HomeEnterprise AI

Enterprise AI Matures When The Entire Workforce Centralizes Context In One Shared Repository

November 18, 2026

Oren Hizkiya, Co-Founder and Chief Technology Officer of Four, on why one shared repository of institutional and process knowledge, not more tools, is what lets AI run autonomous work and keeps headcount flat.

Enterprise AI Matures When The Entire Workforce Centralizes Context In One Shared Repository
Credit: CIOnews

Get the latest from CIOnews.

Enterprise AI, governance, risk, and leadership insights for CIOs, CTOs, CISOs, and technology leaders.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
"Our posture is that the company brain, the central context co-located in the repository, is so important that we make it a point from day one. Everyone gets their changes in, and they learn how to do it over time."

Oren Hizkiya

Co-Founder & CTO
Four

Most enterprises chasing AI at scale keep buying tools and wiring up agents, then watch the gains stall. The knowledge those agents need to work—the data, the processes, and the hard-won corrections—sits scattered across disconnected systems. It's a problem data built for analytics was never designed to solve. The companies pulling ahead treat that knowledge as shared infrastructure instead: consolidated and version-controlled in one place, where every agent and every employee can reach it. That consolidation is what turns an impressive demo into AI that executes real work without a person supervising every step.

Oren Hizkiya is the Co-Founder and Chief Technology Officer of Four, a consumer finance company acquired by PROG Holdings. He built his career in financial technology, with earlier engineering and leadership roles at MKP Capital Management and JPMorgan Chase. At Four, he's made a shared repository the operational core of how the company runs, and his conviction about that starts with a single onboarding rule.

"Our posture is that the company brain, the central context co-located in the repository, is so important that we make it a point from day one," Hizkiya said. "Everyone gets their changes in, and they learn how to do it over time." The rule applies to everyone, not just engineers. A new finance or accounting hire who has never seen a repository is expected to push a change on their first day, because the knowledge they carry is exactly what the company's AI systems need to reach. The repository isn't documentation sitting beside the work. It's the substrate the work runs on.

  • One place by default: The architecture is deliberately simple. Nearly everything the business knows goes into the repository, and the transactional data that doesn't lives in a single consolidated warehouse. "By default, as much as possible makes it into the repository. We lean towards overusing it," Hizkiya said. "Everything about the business, in the context of the business, can be found in one of those two places." The reach is the point: the repository even stores the skills that tell an agent how to query the warehouse, so the context and the instructions for using it live together. That instinct toward a single foundation tracks with what studies of scaling agentic AI keep surfacing, that autonomy stalls without consolidated, governed data underneath it.

  • Portability over lock-in: "There are so many tools coming and going, and they're sort of baiting you into getting your tribal knowledge stuck in them," Hizkiya said. "We make it a very important exercise not to get coupled to any of those." The defense is routine. The team runs what he calls katas, regular drills that eject knowledge, corrections, and the occasional hard-learned mistake back into the repository, so no single vendor's tool becomes the place a critical process secretly lives.

  • Process, not just facts: Facts about the business are one thing. Knowing how the work actually gets done is another, and Hizkiya argued the second matters more. "They're equally important, but I'd argue that sometimes process knowledge is more important. That's the type of stuff that tends to get lost if it isn't formalized." The hard part is extraction. Pulling a seasoned operator's steps into a documented skill means capturing the soft rules too, the judgment calls in how someone runs a particular review, the kind of documented, shared knowledge that outlives any one employee. He framed the whole effort as reinventing each process as if the company had been built around AI from the start.

Capturing knowledge this thoroughly isn't an end in itself. It's what makes the next tier of AI use possible. Once the context is rich and reachable, Hizkiya's teams can hand models work that runs for days rather than minutes, and they can decide, case by case, when a mature process should stop being a flexible skill and become hard code. Both moves lay the groundwork before autonomy scales, and both change what an individual contributor is actually responsible for.

  • Beyond the one-off prompt: Early adoption meant treating AI as a better search box, a one-and-done answer to a single question. That phase is over. "The models are much more capable of long-standing thought, almost like self-reflection, and much more autonomous runs," Hizkiya said. "We're coaching people to a place where they're empowered to do more than the one-off." Instead of asking an agent to do one task, his team now inventories every change they want across an application and sets a model loose on the list, running for 48 hours or more, working through it item by item. Getting people to think at that scale is the harder shift, and it only holds up when the context behind each decision is solid enough to survive contact with production.

  • When to harden into code: A skill written in natural language is flexible, but flexibility costs speed and tokens. Hizkiya watches for the moment a process has earned a more permanent form. "Certain skills jump out at you," he said. "I run this 20 times a day, it costs tokens, it's highly deterministic and mature. Now's the time to turn it into code." The conversion pays off twice: execution gets faster, and token spend drops, a cost discipline more leaders are starting to formalize. Deciding when to make that move, he noted, is its own skill worth paying attention to as the tooling matures.

  • Measured by what ships: Hizkiya has little patience for adoption counts as a success measure. The metric he watches is merged pull requests across the entire team, including from people in roles that never touched source control before. "When you're used to working in a manual manner, your thoughts of scale and what you even think is achievable are limited," he said. "I try to help people have those aha moments, when they realize, 'we don't need to bring on three other people to help with this.' My job gets easier if I think in a more abstract way that lets me do things that are much more scaled." The business result is concrete: a company that keeps scaling without adding much headcount, which he attributes directly to how it uses AI. "We would not get around to a lot of this if it weren't for AI. We never have, in the existence of this business," Hizkiya said. "Look at the power of ambition. Just ask."

None of this works if contributing is hard. Hizkiya treats the platform itself as his responsibility, building it so someone who identifies as non-technical can ship on day one without first mastering source control. His team even coaches people to work by voice, a meandering spoken brain dump of whatever is on their mind, and lets skills handle the organizing, the pull request, the notifications, the tickets. Remove the friction, and the contributions follow.

For Hizkiya, the system comes back to ambition, and to refusing the ceiling that manual work imposes by default. "It's very empowering when people get used to consistently landing their own pull requests, especially in roles that were never opening them before," he said. "It helps consolidate the team."

research report

From the Edge to the Core:
Bringing Agentic AI to the Heart of the Enterprise.