Sometimes — and the cases where it does share a property worth understanding. AI catches deal-killers that are written down in a document nobody was going to read. It does not catch deal-killers that live in judgment, in absent documents, or in the gap between two systems. Almost every real success and every real failure in this category falls cleanly on one side of that line.
What actually kills deals, and where each one hides
| Deal-killer | Where it lives | Findable by screening? |
|---|---|---|
| Change-of-control in a top-10 customer contract | Written, in a document ranked low by folder position | Yes — the core case |
| Undisclosed litigation exposure | Written, often in correspondence or board minutes | Yes, if those documents are in scope |
| Cross-default linking two facilities | Written, but split across two documents | Only with cross-document capability |
| Customer concentration above tolerance | Derivable from contracts, not stated anywhere | Partially — needs aggregation, not extraction |
| Regulatory approval that will not be granted | Judgment about a future decision | No |
| Key person who intends to leave | Nowhere. Not written down | No |
| Oral side agreement with a distributor | Nowhere. Deliberately not written down | No |
| Culture or integration incompatibility | Judgment | No |
Read the middle column top to bottom. The dividing line is not difficulty and it is not model intelligence. It is whether the fact exists in the document set at all. Everything above the line is a coverage problem. Everything below is not a document problem in the first place, and no amount of model capability will change that.
The one it genuinely catches, and why humans miss it
The archetype is a change-of-control or assignment provision sitting in a mid-tier commercial contract — not the top five customers everyone reads, but numbers eleven through forty, which collectively represent material revenue.
Humans miss these for a structural reason that has nothing to do with competence. In a large data room, reading order is set by folder structure, upload sequence, and whoever flagged something on a call. None of those correlate with risk. A senior reviewer with time for 30% of the room reads a 30% selected by arrival, and a provision in the unread 70% is invisible — not judged unimportant, simply never seen.
That reframing matters commercially, because it sets a claim you can actually defend. “Our AI catches what lawyers miss” invites a fight it will lose. “Every document was examined and ranked, so your reviewers spent their hours on the highest-risk material” is smaller, true, and enough.
Timing is most of the value
A deal-killer found in week four and the same one found on day two are worth very different amounts, and the gap is larger than most teams model.
- Day two: you stop. The diligence workstream never runs. Adviser fees are not incurred. The team redeploys to another opportunity. Cost: two days.
- Week four: you stop. Legal, financial and commercial workstreams have all run. Advisers have billed. The team has spent a month. Cost: the entire workstream.
- After signing: you renegotiate or you complete and carry it. Leverage is materially worse and the outcome is priced against you.
This is the single largest financial effect available from screening, and it is almost never claimed in vendor material because it cannot be attributed cleanly — nobody can prove the deal would have run four more weeks. It is nevertheless real, and it is measurable in your own history: take a deal you abandoned, and total the spend between kick-off and the moment the disqualifying issue surfaced.
Note the operational precondition. This value only exists if screening runs before the expensive workstream rather than alongside it. A screening pass that starts in week three, once the data room is “complete,” forfeits most of what it is worth.
Where it fails, stated plainly
The document is not there
The most common cause of a missed deal-killer is not a model failure. It is a seller who did not upload the amendment. Screening reads what it is given, and a clean result on an incomplete data room is a confident answer to the wrong question. This is why coverage reporting — what was examined, what was expected, what is absent — matters more than finding quality.
The issue is an aggregate, not a clause
Customer concentration is not written in any contract. It emerges from reading forty of them and adding up. Similarly, a covenant package may be individually unremarkable and collectively unworkable. Systems built around per-document extraction miss this class by construction, and it is worth asking a vendor directly whether anything in their product operates across the set rather than within a document.
Severity was assessed against the wrong baseline
A clause read correctly and scored against the wrong sector or jurisdiction produces a low-severity finding that looks exactly like a correct one. Nothing on the face of the report reveals it. This is the quiet failure mode in cross-border and specialist-sector deals.
The finding was right and nobody believed it
A system that over-flags trains its users to skim. By the fortieth “critical” finding that turned out to be market-standard, the forty-first is not read carefully. False positives are not harmless noise; they consume the attention the tool exists to direct, and they are the most common reason a genuinely good finding fails to change a decision.
How to configure for deal-killers specifically
Most screening is configured to find everything, which is the wrong objective. If the goal is catching transaction-enders, four choices matter:
- Define blocking categories in advance. A small set — the issues that genuinely stop a deal in your sector — with a score threshold that escalates immediately. Not a long list; if everything blocks, nothing does.
- Make absent disclosure score non-zero. A blocking category with no findings should not read as clean. It should carry an explicit floor and a stated diligence request, so “nothing found here” is visibly different from “nothing was looked for.” This single choice converts the most dangerous failure mode into a visible one.
- Run it on day one, on whatever exists. Partial coverage early beats complete coverage late for this purpose. Re-run as documents arrive.
- Require the source sentence on every blocking finding. A deal-killer that stops a transaction will be challenged, by the seller and internally. It needs to survive being read.
The honest summary for your committee
Do not say AI catches deal-killing issues. Say this instead:
That claim holds up when a deal-killer is later found in something the system flagged as low-severity — because it was never a promise of detection, it was a promise of coverage and ranking, and the record shows both.
Bottom line
AI catches the deal-killer that was written down in a document nobody prioritised. That is a narrower claim than the category markets, and it happens to describe a large share of the expensive surprises in real transactions, because the expensive surprises are usually not exotic — they are ordinary provisions in unread documents.
Run it early, define what blocks before you start, make absence of a finding visible rather than comfortable, and keep materiality with the people who can be accountable for it.
Sources
This page makes no external statistical claims — the arguments are structural and testable against your own transaction history. Where we cite figures elsewhere on this site, we cite only sources we have retrieved and read; where a figure could not be verified, we changed the figure rather than the citation. See our methodology. Nothing here is legal advice.