The checking problem
Speed was never the hard part of research. Trust is.
Everyone is being sold speed right now. Faster summaries, faster drafts, faster answers to harder questions. The pitch behind most AI tools for finance is the same: research is slow, and this makes it fast.
We think that pitch misunderstands the job.
Ask anyone who has actually run a deal process where the time goes. It is rarely the first answer. A capable analyst can produce a view on almost anything in an afternoon. The time goes into what happens after the first answer: confirming the revenue figure against the actual filing, finding the footnote that changes the margin story, reconciling the number in the CIM with the number in the data room, tracing a claim back to something a committee can rely on.
In other words, the time goes into checking. Checking is not the boring part of research that happens around the real work. Checking is the work. It is the difference between a view and a finding, between something that sounds right and something you would put your name on.Good checking does not always win the deal. Bad checking loses it.
This is why the current generation of AI tools has landed awkwardly in professional research. They are extraordinary at producing first answers and mostly indifferent to what those answers rest on. The output reads well. Then someone asks where a number came from, and you are back in the data room at midnight, checking line by line. The tool did not remove the work. It moved the work to the worst possible moment, right before the answer goes in front of a client or a committee. And there is no transfer of responsibility. You cannot blame the model for something you did not catch.
There is a simple rule underneath this. In most fields, unverified work is a draft. In finance, it is a liability. An investment committee does not take vibes. A fund’s LPs do not take vibes. The regulator certainly does not. Every claim that cannot be traced to a source is a small unexploded risk sitting in the memo, and everyone in the room knows it.
So we would suggest a different measure for research tools. Not time to answer. Time to a checked answer: a source, a reference, something you can verify. How long does it take to get from a question to something you can defend, source attached, at the moment someone challenges it? On that measure, a tool that produces instant unsourced answers has not made you faster. It may have made you slower, and possibly worse, because now there is more plausible-sounding material to verify.
This is the standard we have built Analyst One to. Every answer is cited back to the exact page it came from, in the filing, the data room document, the transcript. Checking a claim takes seconds, because the source is attached to the answer rather than waiting to be hunted down. The goal is not to remove judgement from research. It is to remove the midnight part.
Research that shows its work used to be the expensive option. It should be the default.
