Five things no model does on a data room
These are not tuning problems. They are categories of work that sit outside what a language model is.
Decide what matters
A model can find every change-of-control clause in the room. It cannot tell you which one kills the deal. Materiality is a judgement about your structure, your price and your appetite, and none of those is in the data room.
Who does itThe reviewer ranks findings against the deal you are actually doing.
Cite the document that isn’t there
Retrieval reads what was uploaded. The most expensive thing in diligence is often an absence — the contract nobody filed, the year missing from the payroll export. An absence has no passage to quote, so no grounded system can raise it as a finding.
Who does itThe platform checks the room against the documents a buyer expects and lists what is missing. Chasing them is your team’s job.
Read what nobody wrote down
Management sessions, the site visit, a customer reference call, an integrity check on a shareholder. These produce the findings that move bids most often, and none of them starts as a document.
Who does itYour own team, or your advisers. This platform does not do it and does not imply otherwise.
Tell soft from hidden
A number can be weak because the business is weak, or because the seller shaped the process. On the page the two look identical. Telling them apart takes knowledge of how sellers behave, not of what the file says.
Who does itSomeone who has sat on the other side of the table: your own team, or your advisers.
Be accountable
A model cannot be a Representative under your NDA, cannot answer to your investment committee, and cannot sign. Accountability is not a capability you add to a system. It attaches to a person or it does not exist.
Who does itYour own reviewer, by name. No client export leaves the platform without that sign-off.
If a vendor claims their system does any of the five, ask which page it cited.
Why the gate is mandatory
The human step is not a safety blanket bolted onto a finished product. It is where four of the five above actually get done.
- Your reviewer’s name on the export, not a confidence score.
- Materiality decided against your deal, not against a benchmark.
- Unconfirmed claims marked unconfirmed rather than quietly dropped.
- One accountable person, on your own team.
And what the machine is genuinely better at
The honest version of this argument runs both ways.
It reads all of it
A team under deadline reads a sample and calls it coverage. Page 900 gets the same attention as page 9.
It does not get tired
The same schema, the same checks and the same standard of evidence at the end of the run as at the start. Consistency is what people are worst at and machines are free at.
It runs in parallel
4 modules opening at once instead of queueing behind each other is where the days come from — not from any single answer arriving faster.
Three questions for any AI diligence vendor
Including us. The answers tell you more than the demo does.
Who signs?
Here: your own named reviewer. Ask any vendor whether its export can leave without a human sign-off, and who that human is.
What happens to a claim that can’t be cited?
There are three honest answers: it is cut, it is flagged, or it is shown with its evidence gap visible. A system with no answer is showing you unmarked guesses.
What does it do when a document is missing?
Silence is the wrong answer. A missing document should produce an item on a list, not the absence of a finding.