Auditing AI-Generated Company Research: A Review Checklist

Why AI research needs an audit step

A generated company profile reads the same whether it is right or wrong, which is why AI-assisted research needs a separate audit step rather than a quick read-through.

NIST's Generative AI Profile names the failure directly. Section 2.2 defines "confabulation" as a phenomenon in which generative AI systems "generate and confidently present erroneous or false content in response to prompts," and notes the behavior is colloquially called hallucination ^[1]. The same section warns that outputs "may also include confabulated logic or citations that purport to justify or explain the system's answer, which may further mislead humans into inappropriately trusting the system's output."

Two details in that section matter for deal teams:

• NIST says the dynamic is "particularly relevant when it comes to open-ended prompts for long-form responses and in domains which require highly contextual and/or domain expertise." A target profile is that kind of output.
• The profile describes automation bias as "excessive deference to automated systems" and says it can exacerbate confabulation risk ^[1]. A citation next to a sentence raises a reader's trust whether or not anyone has opened the source.

The audit step breaks that link: a claim is accepted only when a reviewer has confirmed the cited material says it, about the right company, as of a known date.

Failure modes in company research

The research literature does not use one definition of hallucination, and a reviewer should not either. Ji et al.'s survey calls the most inclusive definition "the generated content that is nonsensical or unfaithful to the provided source content," and splits it in two: intrinsic hallucinations, where output "contradicts the source content," and extrinsic hallucinations, where output "cannot be verified from the source content" ^[2].

For acquisition research, the useful categories are narrower and map to different review steps:

The misgrounded case deserves the most attention because it survives a quick check. Magesh et al., evaluating commercial AI legal research tools, define a response as "misgrounded if key factual propositions are cited but misinterpret the source or reference an inapplicable source," and count a response as hallucinated "if it is either incorrect or misgrounded" ^[3]. They argue these errors are "potentially more dangerous than fabricating a case outright, because they are subtler and more difficult to spot."

That study reports that the commercial legal research tools it tested "each hallucinate between 17% and 33% of the time" on its preregistered set of legal research questions (Magesh et al., arXiv 2024; the same figure appears in the published version, Journal of Empirical Legal Studies, 2025). Read that figure for what it is: a measurement of specific legal research products on legal queries. It is not a rate for company research, for any other vendor, or for any system you are reviewing. Its value here is the typology: a real citation can still be the wrong evidence.

Not every bad record is a hallucination: a faithful quote about the wrong company is an identity error, and the cause decides what to fix.

What NIST AI 600-1 suggests

NIST released the Generative AI Profile (NIST-AI-600-1) on July 26, 2024, as a companion to the AI Risk Management Framework, which NIST describes as "intended for voluntary use" ^[4]. The profile is a cross-sectoral profile that can be used "to govern, map, measure, and manage risks" ^[1], and its suggested actions carry IDs that follow those four functions (GV, MP, MS, MG).

The actions below are quoted verbatim from the PDF tables as read on 2026-10-28. NIST wrote them for generative AI generally; the right column is our mapping, not NIST's.

MS-2.5-003 is the core of this checklist. The phrase "ongoing monitoring" matters: a review done during a vendor trial does not cover a batch delivered months later, when sources, prompts or models may have changed.

Retrieval is not proof

Retrieval-augmented systems condition their answers on fetched documents. That moves the reviewer's problem rather than removing it.

The original retrieval-augmented generation paper framed the motivation plainly: for large pretrained models, "providing provenance for their decisions and updating their world knowledge remain open research problems," and the authors combined "pre-trained parametric and non-parametric memory" (in their case, a dense vector index of Wikipedia) to address it ^[5]. Retrieval gives the output a place it can point to. It does not guarantee the pointer is right.

Four things can still go wrong after a page is retrieved:

1. The page is about a different entity. Retrieval matches text, and two companies can share a name, a town or a product line.
2. The page does not say what the claim says. This is Magesh et al.'s misgrounded case: the source is real, the attribution is false ^[3].
3. The page is out of date. A cached or archived page can be faithfully quoted and still describe a company that has since been sold.
4. The claim adds to the page. Ji et al.'s extrinsic category: output that the source neither supports nor contradicts ^[2].

Magesh et al. describe what checking actually requires: users must "click through to cited references, read and understand the relevant sources, assess their authority, and compare them to the propositions the model seeks to support" ^[3]. Written about legal research, it describes the company research reviewer's job too.

The reviewer's workflow

Run the review at claim level, not record level. A record with mostly correct fields and one fabricated ownership claim is not "mostly right" for a screening decision; the ownership claim may be the only field that mattered.

Step 1. Split the output into atomic claims. One claim, one value, one citation. "Family-owned manufacturer in Ohio since 1994 with about 60 staff" (a fictional example) is four claims.

Step 2. Classify each claim. The class sets the evidence bar and the sampling tier.

Step 3. Open the cited source. Prefer the retained copy the system used over today's live page; if they differ, record both.

Step 4. Confirm the quote and the date. Find the quoted text verbatim. Record the page's own date, if any, and the date read.

Step 5. Check entity identity. Match the source's company to the target on legal name, address, domain or registry number. A match on trade name alone is not a match.

