SecureLend

Language models write the memo, and the memo is what SecureLend sells. Agents classify the file, extract and spread the numbers, screen the parties, test the deal against the institution's rubric and draft the memo, then route exceptions such as customer concentration, covenant headroom or a filing mismatch to the underwriter.

The work is organized like a staffed desk: Junior Analyst agents read and structure the file, Analyst agents run rubric prechecks, quantitative analysis and fraud flags, and Senior Analyst agents draft the credit, underwriting and investment committee memos. A Desk plan bundles five Senior Analysts with a policy engine, decision packs and audit exports.

Approval is reserved for a person. The agents can draft a cited decline or refer a file, but SecureLend will not approve credit automatically, and policy exceptions go to the underwriter with the evidence behind them. A waiver needs a named person and a reason. On the borrower side, the loan comparison service lets an assistant submit an application to one or several lenders once the borrower gives explicit consent. Institutions set their own rubric, escalation rules and approval gates, and every material action is logged with its evidence and review state.

Every material figure in a memo links to the page or row it came from, so an underwriter can check the work instead of redoing it, and a check that fails is left visible rather than filled in. SecureLend publishes its benchmark protocol and results by task, and describes testing extraction against thousands of real financial documents and running the agent on a lender's closed deals to see whether it reaches the same conclusions, flags and ratios as the credit team.

No model card or validation report is offered. Prompts are version pinned, each memo records the model behind it, and a decline cites the file, such as debt service coverage below policy or occupancy the rent roll does not support.

Results come from SecureLend's own pilots and from public benchmarks it ran itself, and no customer is named. In second quarter 2026 pilots it reports underwriting cycles cut from weeks to hours, document accuracy of up to 99.2 percent in controlled tests on structured documents, and estimated processing costs down by roughly 75 percent.

On public benchmarks it reports each track on its own instead of averaging them: 90.3 on commercial credit files in UWBench, every outcome correct on 24 origination tasks in a second benchmark (20 of 24 on the full rubric), and 95.5 on MortarBench bank statement questions. Each result is labeled as SecureLend's own pilot measurement rather than an official leaderboard entry, and the UWBench protocol is open, so a lender can run its own files through it.

Client files can feed SecureLend's own models in aggregated, de identified form, and a client can opt out in writing and keep the service, though some advanced features may be limited. Personal information never trains models shared across clients, and aggregated data is not sold or licensed to third parties without the client's written consent. In its pitch for the Tag product, SecureLend goes further and says nothing trains shared models, a promise the terms do not make.

Retention terms for the managed OpenAI, Anthropic, Gemini and DeepSeek routes are not given. Customer keys for outside data and model providers are encrypted at rest, prompts are version pinned, and each memo records the model lineage behind it.

In its terms, SecureLend commits to a written information security program that complies with the Gramm Leach Bliley Act and is updated every year. For consumer data it handles for lenders, it acts as a service provider under the California Consumer Privacy Act and a processor under the General Data Protection Regulation (GDPR).

Client data is stored and processed mainly in United States data centers, with standard contractual clauses for any transfer abroad, and is kept for the subscription plus 90 days unless the client asks for earlier deletion. The subprocessor list sits behind the customer login, with 30 days notice of any new subprocessor and a right to object.

Loan comparison requests made inside ChatGPT or Claude pass through the assistant platform first, comparison results are not stored, and personal information moves only when a borrower submits an application.

SecureLend markets a SOC 2 Type II examination, and its terms promise to maintain one and share current reports with clients under confidentiality. Its own compliance program shows both a SOC 2 Type I and a Type II audit as still in preparation with an independent auditor, alongside 83 continuously monitored controls covering encryption, access reviews, logging, backups and vendor risk.

Data is encrypted with Transport Layer Security (TLS) 1.3 in transit and 256 bit Advanced Encryption Standard (AES) at rest, staff access uses multifactor authentication and least privilege, and customer keys for outside services are encrypted before storage. Breaches affecting client data are reported within 72 hours.

By its own terms, SecureLend is not a lender, broker, loan arranger or credit decision maker, and it claims no license or examination of its own. The terms place fair lending under the Equal Credit Opportunity Act and the Fair Housing Act, adverse action notices, and model validation under SR 11-7 and Bulletin 2011-12 of the Office of the Comptroller of the Currency on the lender.

