Can AI Predict Deal Success or Failure?

Learning about predictive analytics capabilities and their current limits

Updated August 2026 · 5 min read · Deal Room Intelligence Series

No, and the reason is worth understanding rather than accepting on authority. Deal outcomes fail every condition prediction requires: there is no agreed definition of success, the sample is tiny, the outcome arrives years later, and the most decisive variables are never written down. A model trained on this produces a number — and the number is confidence, not information.

The four conditions prediction needs, and how deals fail each

ConditionStatus in M&A
A stable definition of the outcomeAbsent. Success means synergies realised, or IRR hit, or strategic position acquired, or the founder retained — and these disagree on the same deal
Enough labelled examplesAbsent. A prolific acquirer does a handful a year; most firms have dozens of outcomes total, not thousands
A short feedback loopAbsent. Outcomes resolve over three to seven years, by which point the market has changed
Predictive variables in the dataAbsent. Integration execution, key-person retention, cultural fit and macro timing dominate — none are in the data room

Any one of these would make prediction hard. All four together make it structurally unavailable, and no improvement in model capability changes that — the constraint is the data-generating process, not the algorithm.

The confound that ruins the training set

Even with sufficient volume, there is a problem specific to this domain: the deals that would teach you the most never closed.

A well-run diligence process kills bad deals. Those transactions have no outcome to label — they are neither successes nor failures, they are absences. So a model trained on completed deals learns from a population that has already been filtered by the very judgment you are trying to automate.

The result is a system that learns what closed deals look like, not what good ones look like. Those are different things, and the difference is invisible in the training data.

What people mean when they say “predictive”

Most claims in this area, examined closely, are one of three things — all useful, none predictive:

The useful reframe: screening measures what is true now; it does not forecast what will happen later. A change-of-control provision in a top-ten customer contract is a fact with consequences. Whether the deal succeeds depends on things no document contains.

Why a false prediction is worse than none

An unreliable forecast is not neutral. It is actively harmful for two reasons.

It anchors the committee. A “73% success probability” on the front page changes how everything behind it is read. Findings that contradict the score get discounted; findings that support it get weight. The number does work it has not earned.

It creates a false audit trail. If a deal fails, “the model scored it 73” is not a defence — it is evidence that a decision rested on an unvalidated number. ABA Formal Opinion 512 (29 July 2024) holds that professionals using generative AI must “fully consider their applicable ethical obligations,” including competence.[1] Relying on an output nobody validated is difficult to characterise as competent reliance.

What is genuinely knowable from the documents

Quite a lot, and it is more decision-relevant than a forecast:

None of these predict success. All of them change the price, the structure, or the decision to proceed — which is what diligence is actually for.

The honest version of a scoring output

A risk score is not a prediction and should never be presented as one. It is a ranking device: an ordering that tells you which documents deserve expert attention and which have been examined and ranked below a threshold you set in advance.

The distinction matters operationally. A ranking is falsifiable — a reviewer can say “that clause is market-standard, this should score 30, not 82,” and the disagreement is recordable and improves the rubric. A prediction is not falsifiable until years later, when nobody can attribute the outcome to it anyway.

It also matters legally. “This document scored high on our defined rubric and a named reviewer examined it” describes a managed process. “Our model predicted the deal would succeed” describes reliance on something nobody can validate.

The question to ask a vendor

If a product claims predictive capability, ask one thing: “Validated against how many completed deals, over what period, with success defined how?”

The honest answers are all disqualifying. Either the sample is too small, or success was defined post hoc, or the validation set was the training set, or — most commonly — there was no validation and the score is a weighted sum someone designed.

Ask the same question about failure: how many deals in the validation set actually failed? A model trained overwhelmingly on successes cannot distinguish them from anything.

Bottom line

Deal outcomes cannot be predicted from diligence documents, and the obstacles are structural rather than technical: no stable outcome definition, no sample size, no feedback loop, and the decisive variables absent from the data entirely.

What AI does well here is measure the present — every document examined, every category scored, every finding traceable to its source. That is not a forecast. It is a better basis for the human forecast, which is the only kind there is.

Sources

  1. ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512: Generative Artificial Intelligence Tools, 29 July 2024. americanbar.org

We publish no deal-outcome prediction and make no predictive claim. We cite only sources we have retrieved and read — see our methodology.

Measure what is actually in the document →

Anweshna Portal
Anweshna Demo