The documents you cannot paste
The valuable half of the job never reaches the tool.
The documents you cannot paste
There is a quiet arrangement inside most deal teams right now. People are using AI, and they are using it on the material they feel safe using it on. Public filings. A quick industry primer. Rewriting an email so it sounds less annoyed than the first version. Then they open the CIM, or the data room, or the management pack, and the tab stays closed.
Nobody announced this. There was no memo. Everyone simply knows roughly where the line sits, and nobody fancies being the person who establishes its exact position after the fact.
So the tool ends up working well on the least valuable part of the job.
This is what gets missed when a firm reports that AI has not changed much for them. Usually the model is fine. The model just never saw anything that mattered. In private capital the work that carries the value sits inside material that cannot leave the building: provided under an NDA, restricted by a side letter, covered by a confidentiality undertaking the analyst has not read but the general counsel certainly has. A tool that cannot go near any of that is not a research tool. It is a writing assistant.
We argued last week that the right measure for these tools is time to a checked answer. This is the question that sits before it. You cannot check an answer about a document the tool was never allowed to read.
Firms have mostly responded with policy. An AI policy gets written, a list of approved tools goes round, training happens. That is worth doing and it does not fix this, because the issue was never that people are confused about the rules. The rules and the tools are pointed at different things. An analyst who understands the policy perfectly still has no way to ask a question of a 400 page data room without moving the data room somewhere it should not go.
What would actually change the picture is narrower than it sounds. Three things.
The documents stay where they are. Not copied into a general purpose model, not sitting in a consumer account under someone’s personal email. If the material has to move before the tool works, the tool has already failed.
Someone can see what happened. Which document was asked about, what came back, who did it. Not because anyone expects wrongdoing, but because a firm that cannot answer that in an audit will not approve the tool, and should not.
Permissions carry over. An analyst who cannot open the folder cannot query the folder either. That sounds obvious and it is the thing most tools get wrong, because retrieval is far easier to build when everything sits in one bucket.
None of this is exciting. It is the difference between a tool people are allowed to use on real work and a tool that lives in the margins of the day.
The adoption question in this industry has been framed the wrong way round. It is not whether analysts will use AI. They already do, occasionally further past the line than their firm would like. It is whether anything exists that can be pointed at the documents where the value actually sits. Until that is true, every productivity claim in this market is being measured on the easy half of the work.
Analyst One runs on the documents you cannot paste anywhere else. That is the point of it.