Referral fees from lenders arise only on business and commercial financing; SecureLend takes no compensation on consumer, personal or residential mortgage loans, where its comparison tools are informational.

Fair lending work sits with the lender in SecureLend's terms, which require the lender to monitor fair lending, run disparate impact analysis and send adverse action notices, and which warn that the models may carry bias from their training data. SecureLend publishes no disparate impact testing of its own memo, decline or loan matching models.

For loan comparison, it says offers are ordered on stated criteria such as price, fees, terms and eligibility, that referral fees do not shape how offers appear, and that the tools are not built to steer borrowers toward harmful products, with no testing shown behind any of it. The terms also commit it to a model risk management framework covering development, validation and monitoring.

What a lender gets is traceability: every figure in a memo cites its source and every material action is logged with its evidence and review state, so a reviewer can challenge a finding line by line. The terms disclaim the accuracy of model output and put its use at the client's own risk.

SecureLend's indemnity covers intellectual property claims, breaches of its data security duties and gross negligence, not an error in a memo, and liability is capped at the fees paid in the prior twelve months. For a borrower who applies through the comparison service, any loan is strictly between the borrower and the lender, and disputes under the terms go to binding arbitration with a class action waiver.

Model choice is part of the product. SecureLend runs managed routes to OpenAI, Anthropic, Gemini and DeepSeek, and a tenant can connect its own OpenRouter account to reach Kimi, MiniMax, Qwen, xAI and Mistral models or inference hosts such as Together AI and Fireworks. Which model each step uses by default is not stated. Policy can set the model by workload, quality tier and cost, the evidence workflow stays the same when the model changes, and SecureLend absorbs the model cost inside its per call prices.

SecureLend sits beside the loan system rather than inside it. A lender pushes files in by spreadsheet export or by pasting nCino or Salesforce record IDs, and SecureLend does not scrape the core. No live connector to a core banking system is named. Agents are reachable through the Model Context Protocol and an OpenAPI interface, with a TypeScript software development kit (SDK) for developers, and the team can tag the agent from Slack.

Venture teams can connect Harmonic company data, Affinity customer relationship management (CRM) records, Visible portfolio metrics and DocSend decks on their own keys, and sanctions screening runs on the lender's own ComplyAdvantage, Refinitiv or similar account. A finished workspace can be promoted into a case on SecureLend's own loan system.

SecureLend runs on Amazon Web Services: the loan comparison server on Fargate containers in an eastern United States region with a 99.9 percent uptime commitment, and the agents on AWS Lambda and S3 storage, with sign in through Amazon Cognito and keys held in AWS Secrets Manager. Client data is stored and processed mainly in United States data centers. SecureLend describes isolated compute, tenant isolation and customer managed encryption keys as part of the platform.

No private cloud or on premises edition is offered. A team that wants model calls on its own account can bring an OpenRouter key, and running models on its own hardware through Ollama is on the roadmap rather than available today.

A lender can start without a sales call: the Senior Analyst trial runs seven days with no card, after which plans renew monthly or yearly until canceled. The plan prices themselves are not published. The loan ledger charges a book fee in basis points only on loans SecureLend holds, and nothing when the loan sits in nCino or arrives as a spreadsheet export, though the rate is not given.

The developer agents carry a full price list: $0.50 per pitch deck precheck, $4.99 per memo, $0.06 a page for document reading, $1.50 per quantitative analysis and $0.60 per sanctions and know your business check, with a monthly free allowance for people and a one time trial for machine accounts. Developers pay by card or in stablecoin on the Base network.

SecureLend pitches one harness to community banks, commercial and specialty lenders, private credit and factoring desks, insurers, and venture, private equity and investment banking teams, across business loans, commercial real estate, mortgages and public equity research. The community bank offer is specific: a seven day trial in place of a long implementation, document reading included and nCino left in place. No customer in any of these segments is named. The loan comparison service adds consumers and small businesses comparing personal, auto, student, mortgage and business loans, and the consumer side is offered for information only.