Campbell, California
There is a certain kind of company that spends years doing something unfashionable, and then one day the fashion arrives and everyone acts as though the company just showed up. HerculesAI is that kind of company. It was founded in 2017 as ZERO Systems, back when the frontier of legal technology was a mobile app that quietly watched lawyers answer email and recorded how long it took, so the hours could be billed. This is a deeply unglamorous problem. It is also, if you have ever seen a law firm's unbilled-time report, a very expensive one.
What makes the Hercules story worth telling is that the founders - Alex Babin, the CEO, and Gevorg Karapetyan, the CTO - did not treat that mobile app as the whole business. They treated it as a wedge. The real project was teaching machines to read the documents that keep regulated industries running: invoices, contracts, rate cards, insurance bordereaux, remittance advices, the paper sediment of commerce. And to do that, around 2020, they started training their own language models. Small ones, by today's standards - half a billion to two billion parameters - at a moment when almost nobody outside a handful of labs thought this was a reasonable use of a startup's runway.
Then large language models became the only thing anyone in technology wanted to talk about, and Hercules found itself holding a hand it had been dealt years earlier. In July 2024 the company raised a $26 million Series B led by Streamlined Ventures, with Thomson Reuters Ventures, Proof VC, Alumni Ventures, and a roster of industry angels along for the ride. It also did the thing companies do when they outgrow a name that starts with the word "Zero": it rebranded. The new name, Hercules, had been the internal code name for the AI engine all along, chosen because Hercules, per the company's own telling, "did all those great labors and unthinkable, unimaginable things that everyone else thought impossible." It is a slightly grandiose name for software that checks whether a partner billed at the right rate. It is also, arguably, exactly right, because checking whether a partner billed at the right rate, across millions of line items, forever, is a labor.
The verification layer for every document and transaction - catching errors before they reach your customer, counterparty, or bottom line.
- HerculesAI, on what it actually does
The contrarian architecture
Here is the technically interesting bet. Most companies selling enterprise AI want to sell you access to one enormous, general-purpose model and let its raw intelligence solve everything. Hercules does close to the opposite. It orchestrates ensembles of specialized, mid-sized models, each trained for a narrow job - extraction, transformation, decisioning, rule enforcement - and combines them into something the company likes to call an "AI worker." The pitch is that in regulated industries you do not primarily need genius. You need reliability, an audit trail, and an error rate low enough that a compliance officer will sign off. A team of narrow, predictable models turns out to be easier to trust than one brilliant, occasionally hallucinating one.
The other half of the pitch is where the software lives. Hercules brings large language models inside the security perimeter of the enterprise, on-premise if the customer demands it, so the sensitive documents never leave the building. For a bank, an insurer, or a law firm, this is not a nice-to-have feature bolted on at the end. It is frequently the entire reason a deal can happen at all. Trust, in these industries, is the product.
Selling the result, not the software
The clearest signal that Hercules believes its own accuracy claims is how it prices. Its billing-compliance product, Verify, launched in early 2025 with an ROI guarantee: if the customer does not reach a specified return on investment, they owe nothing. This is an unusual and slightly nervy way to sell software, because it only works if the product genuinely catches errors - Verify claims to detect up to half of billing-compliance issues before they turn into client deductions. Alongside it, the Rule Library converts a firm's Outside Counsel Guidelines and internal policies into machine-enforceable checks that run automatically against work in progress. It is compliance, translated from PDF into code.
The customers suggest the pitch is landing. Hercules says it serves roughly 30% of the top 100 US law firms and counts Fortune 500 names - Mercer, S&P, State Farm - among its clients, with revenue growing 4x in the year before the Series B. These are not organizations that adopt AI because it is exciting. They adopt it when the numbers reconcile and the auditor is satisfied.
The plot twist
In August 2025 the story took a turn that would end a lesser company. Aderant, a major player in legal practice-management software, signed a deal to acquire Hercules's legal tech assets - Apollo, Verify, and Athena, the whole legal vertical the company had spent nearly a decade building. You could read that as an exit. Hercules read it as a spin-out. The company kept its platform and its research team and pointed them at finance and insurance, where the same core problem - unstructured documents, transactions that need verifying, revenue that quietly leaks - is, if anything, larger. The new framing is "zero-defect order-to-cash," which is a mouthful, but it names a real and unloved corner of every enterprise: the place between doing the work and getting paid for it, where hours go unbilled, rate cards drift, and payments arrive without a remittance anyone can match.
That is the through-line across nine years and two brand names. Alex Babin and Gevorg Karapetyan did not pivot with each new trend. They kept building the same thing - machines that read and reconcile the documents people would rather not - and waited for the rest of the market to decide it was a good idea. It eventually did.
