What Broke When We Put AI Into Our Research Process

• Reliable AI research requires clear sources, approved rules, and traceable inputs.
• Missing evidence and unfinished work must remain visible.
• Corrections must become procedures to improve future research.

We built our AI-enabled research process for a fairly ordinary reason: consistency. Running it taught us that consistency was only the beginning.

The harder task was preserving all the small distinctions that make a piece of research trustworthy enough to inspect.

Read one company in March and another in September and it is surprisingly easy to end up with two impressions rather than two comparisons. The questions drift. The emphasis changes. What one company taught you affects what you notice in the next.

So, we formalized the research process. Evidence from filings would remain separate from calculations. Calculations would follow defined rules. Investment judgement would remain identifiable as judgement. Every company would pass through the same questions in the same order.

That was the design. Then we started running it.

The experience is still modest: one quarterly cycle across roughly twenty companies. But what surprised me was that consistency was not the hard part. Separating evidence, calculation and judgement was necessary, but not sufficient.

The process also had to preserve things we had barely thought about: which source governed, whether a statement was provisional, whether an action had actually happened, who owned a piece of metadata, and whether unfinished work still looked unfinished.

Most of the failures were mundane. That made them more useful.

At one point, two sentences written by an AI agent as a working thought ended up in a governance record. Nobody had approved them as a rule, but nothing in the record said they were provisional either.

Later agents therefore treated them as authoritative. Eventually the supposed rule stopped a piece of work I wanted to proceed. It took three agents and four exchanges to trace the restriction back to those two sentences.

The system had not invented a rule. It had lost the status of the information. Once work is divided across separate sessions and agents, the fact that something was tentative does not travel with it unless that fact is recorded. A sentence in the wrong place can acquire authority simply because later work has no other way to interpret it.

We changed the process so that a constraint must point to the rule it comes from or be marked provisional.

Another company went through the research process using a results release rather than the annual report that was supposed to govern the reading.

The work looked good. The questions had been answered, the extraction was competent, factual claims had source locations and nothing in the finished research looked obviously defective. We had asked the right questions of the wrong document.

This is different from a hallucination or a reasoning error. Repeating the same process could reproduce the mistake perfectly. The governing filing is now identified before the reading begins: an available source is not necessarily the authoritative source.

A different failure occurred during valuation. The calculation stage correctly reported that no verified market price had been supplied, so two calculations could not be completed.

The next stage was supposed to assemble the research record. Instead, it found a price itself and completed the missing sections.

The document looked better. The research became less trustworthy.

Whether the price happened to be correct was beside the point. We could no longer show how it had entered the analysis. A correct number with no provenance is difficult to distinguish from one that was invented. So a missing input now remains missing; a later step cannot silently repair the gap.

An agent could submit an edit to our workspace and receive a successful response even though the intended change had not appeared in the record. From the agent's perspective, the task was complete. The interface had said so, and subsequent work could proceed from stale information.

We began checking the records after important updates to verify that the intended changes had actually taken effect. That precaution has proved more general than the software defect that prompted it. Agentic systems constantly report on their own activity: the tool call succeeded, the task was completed, the record was updated. Those are useful signals, but they are not independent evidence that the intended result exists.

Where the consequence matters, we now distinguish between confirmation that an action was attempted and evidence that its effect is present.

A field in our company records identifies whether a company is already held or is still being considered. Initially, nobody had explicit responsibility for keeping that field current.

It looks like administrative metadata. It is not. For a candidate, the research is asking whether the company should be bought. For an existing holding, it is asking whether ownership should continue. A stale label therefore changes the question being answered.

The error was not in the analysis that followed. The analysis was answering the wrong question correctly. Once an automated process branches on a field, that field becomes part of the analytical process, and its ownership matters too.

The workflow itself remains simple: identify the sources, extract the evidence, perform the defined calculations, assemble the research record, then make the investment judgement.

Running it produced a second layer of rules:

• A statement needs a status: provisional is not authoritative.
• A reading needs a governing source: available is not the same as applicable.
• A missing input stays missing: a later stage does not manufacture completeness.
• An important action needs verification: reported success is not observed change.
• Contextual fields need owners: metadata can determine the analysis itself.

These were not principles we wrote down and then confirmed through experience. Most were written down because the first version proved incomplete.

The same applies to unfinished work. A polished record can create a false signal of completion, so unresolved items stay visible and proposed judgements remain proposed. The aim is not to eliminate uncertainty but to stop it disappearing behind a finished-looking document.

The original objective still matters. Fixed questions reduce the drift between one company and the next. Three other properties have proved useful: factual statements remain traceable to their supporting evidence; changed calculation conventions can be rerun from recorded inputs; and missing evidence can remain visibly missing rather than being replaced with something easier to obtain.

None of this establishes that the process is faster, cheaper or more accurate than conventional research. We do not yet have a basis for those claims.

The harder problems concern context. Our agents work from a common workspace containing procedures, research records and open issues, but common storage does not produce shared awareness. Information can be present and retrievable without an agent having any reason to retrieve it. A person notices a comment, remembers an earlier discussion or wonders whether something elsewhere might matter. An agent usually begins from the task it has been given.

I still find myself connecting those dots. The missing capability is not merely memory; it is recognizing that there may be relevant context outside the boundaries of the current request.

A second lesson concerns correction. When an agent gets something wrong, it is easy to spend time explaining the mistake. The agent agrees, restates the rule and describes what it should have done. That can feel like progress while changing nothing about the next run.

We have therefore increasingly moved important corrections out of conversation and into the written procedures used when the work runs again. Correcting the conversation improves that conversation. Correcting the procedure can change subsequent behavior.

Much of the discussion about AI reliability in research naturally focuses on hallucination, reasoning errors and model capability. Very little of our trouble has looked like that.

The more persistent failures have involved provenance, authority, status, context and completion: a number may be correct but unsupported; the analysis competent but based on the wrong filing; an instruction faithfully followed but never approved; an update reported as complete but never applied. Those failures arose in the research process rather than from model performance itself.

There is also an important limit. Better provenance does not make an investment judgement correct. An annual report is not the company. Customers, competitors, suppliers, managers and circumstances outside the filing still matter.

The purpose of the process is narrower: to make clear what the record supports, how a conclusion was reached, what remains unresolved and where judgement begins.