The AI Task Force at One Year

Earlier this year, positive feedback from management on the James 1.0 prototype became the starting point for expanding a personal technology experiment into an organizational AI platform. James is Centroid’s proprietary AI platform, designed to structure data across investment, operations, and legal functions and connect it with AI agents in real time.

At the outset, the AI Task Force was focused on what could be built and how quickly it could be delivered. Looking back after a year of execution, however, this has been as much a year of developing AI services as it has been a year of learning the decision-making structure behind private equity investing and operations.

I joined the firm without fully understanding what private equity professionals actually do. As a developer, I was accustomed to sprints, release cycles, and user reactions that could be measured quickly. This was different. A single deal may be reviewed over months—or years—and teams may spend days examining one number or one clause in a contract.

At first, it felt slow. I now understand that this is how people work when they are entrusted with managing other people’s capital. When a decision cannot simply be rolled back, caution is not inefficiency; it is responsibility.

The AI Task Force’s first attempt was an agent designed to prepare investment review reports. The expectation was simple: if we could turn a specific task into an agent, useful results would follow quickly. In practice, that assumption did not hold.

Investment review is not a standardized task. It brings together deal context, the reliability of source materials, an investment professional’s hypotheses and decision criteria, and knowledge spanning investment, operations, and legal functions. The demo looked convincing and the output resembled a report. Yet after trying it once or twice, users quietly returned to their original way of working. No one said it was bad. They simply did not use it.

The experience left us with a clear lesson. Rather than automating isolated tasks one by one, we first needed to connect the organization’s knowledge and materials—and enable people to collaborate with AI agents using their own expertise to create better outputs. James was built as a general-purpose AI agent platform from this realization.

James evolved from a personal tool into an organizational work platform through its connection to real deal contexts and feedback from early users. We linked servers, archives, and work histories into a knowledge base, and expanded the scope of use to include disclosure and patent research, contract review, VDR analysis, and recurring work automation.

But the success of AI transformation is not determined by the number of features. It depends on the source and timeliness of data, access permissions and security, and the business context of the person asking the question. Designing who can see what, under which standards outputs are validated, and who takes responsibility for the result is the essence of data readiness.

This principle became even clearer in portfolio-company settings. To design practical AI use cases, we needed to understand how critical business variables—production, quality, scheduling, pricing, demand, and competitive conditions—feed into management decisions. In industries where data is highly sensitive and regulation is strict, document archiving, access controls, and legal and compliance boundaries must be addressed before the technology itself.

AI is therefore not a solution that can simply be installed. It is a process of redesigning workflows, data, training, and operations together. This is also the foundation for commercializing selected capabilities in other industries, including real-estate-focused applications.

The most difficult part was not technology, but human habit. Very few people openly opposed the initiative. Most agreed with the direction in meetings, but often returned to familiar ways of working afterward.

“The deal is urgent.” “This material is too sensitive to upload.” “It is faster if I do it myself.” “If the AI is wrong, who is responsible?” Each concern is valid. Deals are urgent. Security matters. Excel and Word can be faster for people with established routines. But when these concerns accumulate, change does not happen.

There are understandable reasons for this. Private equity is an industry where a small group of professionals generates performance through judgment built over long, apprenticeship-like training. Introducing an unproven tool into that workflow can itself be perceived as risk. Similarly, while sharing code is natural in development, sharing deal materials and professional know-how can feel very different in an industry where such knowledge is closely tied to individual competitiveness.

When AI transformation was a focus of company-wide meetings, workshops, or management engagement, James usage increased and improvement requests followed. But when a major deal emerged or organizational attention shifted, users tended to drift back to individual folders and familiar workflows. The point is not to blame anyone. The pattern showed that trying a tool once is very different from making it habitual. Closing that gap requires time, repetition, and operational support—not feature development alone.

Management support and active use were therefore critical in turning the AI Task Force into an organizational priority. There is a meaningful difference between a working-level manager suggesting that a team try a new tool and a CEO asking in a meeting, “Have you run this through James?” The former is a request. The latter becomes part of the work.

Management’s repeated questions and specific feedback showed us what kind of support AI should actually provide in their decision-making. AI does not grow through hypothetical users imagined behind a screen. It grows through users’ real questions, frustrations, and recurring work.

Business leaders also served as a bridge between operating reality and technology. They helped reframe AI initiatives across investment review, due diligence, PMI, portfolio operations, and commercialization. The key question was not simply how complete a feature was. It was which problem should be addressed first, which capabilities should be standardized as shared assets, which elements should remain tailored to individual clients, and how technology could become a repeatable organizational capability.

The AI Task Force was never the work of the AI/Data Cell alone. It has been the result of collaboration among the investment team, the operations team, and the AI/Data Cell—each learning from and connecting with the others.

The operations team helped link portfolio-company challenges, field data, and priorities, while establishing a clear standard that AI must ultimately contribute to operating performance. Finance and accounting perspectives helped test the quantitative and qualitative strength of outputs, challenging assumptions, numbers, and execution feasibility.

That discipline was an important learning experience. In development, it can be tempting to believe that demonstrating something that works is sufficient. In investing and operations, the first question is often: “Where did this number come from?” An output without a traceable source cannot be used in this industry.

Legal and investment professionals were among the most practical users of James. They clarified which questions were genuinely useful, where AI-generated outputs fell short, and where AI support should end and human judgment must begin. Some team members have also begun creating small AI agents of their own and applying them to their work.

This is where the real value lies: business professionals reshape technology in their own language, while technology learns the standards by which the business makes decisions.

There is still much to do. Work experience and materials across the organization must be better captured as common resources, and AI must become more capable of evolving on top of that accumulated knowledge. We must continue developing standards and training suited to different industries and clients, while building use cases that translate into measurable portfolio operating performance.

Our role is to connect the resources people need to perform at their best—to provide well-crafted tools and an environment that makes adoption easier. The goal is to create a stronger cycle in which people produce work alongside AI agents, and the knowledge generated through that process becomes an asset that supports the next person’s judgment and execution.

Compared with a year ago, the change is unmistakable. Centroid is becoming an organization where people can engage with AI more naturally, explain their work more clearly, ask better questions, and validate results together.

This is more than a productivity initiative. It is a shift toward a structure in which investment, operations, legal, and AI/Data professionals understand one another’s work and turn their expertise into outputs through collaboration with AI agents.

When technology and the language of the business begin to meet in the same place, organizational knowledge can become a shared asset. That is where organization-wide value begins.