In 2017, the future of finance had a peculiar loading screen. Customers could open an account from a sofa, upload a bank statement and admire a spotless mobile interface. Then the application disappeared backstage, where a human might spend half an hour checking names, dates, balances and whether somebody had nudged a number in Photoshop. Sometimes the answer came in days. Sometimes weeks. Inscribe's founders did not begin with a grand theory of agentic AI. They began by asking what on earth was taking so long.
Twin brothers Ronan and Conor Burke had encountered the problem from opposite sides. Conor had worked around document-heavy onboarding at an Irish bank; Ronan had felt the customer's delay. Contemporary reporting also names Oisín Moran and James Eggers in the original founding group. The team moved to San Francisco, joined Y Combinator's summer 2018 batch and found an early customer in Bluevine. The wedge was narrow: help financial companies verify documents without building a data-science department of their own.
That remains the cleanest description of Inscribe. Banks, credit unions, lenders and fintechs send it pay stubs, tax forms, IDs, business filings and bank statements. The software looks for tampering, fabrication, recycled templates and AI-generated fakes. It also extracts transactions, categorizes cash flow and checks whether facts agree across the application. The result is not merely a red light. Reviewers see ranked evidence and an explanation they can inspect.
The first product was an x-ray for PDFs
A forged statement is rarely considerate enough to announce itself. One applicant changes a salary. Another buys a template. A third generates an entire document whose logo, typography and polite little rounding errors look plausible. Inscribe attacks the file from several directions: pixel-level traces, metadata, layout, language, transaction behavior, known template networks and contradictions between documents or outside records. The point is not that one detector is brilliant. The point is that each detector gets a vote.
One application, four useful questions
This layered approach is the company's real distinction from a generic OCR service. OCR can tell a lender that the number on line seven is $84,000. It cannot, by itself, tell the lender whether the number belonged there yesterday. Nor is Inscribe exactly an identity vendor or a decision-orchestration platform, although it overlaps both. Its home turf is the document evidence inside underwriting, account opening, KYC and KYB workflows.
“Start with data, not machine learning.”Ronan Burke on the cold-start problem
The line is more than founder folklore. Fraud models need examples of real files, real manipulations and emerging attacks. Inscribe says its systems learn from millions of production applications, known fraud templates and cases hand-labeled by in-house risk analysts. A single lender sees only its own queue. A specialist vendor can spot an account number, template or editing pattern reappearing across institutions. That network can become a moat, provided the data is governed carefully and customers keep sending enough varied examples.
What it did, what it cost, what broke
The company funded that data-and-distribution loop with roughly $3 million in seed capital, a $10.5 million Series A in 2021 and a $25 million Series B led by Threshold Ventures in January 2023. The last round brought its reported total to about $38 million. Inscribe planned to double a workforce that was then around 50 people. Public pricing for the product, however, remains undisclosed. It is enterprise SaaS sold through demos and order forms, available in a web application, a REST API and partner integrations. The sensible buying unit is not “one AI seat.” It is the cost of reviews, fraud losses and abandoned good applicants.
Then the plan met the interest-rate cycle. Fintech customers slowed, Inscribe missed revenue targets for more than a year, and in January 2024 the company cut just under 40 percent of its staff, mainly in go-to-market and operations. This is the part many company profiles airbrush away. The first thing to fail was not the detector. It was the growth assumption attached to the Series B.
What changed management's mind was a two-sided jolt. Higher rates forced customers to do more with smaller operating teams. At the same time, large language models became capable of reasoning across text and images, while generative tools made forgery cheaper. In the fourth quarter of 2023, Inscribe set a new product strategy. By September 2024 it was selling AI Risk Agents - software meant to work through a review, call specialized tools and prepare a case, rather than merely return isolated fields and alerts.
The agent is an investigator, not a tiny loan officer
Inscribe's 2026 architecture is refreshingly pragmatic. Routine parsing and classification can go to a smaller, cheaper model. Transaction enrichment can run on an open model when its performance is comparable. A stronger model coordinates cross-document reasoning and writes the audit-ready report. Proprietary models run in parallel for image forensics, metadata and network patterns. Amazon Web Services described the system as completing a full investigation in under 90 seconds, around 20 times faster than a traditional manual review.
That model routing matters. Asking the most expensive model to read every comma is wasteful; asking the cheapest model to reason through a coordinated fraud ring is reckless. Inscribe says it cut inference cost on routine tasks by roughly 40 percent by matching the model to the job. The company's newer Assistant lets an analyst ask why a document was flagged, compare two submissions or inspect a strange transaction without hunting through screens.
Customers provide the practical proof, with the usual caveat that case studies are selected success stories. Plaid reported getting document results in roughly 30 seconds instead of one or two days and automating half of reviews. Ramp said it saved 30 minutes per application and $300,000 in fraud. Logix Federal Credit Union reported more than $3 million in potential losses prevented within eight months. BHG Financial said review became more than 90 percent faster. Those figures explain the sales pitch better than a hundred slides about agents.
“Inscribe comes back with a response on each document in under 30 seconds.”Katie Randolph, Plaid
The customer list also clarifies the market. Inscribe is not for a consumer trying to check one suspicious PDF. It sits inside institutional workflows at Plaid, Ramp, Bluevine, BCU, Coast, Camino Financial, Navan, Rapid Finance and others. It can connect by API or through partners including Taktile and Alloy. The business model works when the queue is large, the review is repetitive, a bad approval is expensive and the buyer can measure the before-and-after result.
The playbook worth stealing
Build from the queue outward
- Shadow the ugly manual workflow before naming the AI product.
- Win one narrow job where speed and accuracy are measurable.
- Use pilots to earn domain data and learn the buyer's real scorecard.
- Combine specialist tools; do not force one model to perform every task.
- Turn every output into evidence a human can challenge and audit.
Plaid's selection process is especially copyable. Its team compared vendors on latency, accuracy, document variety, pricing flexibility and lending experience, then ran a pilot against both another tool and manual review. A startup can borrow that exact shape: ask the customer to define five buying criteria, establish a control case, and measure the pilot on their documents. The demo becomes an experiment instead of theater.
Another lesson is less comfortable. Product progress does not immunize a company against market timing. Inscribe raised during fintech expansion, planned a larger organization, then cut back when lending slowed. Its recovery bet was not “add AI” as decoration. It changed the unit of work from a document to an investigation. That is a meaningful product shift, but the layoffs remain part of its cost.
Where the magic stops
No detector proves a negative. Novel fraud can evade learned patterns; legitimate documents can look odd; low-quality scans can erase clues. A risk team that automates denials from an unexplained score merely converts a slow problem into a fast liability. Inscribe's own product language emphasizes explanations and human escalation, and its terms put boundaries around public web-search functionality in credit, employment, tenant and other regulated screening contexts.
Applicants provide no documents, or direct bank and payroll connections already deliver authoritative data.
An automated flag can deny credit, housing or work without a trained person reviewing the evidence.
Volume is low, fraud losses are small, or integration work could cost more than the review queue.
The input is damaged, the attack is genuinely new, or outside records are incomplete or wrong.
The strongest version of Inscribe is therefore modest: a fast, tireless first-pass investigator that makes a human analyst's case file better. It does not abolish judgment. It reallocates attention. The company began by noticing people waiting behind an “instant” application and learned, painfully, that neither growth nor automation is instant either. Its useful idea is simpler. Find the queue. Learn its exceptions. Build the evidence. Then choose the model.