Much of the conversation around AI in financial research still starts with a chat window. You ask a question, wait for an answer, and decide what to ask next.
That is useful. But when I think about how company research should work in CEEWire, I imagine something that can continue while I step away: a research process with a clear objective, several lines of inquiry running in parallel, and findings ready for review when I return.
The interesting question is what happens around that work. How should the system use my feedback? What should it remember? And how do we prevent a system that understands my preferences from becoming too willing to confirm them?
From a conversation to a research process
Chat is a largely synchronous way of working. Even when a model takes time to answer, the user often remains in the loop, reading each response and supplying the next instruction. The conversation advances when the user returns to it.
Company research rarely follows such a tidy sequence. Understanding the business model, examining ownership, and tracing the use of cash are related tasks, but each can begin before the others are complete.
A background workflow should reflect that. Given a research objective and the right context, agents could work across several dimensions of a company:
- •Financial performance: how revenue, margins, and the balance sheet have developed.
- •Ownership and control: who owns the business and how those relationships affect minority shareholders.
- •Cash flow and capital allocation: where cash comes from, where it goes, and how management balances investment, distributions, and financing.
- •Capital expenditure: what the company is building, maintaining, or committing to.
- •Business model: how the company earns money and what could change its economics.
The findings then need to be brought together. A capital-expenditure figure means more when it is connected to the business strategy, the funding required, and management's explanation. Parallel work is valuable only if those connections survive.
What an hour of background work could mean
My view is that a well-supported research run lasting 30 minutes to an hour can produce a first pass comparable to work I would ask a junior analyst to prepare. The qualification matters: the data must be available, properly structured, and connected to its sources. The output still needs review.
When several research tasks run in parallel, I believe the preparation saved could amount to more than a week of manual work on a company. That is my estimate of the opportunity, rather than a measured productivity benchmark. The result will depend on the company, the available evidence, and the depth of the questions being asked.
The benefit is having material ready to examine: the relevant evidence assembled, the main relationships explored, and gaps made explicit. An analyst can then spend more time testing an interpretation and deciding which questions deserve further work.
This is why I see background agents as central to the value of AI in research. You should be able to define the work, let it proceed, and return when there is something substantial to review.
Feedback should change the analysis
A research workflow should continue beyond its first report.
Suppose I review an analysis and point out that a cash outflow treated as recurring was related to a one-off investment. The useful response is to revisit the cash-flow interpretation and identify which conclusions depended on it. If that interpretation also informed a view on dividend capacity or financing needs, those parts should be reconsidered too.
The system should make those changes visible: what was revised, why it was revised, and which questions remain unresolved.
Over time, some feedback should also improve the research process itself. A correction to one company's figures is specific to that analysis. A request to distinguish maintenance from growth capital expenditure may be a reusable instruction. A preference for a particular presentation format belongs in another category again.
Playbooks offer a way to express that reusable process. They can capture the questions a team wants answered, the evidence it expects, and the checks it considers essential. Learning from feedback should make those instructions more useful and explicit.
The risk of learning what I want to hear
There is a tension here that I have not fully resolved.
In the philosophy behind CEEWire, I argued that private context makes a research system more valuable. Internal notes, assumptions, and investment processes help it understand how a team works. But the more a system learns about an investor, the easier it may become to reproduce that investor's biases.
If I repeatedly focus on a certain kind of business or favor a particular interpretation, the system may learn to give those ideas more attention. That can make the output feel increasingly relevant while narrowing the range of explanations it considers.
Investment research benefits from unexpected connections, competing interpretations, and evidence that makes a preferred thesis uncomfortable. Some of the value of a probabilistic system lies in its ability to approach a question differently. I would not want personalization to remove that variation.
A system should learn how I investigate a company while continuing to test what I believe about it.
That means preserving space for alternative explanations and asking what evidence would weaken the thesis. It also means treating my feedback according to what it contains. A factual correction should update the analysis. An opinion should remain an assumption that can be challenged.
Where this leads for CEEWire
This is the direction I want research workflows in CEEWire to take: background work grounded in sources, parallel analysis brought into a coherent view, and feedback that improves both the immediate result and the process behind it.
These are design goals. The ability to run a long task is only one part of building a system that can carry out useful research, respond well to corrections, and retain enough independence to disagree.
The test I keep coming back to is simple. When I return to the work, does it help me understand the company better? And when I respond, does the next analysis become more rigorous—or merely more agreeable?
I would like to hear how other investors think about that balance. Reach out to [email protected].