A refrigerated truck can finish its journey while its paperwork is still looking for a destination. At Hirschbach, the freight documents feeding billing once helped stretch the average time to bill to nine days. The cargo had moved. The information needed to charge for moving it had not. There is something exquisitely modern about an industry that can track a vehicle yet still wait for someone to read its receipt.
- Hyperscience turns complicated documents into usable business data.
- Its customers include insurers, banks, healthcare providers and public agencies.
- The useful measure is a finished workflow, with uncertain results reviewed.
01 The receipt was the traffic jam
Hyperscience’s published Hirschbach case study reports that billing time fell to three days after document automation and workflow changes. Bills of lading, delivery proofs and receipts entered a pipeline that classified them, extracted fields and passed structured results into downstream systems. Staff handled exceptions. The intervention reached beyond a clever reader: it connected reading to the work that followed.
Company-published customer result. Billing cycle time, not model processing time.
One detail makes the account more persuasive than a pristine demonstration. Signature detection initially sat in the mid-60% accuracy range. Combining specialized models with Hyperscience’s ORCA vision-language model pushed it above 77%. That is an improvement with plenty of uncertainty left. A missing signature is precisely the sort of small nuisance that can stop an otherwise impressive system.
02 The robot needed a reader
Hyperscience occupies the space between a document arriving and another system being able to act on it. Its Hypercell platform classifies documents, extracts information and coordinates validation and subsequent steps. An insurance claim, handwritten form or invoice becomes fields that software can use. Reading is only useful if the account number ends up in the account-number field.
CI Financial illustrates the distinction. Its existing robotic process automation handled rules-based tasks but struggled with unstructured document data. Hyperscience supplied the reading layer alongside that automation. The company’s account reports a 60% reduction in the time agents spent typing case documents. Replacing the whole stack would have answered a different question from the one the customer actually had.
The governing choice is how much uncertainty to tolerate. Hyperscience’s documentation distinguishes accuracy from automation, also called skip rate: the proportion of work completed without human supervision. Raise the required accuracy and more uncertain results may need a person. Buyers should ask to see both numbers together. Otherwise, a beautiful accuracy statistic can conceal a very busy review desk.
03 A tax form is somebody’s Tuesday
The customer list explains the emphasis on review. Hyperscience names Charles Schwab, MetLife, Stryker, the Social Security Administration and the Department of Veterans Affairs among organizations using its technology. These are places where a document can influence money, treatment or access to a service. Their paperwork has consequences beyond the cost of typing it.
At Britain’s HMRC, the company describes an on-premises implementation trained on tax forms, connected to scanning and logic-processing systems. Low-confidence fields went to human reviewers. Its case study reports turnaround falling from up to 45 days to one day in many cases, with more than 300 staff redeployed. A taxpayer waiting for a refund experiences the queue as a delay in ordinary life. The platform’s mission to improve front-office experiences through back-office work becomes rather literal here.
04 Money arrived before the reset
Founded in 2014 by Peter Brodsky, Krasimir Marinov and Vladimir Tzankov, Hyperscience built around machine learning before the current generative-AI rush. Investors financed expansion: a $60 million Series C and an $80 million Series D in 2020, followed by a $100 million Series E announced in December 2021. The last announcement described nearly 400 employees.
In March 2022, Brodsky stepped down. The company announced a reorganization and named operating chief Charlie Newark-French interim CEO. Andrew Joiner, previously CEO of InMoment, took over in April 2023. Those events establish a reset; they do not establish its private causes. They also complicate the pleasing assumption that a large funding round proves a company has settled the difficult questions.
05 Give the expensive model a smaller job
By 2026, the product argument had expanded. Spring’s release emphasized routing different tasks to different models. Straightforward, high-volume extraction could use specialized models; harder documents could call on ORCA or frontier models. The appeal is economic as well as technical: paying for deeper reasoning on every ordinary field can become an expensive habit.
The September 29 Fall release introduced ORCA 2, zero-shot table extraction, address verification, redaction and in-workflow translation across 26 languages. Hypercell for GenAI prepares document information for downstream AI applications. Hyperscience sells enterprise software with managed SaaS and customer-controlled deployment options; its partners include AWS, Google Cloud, Microsoft, IBM and Palantir.
“Enterprise AI doesn’t have a model problem. It has a data problem.”Andrew Joiner / September 2026
That is Joiner’s diagnosis. The competitive field includes ABBYY, UiPath, Tungsten Automation and Rossum, as well as cloud document services. Hyperscience’s pitch centers on combining document expertise, model routing and review inside an operational platform. An extraction API can be a useful component. Someone must still build the surrounding workflow.
06 Count the queue, then count the cost
An IDC study discussed in a Hyperscience webinar interviewed six customers and calculated 615% three-year ROI. Its discounted three-year investment averaged approximately $2.8 million. That is a total investment measure, not a software price, and the interviewed customers’ return is no promise for a new buyer.
The method worth copying is to establish a baseline: elapsed time, manual touches, errors and cost per completed case. Test actual awkward documents, connect the output to a working process and budget for review and change management. For a small workload or an already structured input, a full enterprise platform may be excessive. When the next system cannot use extracted data, faster reading merely moves the queue. Follow the document until somebody can finally do something with it.