Step 6. Test support, not topic. Ask whether the quoted sentence, read alone, establishes the claim. A markets page listing "marine and industrial" does not establish "aerospace."

Step 7. Assign a disposition and a cause. Supported, narrowed, unsupported, contradicted, wrong entity, or open gap. Record the cause (confabulation, identity, stale, misgrounded, extrinsic addition) so errors can be tracked under MG-4.3-002-style logging.

Illustrative example (fictional). A generated profile of "Tamsin Ridge Coatings," a fictional company, reviewed claim by claim. Company, sources and values are invented.

Claim 4 alone would have changed the screening outcome, and it carried a working citation. Identity is checked before support for that reason.

Sampling and escalation

Whether to review every claim or a sample depends on what each claim decides. The arithmetic of sample sizes, acceptance numbers and confidence bounds is covered in Docket's audit sampling worksheet; this section covers how to tier the review and when to stop trusting a sample.

A practical tiering:

• Tier 1, review 100%: identifier, ownership and status claims that decide whether a company advances or is contacted. A wrong-entity error here can send outreach to the wrong owner.
• Tier 2, review all used in a decision: financial claims that drive a size screen. Financials shown for context only can be sampled.
• Tier 3, sample: narrative summaries, using a sample size and acceptance number written down before the review starts.

Stop rules should also be written in advance. Examples that do not require any assumed error rate:

• Any citation that does not exist, or a quote not found in its source, stops the sample. Treat the batch as unreviewed and move it to full review for that claim class.
• Any wrong-entity finding in Tier 1 triggers an identity re-check across the batch, because an identity error may come from a shared upstream step.
• Defects above the pre-set acceptance number expand the sample or reject the batch, per the worksheet.
• A change in model, prompt, source set or vendor resets the baseline. Earlier review results do not carry over (MS-2.5-001 and MS-2.5-003 above).

Do not turn review results into a system-wide accuracy percentage unless the sample was designed for it. A clean narrative sample says little about ownership claims, and a handful of spot checks is the "narrow, non-systematic, and anecdotal" assessment NIST warns against ^[1].

Sign-off and record keeping

The sign-off should let someone who was not in the room see what was checked, against what, and what remains open. The field-level structure for recording observations, conflicts and dispositions is laid out in Docket's claim ledger; the sign-off sits on top of it.

Sign-off template

Three record-keeping rules make the file useful later:

1. Keep the struck claims. A deleted error teaches nothing; a logged one shows which claim classes fail and why, the kind of record MG-4.3-002 describes.
2. Keep the source copies. Live pages change. A quote that cannot be reproduced from a retained copy is, for audit purposes, unsupported.
3. Keep gaps visible. An open "unknown" is a finding. Filling it with an inference to make the record look complete reintroduces the failure the review was meant to remove.

Limits of this checklist

This checklist is a working method, not an assurance standard, and passing it does not certify a file. It describes voluntary guidance and is not legal or compliance advice.

• NIST AI 600-1 is voluntary. The AI RMF is "intended for voluntary use," and NIST's framework page states: "The AI RMF 1.0 is being revised as part of the White House AI Action Plan" ^[4]. As of October 28, 2026, the AI 600-1 PDF on NIST's site is still the July 2024 profile and the six action IDs above match it; IDs and wording may change in a future revision, so re-check them before citing them in a policy.
• The literature is from adjacent domains. Ji et al. survey natural language generation broadly; Magesh et al. measured legal research tools on legal questions; Lewis et al. evaluated retrieval on general knowledge-intensive NLP tasks. None measured company research, and none of their figures transfers to a system you are reviewing.
• No accuracy claim is made for any system, including Docket's. The only defensible accuracy statement about a research batch comes from a sample you designed and reviewed yourself.
• Out of scope: adversarial testing of research agents (for example, target websites that carry instructions aimed at the agent) is covered in Docket's prompt injection test suite, and sample-size calculation in the sampling worksheet.

Docket runs this kind of evidence check on target lists before outreach: its Audit agent re-reads retained pages with fetch tools denied, strikes anything it cannot quote and leaves gaps open; the method is described in Docket's evidence standards.

Sources

1. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)" (July 2024). https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf. Accessed 2026-10-28.
2. National Institute of Standards and Technology, "AI Risk Management Framework" (web page, undated). https://www.nist.gov/itl/ai-risk-management-framework. Accessed 2026-10-28.
3. Ji, Z., Lee, N., Frieske, R., et al., "Survey of Hallucination in Natural Language Generation", ACM Computing Surveys 55(12), 1-38 (2023), https://doi.org/10.1145/3571730; quotations from the open arXiv version 2202.03629 v7 (July 14, 2024), https://arxiv.org/abs/2202.03629. Accessed 2026-10-28.
4. Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., Ho, D. E., "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools", Journal of Empirical Legal Studies 22(2), 216-242 (2025), https://doi.org/10.1111/jels.12413; quotations from the open arXiv version 2405.20362 (May 30, 2024), https://arxiv.org/abs/2405.20362. Accessed 2026-10-28.
5. Lewis, P., Perez, E., Piktus, A., et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (NeurIPS 2020; arXiv 2005.11401). https://arxiv.org/abs/2005.11401. Accessed 2026-10-28.