implementation provider delivers the platform into the brokerage channel. That combination covers the underwriting decision, the system of record and the verification layer.

Delivery is cloud hosted software serving domestic lenders and brokerages, so cross border complexity does not arise. Residency and retention still matter because mortgage files, bank statements and payroll records carry multi year record keeping obligations, and the identifier stripping practice suggests deliberate thought about what leaves the environment without describing where anything sits. No hosting regions, tenancy model, residency options or subprocessor chain were located.

No rates, tiers, billing unit or minimum were located. The unit question is straightforward in this market, where point of sale systems are conventionally priced per loan officer seat or per funded loan, and the two produce very different economics for a brokerage with 3,200 originators. Nothing public indicates which applies.

Coverage runs across the mortgage origination market's different shapes, reaching independent mortgage banks, brokerages and their loan officers, with a partner delivering the platform specifically to the brokerage channel, and both the borrower and the originator are treated as users of the same system.

The boundary is singular and deliberate: this is residential mortgage origination in the United States, with nothing addressing servicing, other consumer lending, commercial real estate finance or any non mortgage product.