Carl Bate Wants AI to Show Its Work

The origin story fits inside a small black box. Carl Bate was about ten when his parents brought home a Sinclair ZX81 and its 16K RAM pack. The machine had less memory than a blank modern document and none of the forgiving surfaces that later made computing feel effortless. It invited a child to type instructions, see what happened, and try again. For Bate, the important effect was not access to a device. It was the discovery that an idea could be made executable.

Nine years later, in 1989, he reached a fork that now reads like a tidy piece of narrative engineering. The University of Edinburgh offered him a place to study artificial intelligence. A software engineering job was available too. He chose the job. His first professional work involved PL/1, COBOL and CICS at Guardian Royal Exchange, the insurer that became part of AXA. Long before a chatbot could produce a paragraph on command, Bate was learning what production software always teaches: rules have edge cases, systems inherit history, and somebody remains accountable when the output reaches the real world.

The translator in the machine

Bate advanced from developer to director of software engineering, then moved through a sequence of jobs that placed him at the boundary between technology and the people expected to use it. He became chief technology officer at Javelin Group in 2000. Three years later he took the CTO role at Capgemini UK, sitting on an executive board responsible for roughly 2,000 staff. His description of the job was notably free of gadget worship: extract value from technology trends, minimize risk, and resolve the tension between the labels “business” and “IT.”

That sentence contains much of his later work in miniature. Software is never merely software once it enters an organization. It changes incentives, redistributes judgment and creates new handoffs. Bate was already arguing for a people-centered view of information systems when enterprise technology discussion tended to treat users as the last box in a diagram. He also resisted executive-tech theater. Asked in one interview how often he checked his BlackBerry, he replied that he did not own one.

The more revealing answer came when he was asked about useful training. Bate did not name a certification. He pointed to colleagues and to skills developed by working with people rather than machines. The engineer had become a translator. His later career at Atos Consulting and Arthur D. Little formalized that role: helping organizations frame difficult problems, break stale patterns and connect technical possibility to human consequence.

“The humble ZX81 gets my vote.”Carl Bate, on the technology that most shaped his working life

From best practice to the next question

At Atos, where he led CIO advisory work and later served as managing partner, Bate wrote about “next practice.” The phrase challenged a comfortable corporate instinct: when faced with uncertainty, borrow the method that worked somewhere else. Best practice is useful when the problem is familiar. It becomes a trap when the category itself is changing. Bate’s alternative started with a more demanding task - finding the question that the inherited playbook had failed to ask.

At Arthur D. Little, he co-led digital problem solving across London and New York, worked on artificial intelligence and led digital health globally. His published methods included Pattern Breaking Pattern Matching, an approach to inventive problem solving, and VPEC-T, a framework for examining values, policies, events, content and trust in complex systems. The acronyms are consulting-era artifacts. The durable idea underneath them is plain: look at the full system, including assumptions and behavior, before reaching for a technical fix.

This is useful context for understanding why Bate’s current company does not begin with a chat interface. Evidium begins lower in the stack, with the representation of knowledge itself. The company’s premise is that a language model can discuss a body of knowledge fluently while still lacking a dependable model of its dynamics. A polished answer may conceal where a claim came from, how concepts connect, which transition is being assumed and where an expert might disagree.

Making knowledge computable

Bate founded Evidium in San Francisco in 2021. The company describes its work as building world models for healthcare: computational maps of states, the transitions between them and the factors that cause those transitions. Its systems are designed to join structured knowledge with real-world observations, then expose a reasoning path that a qualified user can review. The ambition is not to remove expert judgment. It is to give that judgment a clearer object to inspect.

The patent record makes the abstraction concrete. Bate and a group of colleagues are named on inventions concerning computational evidence, the determination of a patient state, and the evaluation of possible next actions. The described systems extract concepts, connect them through ontologies, match structured factors to observations and keep links back to the underlying evidence. Several patents in that family were granted between 2023 and 2026. Their shared design instinct is traceability: an answer should carry a route back through the machinery that produced it.

“We’ve focused on knowledge integrity from day one.”Carl Bate, after Evidium’s Series A

A company built around inspectability

Evidium emerged publicly from stealth in 2023 with the phrase “Referenced AI.” That same year it announced work with Syntropy, a data-governance initiative associated with Merck KGaA, to connect governed data with evidence-grounded systems. In a 2024 conversation with Syntropy’s James Kugler, Bate returned to a risk that has become more visible as generative AI spread: the same technology can extend well-curated knowledge or amplify the gaps and biases already present in the material beneath it.

His response is architectural. Ground the model in computational knowledge. Keep experts in control of configuration. Make the evidence legible. Treat reliability as a product property rather than a disclaimer. The approach reflects his old belief that people belong inside the information system. A human review step added at the end is weaker than a system designed from the start around human scrutiny.

In October 2025, Evidium introduced a modeling product connecting clinical states, care pathways and cost for insurance and risk organizations. The following month, the company announced a $22 million Series A co-led by Health2047, the venture studio powered by the American Medical Association, and WGG Partners, with participation from Interwoven Ventures and Mindset Ventures. Evidium said the money would support product development, expanded modeling and hiring across research engineering, clinical science and business development.

Funding news can flatten a founder’s work into a number. Here the number matters mainly because it gives a long-running thesis more room to operate. Bate did not arrive at inspectable AI after the category became fashionable. His career has repeatedly returned to the same junction: complex technology on one side, accountable human judgment on the other, and a translation problem in between.

The philosophy followed him to work

There is a lighter version of Bate in an old executive questionnaire. If he were not in IT, he said, he might choose philosophy, play guitar with his blues band and spend time in beach bars, assuming his family was willing to move out of London. The answer is funny because the alternative life is already threaded through the real one. The guitar offers patterns and improvisation. The philosophy asks what can be known. The family clause keeps the thought experiment grounded.

His public X biography now says he is “building epistemological systems.” It is unusually accurate founder shorthand. Evidium is a software company, but Bate’s chosen problem sits upstream of software features. How does a machine represent knowledge? How can a person examine the path from evidence to conclusion? What happens when new observations change the state of the model? Which decisions should remain visibly contestable?

Those questions do not produce the instant gratification of a clever demo. They demand ontologies, data structures, feedback loops and the patience to make invisible infrastructure work. They also reveal the continuity in Bate’s path. The child at the ZX81 learned that instructions create outcomes. The enterprise CTO learned that people and incentives sit inside every technical system. The consultant learned that old answers fail when the question changes. The founder is trying to combine all three lessons.

Bate’s aspiration is larger than a single model and more restrained than the language of machine omniscience. He wants explanatory knowledge to move faster between research, practice and outcomes, while preserving the ability of experts to see and shape the reasoning. It is a bet that the future of consequential AI will depend less on how confidently a machine speaks and more on whether a person can understand why it spoke at all.