M&A · investment screening
AI Perception Due Diligence for M&A and Investment Teams
The data room shows the company’s version. AI shows the market’s. The gap between them is already shaping attention.
When a venture associate prepares for a first call, when an M&A analyst builds a screening memo, or when a buyer compares vendors, there is a growing probability that the first research step is a prompt to a large language model rather than a visit to the company’s website or data room. The model returns a compressed judgment synthesized from everything it has absorbed about the firm. That judgment arrives before anyone at the target has spoken and before the formal diligence process has begun.
Different models return different judgments. One may describe a defensible platform; another may treat the same company as a commodity tool or omit it from category lists entirely. The variation is currently invisible in most deal files. The memo records conclusions; it does not record which model produced them or how widely models disagreed.
VeritasLinks measures the distribution across models, the specific positioning and risk language they attach, and the separation between brand perception and narrative evaluation. The output is designed to be retained alongside other diligence materials so that the first AI judgment becomes a recorded input rather than an untraceable influence.
Live report surface
VeritasScore · multi-model distribution
Northline Analytics · illustrative fixture on the real UI. Not a screenshot.
Score by LLM model
620Overall · 300–870
Two distinct entry points in the deal process
Sourcing versus screening
Sourcing and pipeline construction
Category-shaped and “best tools / best platforms” queries produce the short lists from which many pipelines are built. Generative answers are winner-take-few: the model names a small set of entities and stops. An omitted company is not ranked low; it is absent, and its exclusion generates no signal to the omitted party or to the analyst who never sees the name.
In measured data, 6.4% of companies with category-query coverage from at least two models were simultaneously omitted entirely by one model and ranked in the top three by another. For those firms, whether they appear in an AI-assisted sourcing list is decided by which assistant the analyst happens to use. That is a structural feature of the current AI surface, not an edge case.
Screening and pre-call framing
Diligence-shaped queries (“risks and weaknesses of Z”, “concerns about Z as an acquisition target”) function as a draft of the pre-call or early screening memo. The median cross-model score range of 100 points (90th percentile 134 points) means the same target can read as belonging to different quality or risk tiers depending on the model consulted. The memo that results typically records the conclusions, not the tooling that shaped them. The variation is invisible ex post.
Both stages benefit from measurement. Sourcing measurement surfaces omission risk. Screening measurement surfaces positioning splits, risk language, and the gap between how models describe the firm and how they evaluate its narrative when placed in an explicit analyst role.
Live report surface
Perception affect field · split detected
Northline Analytics · illustrative fixture on the real UI. Not a screenshot.
Perception affect field
Models split into two emotional camps
Polarization
64%
Division is material, not noise.
Dominant zone
Activated
Brand-scope versus narrative-scope assessment
Two objects, separable scores
Models assess a company and the account the company gives of itself as separable objects.
Brand-scope measurement captures how models describe and position the firm across open buyer-shaped, category, comparative, and diligence queries. It answers: is the company named, where is it placed, what sentiment and risk language attach to it, and how much do models disagree?
Narrative-scope measurement (AI Investor Review) places models in an explicit evaluator role — investment reviewer, acquirer analyst, or similar — and scores a pitch deck, business plan, or sale memorandum against a fixed rubric. It answers: how does the story land when models are asked to evaluate it rather than merely describe the company?
Because the story is scored by its own rubric rather than folded into brand signals, the two scores can diverge on the same measurement date. The gap locates in the evaluative layers (synthetic panel and narrative review) rather than in simple mention statistics. Which of the two objects an associate’s research actually consulted is determined by query shape and is almost never recorded in the deal file. Measuring both makes the distinction visible.
For the target company the narrative review functions as advance notice of reservations that human reviewers later raise. For the deal team it functions as a structured summary of how the AI layer their own analysts already consult perceives the target’s story.
Live report surface
AI Investor Review · narrative-scope
Northline Analytics · illustrative fixture on the real UI. Not a screenshot.
AI Investor Review
Investor Lens rubric — 5 parameters
72/100
Investor Readiness
Top 3 fixes
- Add current ARR, paid customers, or signed pilots on a dated traction slide.
- State the raise amount, instrument, and a 100% use-of-funds split.
- Replace vanity user counts with an active / paid breakdown.
What the full assessment delivers to a deal team
From single-model draw to reviewable distribution
A complete assessment produces:
- Overall VeritasScore (300–870) with full component decomposition.
- Per-model scores and the explicit range (the split).
- Mention, position, sentiment, and omission data across query families.
- Affect-field map and polarization index.
- Coded objections from synthetic panels with retained transcripts.
- Citation and authority profile.
- When a deck or memorandum is supplied: separate narrative score and rubric-level reservations.
The median observed split of 100 points is large relative to the 200-point width of the reporting bands. Treating that spread as a finding — rather than an inconvenience — is the core procedural shift. A wide split is diagnostic of a thin or contested evidence base, which is itself due-diligence information.
Immediate procedural mitigations
What teams can do even before a full assessment
The paper and the product both emphasize mitigations that require nothing beyond access to several models and a record of what was asked:
- Query several models on the same diligence-shaped and category-shaped questions and record the spread.
- Treat a wide split not as noise but as a finding about the evidence base.
- Where the assistant is retrieval-grounded, repeat the query across several days before treating any single answer as stable.
- Record model name and version in the deal file so that the tooling that shaped attention is at least reconstructible.
- When the target reaches deeper screening, run a full VeritasLinks assessment (brand-scope plus narrative review if materials exist) and retain the timestamped report alongside other diligence materials.
These steps convert an invisible influence into a recorded input.
Research foundation
Measured, not anecdotal
All quantitative statements about divergence, stability regimes, and omission rates are drawn from the working paper “Split-Perception: Measuring Divergence in How Large Language Models Assess Company Credibility” and the underlying VeritasLinks measurement dataset. The paper frames split-perception explicitly as an unpriced information risk for M&A and investment screening workflows.
Make the first AI judgment visible in the deal file
The first impression is already being formed by models. The operational question is whether that impression, and the disagreement around it, is measured and retained or left untraceable.
Free analysis
Run it on any company
Public URL. Same funnel as the homepage. No card required.
Enter a public website
Multi-model · VeritasScore · ~5 minutes
See a sample dossier →FAQ
Short clarifications — positioning and proof live in the sections above.
At what stage of a deal is this most useful?+
At sourcing it surfaces omission and short-list risk. At early screening it surfaces positioning splits and risk language. At deeper diligence the full battery plus narrative review of the data-room materials adds the most complete picture. All stages benefit from recording the model distribution.
Does this replace human diligence or the data room?+
No. It makes the AI layer that is already influencing human attention measurable and reviewable. The data room and human analysis remain primary.
What does a 100-point or larger split practically mean?+
The same company is being assessed as belonging to different quality or risk tiers depending on which model was used. That is material information for any process that currently relies on a single assistant for first-pass research.
Can we run narrative review on a confidential deck?+
The module is designed to accept the document the deciding audience would receive. Handling of confidential materials follows the same controls the firm already applies to other diligence tools.
How stable are the scores across repeated runs?+
Four of five judge models show individual stability well below the typical between-model range. The retrieval-grounded model can exhibit high volatility on thinly cited firms; that volatility is itself a signal of evidence thinness.
Is the output suitable for retention in the deal file?+
Yes. Assessments are timestamped, methodology-versioned, and include component decomposition, per-model scores, and coded objections with transcripts. The perception state at the time of screening can be reconstructed later.