Over the past two months, we have been turning our research process into an AI-enabled, agentic workflow.
The problem is simple. Read one company in March and another in September, and you can easily end up with two impressions rather than two comparisons. The questions drift. The emphasis changes. The second analysis is shaped by what the first one taught you.
We have built a process designed to prevent that.
It separates three things that are often blurred together:
- Evidence comes from the filings.
- Calculations follow defined rules.
- Judgement remains explicit and attributable to a person.
When the evidence does not determine the answer, the system does not manufacture one. It records the question, sets out the competing interpretations and escalates the decision.
What is live today
The first company has now completed the process end to end. The workflow is operational, although the final Company Sheet format remains in testing and will evolve as more businesses pass through it.
Specialised AI agents acquire the documents, apply the research procedures, perform the calculations and assemble the record. Each stage is gated on the one before it and cannot proceed until the required work is complete.
Every company is reviewed through the same Company Sheet, using the same questions, in the same order and to the same standard of evidence.
Guided by the Steward Productive System Framework, the research covers the economics of the business, management incentives and the quality of its finances. Every finding is traceable to the company’s own filings.
The next milestone
We are now working towards an agentic investment committee process: every position supported by research produced to the same standard, and every decision recorded alongside the evidence considered and the judgement applied.
Quality research first. Then decisions that stand up to inspection.
