He was studying how AI fails before most people believed it worked. Now he builds the guardrails that keep everyone else's models honest.
Most people fall for artificial intelligence by watching it do something clever. Adam Wenchel fell for it the other way around. He started at DARPA in the late 1990s, a University of Maryland undergrad let loose on research into agent-based systems, and the thing that stuck with him was not the magic. It was the breakage. Two decades in, he can tell you that the fastest way to understand an AI system is to become intimately familiar with every way it fails.
That instinct - stare at the failure, not the demo - is the whole reason Arthur exists. Wenchel co-founded the company in 2019 and runs it as CEO from New York. Arthur makes the software that watches other companies' AI: monitoring models in production, catching bias and drift, flagging hallucinations, measuring whether the machine is actually saying the right thing. It is the unglamorous layer under the glamour. And it turns out the unglamorous layer is exactly what banks and defense agencies are willing to pay for.
Wenchel grew up in the 1980s teaching himself to program on an Atari 800, pulled in first by games and then by the promise everyone kept repeating - that computers were about to remake the world. "Growing up in the 1980s, there was a lot of hype around how computers were going to revolutionize the world," he has said. The hype was mostly right, just early. He carried that pattern-recognition into everything after: the difference between a technology that is over-promised and one that is under-built.
He earned his B.S. in Computer Science from Maryland in 1999 and kept working at DARPA for about two more years on COABS - Control of Agent Based Systems - a program run by Jim Hendler that imagined software agents smart enough to compose themselves out of whatever resources they found. Read that description now, in the middle of the agentic-AI boom, and it sounds less like history and more like a preview.
In 2000, at 24, he was asked to lead software development for a social-networking startup called HeyMax Interactive and did not hesitate. That set the pace for the next fifteen years: design and engineering leadership across a run of companies - Govolution, Positive Development, Everfi, Endgame. He describes himself plainly. "I've been a focused entrepreneur for at least the last 15-plus years." The through-line is not a sector. It is the verb. He builds.
In 2014 he started his own: Anax Security, a Washington, DC startup using machine learning for large-scale defensive cybersecurity. It worked well enough that Capital One bought it in 2015 - and that acquisition turned into the job that would set up everything after.
After the acquisition, Wenchel got a rare chance: work directly with Capital One's CEO and CIO to stand up an AI division from scratch. He launched the Center for Machine Learning in 2016 and, as VP of AI & Data Innovation, scaled it to nearly 300 people. The team put models into the highest-leverage parts of the business - credit, fraud, cybersecurity - decisions that touched millions of customers.
Which is precisely where the discomfort set in. Deploying AI at that scale meant living with its failure modes in public. Success stories were everywhere, but, as he later wrote, "the same failure modes still impact models every day." Organizations were "stuck trying to cobble together piecemeal solutions." He kept arriving at one question and finding no good answer for it.
He decided the reason there was no good answer was that the tooling did not exist. It was, in his words, "an unaddressed part of the AI stack." So he left the comfortable job to go build the missing piece.
Arthur launched in 2019 - before ChatGPT, before "hallucination" was a boardroom word, back when GPT-2 was still the frontier. Wenchel assembled a founding team of people who had thought hard about the problem: responsible-AI advocate Liz O'Sullivan, machine-learning academic Dr. John Dickerson, and former Capital One colleague Priscilla Alexander. The pitch was simple and unfashionable: if you put AI into the world, you need a way to watch it.
The market caught up. Arthur's platform now covers large language models, computer vision, tabular data, and NLP, measuring things like response relevance, hallucination rates, bias, latency, and drift. The company has raised more than $63 million from investors including Index Ventures, Greycroft, and Work-Bench, with a $42M Series B in 2022. Its customer list reads like a stress test: three of the top five US banks, Humana, and the Department of Defense.
Here is the move that makes people tilt their heads. A venture-backed company took some of its most valuable tools and made them free. Arthur open-sourced Bench, a tool for comparing large language models, in 2023. Then in March 2025 it released the Arthur Engine, billed as the first open-source, real-time AI evaluation engine - no black box, no third-party dependencies, no data-privacy risk. Free to run.
The logic is trust. The hardest part of shipping AI is not building the model; it is the last mile - convincing an organization it can rely on the thing in production. You do not earn that with a closed box. You earn it by showing your work. Later in 2025, Arthur extended the idea to autonomous agents with a discovery-and-governance platform, built to find every agent running loose in an enterprise and keep it inside the rails.
Wenchel never fully left College Park. He chairs the University of Maryland's Computer Science Advisory Board, a role he took on in 2017, and in 2025 returned to give a commencement keynote. He endowed a $3M gift, negotiated a Capital One tech-incubator partnership for the school, and established the Rosemary Wenchel Memorial Scholarship in honor of his mother. He credits the place with the thing he values most in engineers: "one of the things that Maryland does very well is teach computer science at a really fundamental level."
His advice to the graduating class was less about AI than about how to move through a career where the ground keeps shifting. Have a plan, he told them, but leave room for luck and go after the problems that are actually fun. "Whatever the new trend is," he said, "you have the tools to adapt and understand and thrive." Coming from someone who went from an Atari 800 to agentic AI without ever changing his core job description - build things worth being proud of - it lands as earned, not tidy.
The tell about Wenchel is that he never got into AI for the reasons most people do. Not the demos, not the hype cycle he grew up inside. He got in through the failures - and then he built a company whose entire job is to catch them before they reach you. That is a quieter story than most AI founder arcs. It is also, if you run your bank on a model, the one you most want to be true.