Why finance AI needs verified private company data

An LLM can read faster than an analyst. It can summarize filings, compare descriptions and draft a screening memo. But it cannot turn an unverified company list into verified evidence. If the retrieval layer contains the wrong legal entity, a registered office presented as an operating location, or modelled revenue presented as filed revenue, the answer can be fluent and still be wrong.

That distinction matters in finance because the error rarely stays in one field. A bad identity match can contaminate a valuation, a buyer list, a credit review and every memo generated from the same record.

The language model is not the source of truth

The US National Institute of Standards and Technology describes generative-AI confabulation as confidently presented false or erroneous content. Retrieval does not remove that risk by itself. It changes the question from “what does the model remember?” to “what evidence did the system give the model?”

For private-company research, the useful architecture has three separate layers:

1. Evidence layer: source documents and fields tied to a legal entity, source and date.
2. Reasoning layer: an LLM that extracts, compares, summarizes or flags contradictions.
3. Decision layer: an analyst who decides whether the evidence is sufficient for the task.

Collapsing those layers encourages the model to fill gaps. Keeping them separate lets it say that a fact is supported, contradicted or not publicly available.

Four data errors that change the answer

Weak input	What the AI may conclude	What the evidence should preserve
Company name without a registry identifier	Facts from two similarly named entities belong to one business	Legal name, company number, jurisdiction and former names
“Active” registry status	The company is currently trading	Legal status plus a separate, recent operating signal
One generic address field	The company operates from its accountant’s office	Registered office and trading address as separate fields
Revenue without a source type or period	The figure is current, filed and comparable	Filed or modelled status, currency, period end and source

These are not cosmetic data-cleaning issues. Imagine an acquisition screen where a model is asked to find regional operators above a revenue threshold. If the location is a registered office and the revenue is an estimate attached to the wrong entity, the model can satisfy the prompt perfectly while returning the wrong target.

Entity resolution comes before analysis

A private company can have a legal name, a trading name, former names and several related legal entities. The website may use the group brand while accounts are filed by a subsidiary. A press release may name the brand and omit the company number.

Before an AI system compares companies, its record should answer:

• Which jurisdiction and legal identifier anchor this entity?
• Is the website operated by that entity or by another group company?
• Is the name a legal name, trading name or former name?
• Which entity files the financial statements being used?
• Are parent, subsidiary and operating-company relationships evidenced or merely inferred?

Fuzzy name matching is useful for finding candidates. It is not evidence that two records describe the same company. The system should surface uncertain matches for review rather than silently joining them. For a step-by-step approach to building resolved records from registration numbers, names and group structure, see entity resolution for private company data.

A registry record is a starting point

Companies House makes company status, registered office, filing history and other statutory fields searchable. Its disclaimer also explains the limits of information on the register. A registry is authoritative for what was filed; that does not make every filing proof of current commercial activity.

The safer pattern is to keep two kinds of evidence:

• Structural evidence: incorporation, legal status, company number, legal form and filed address.
• Operating evidence: a current company website, recent accounts indicating activity, a regulator entry, a current tender, or another attributable sign that the business operates.

The model can then explain why a company passed the filter instead of turning “active” into a stronger claim than the source supports.

Provenance needs to travel with the field

A source list at the bottom of a report is not enough. The analyst needs to know which source supports which value. The Basel Committee’s data-aggregation principles were written for banking risk data, not private-company LLM research, but the underlying disciplines are useful here: accuracy, completeness, integrity and timeliness have to be designed into the data flow.

A decision-grade field should ideally retain:

• the value exactly as observed;
• the source URL or document identifier;
• the observation or filing date;
• the legal entity to which it applies;
• whether it was reported, calculated or inferred;
• the transformation applied, if any; and
• a reason when the field is blank.

This makes later corrections possible. If a company changes its name or a newer set of accounts arrives, the record can be updated without guessing where the old value came from. For how source dates, observation dates and filing periods differ in practice — and how different field types age at different rates — see how current is private company data and how would you tell.

Missing should remain missing

LLMs are very good at producing the sort of value that would plausibly fit a gap. That is precisely why a retrieval system must not invite them to do it.

An absent revenue figure does not mean zero revenue. An absent ownership field does not mean independent. No published trading address does not make the registered office the trading address. The clean response is “not established from the available public evidence.”

If modelling is appropriate, keep the output in a separate field and label it as an estimate with its method and range. Never overwrite the sourced field. Otherwise a downstream model, analyst or client cannot distinguish disclosure from prediction.

For a taxonomy of the different types of blank — not-filed, not-applicable and not-researched — and how each one should be encoded and handled in AI screening workflows, see what blank fields in a private company record actually mean.

A review checklist for AI-assisted company research

Before using an output in valuation, target screening or a client deliverable, check:

1. Every company is anchored to the correct jurisdiction and legal identifier.
2. Legal, trading and group names are not silently merged.
3. Registry status is not described as proof of operating activity.
4. Registered and trading addresses remain separate.
5. Financial values show their period, currency and whether they were filed or modelled.
6. Ownership and group relationships have attributable evidence.
7. Blank values have not been converted into zeros, negatives or invented facts.
8. Each material claim can be traced back to a source and date.

This does not make an LLM infallible. It makes its answer auditable and gives the analyst a way to find the point where an error entered the chain.

For an implementation-level workflow, see how to prepare private-company data for AI research, including the minimum entity key, field-level provenance model and explicit missing-data states.

SeekCompanies is building the evidence layer for private-company research: legal identity, operating evidence and public contact routes kept distinct rather than blended into a deceptively complete row. See the coverage methodology and current access path. For a deeper look at the underlying inputs, read where private company data actually comes from.

Frequently asked questions

Can an LLM verify whether a private company is operating?

Not from a company name or registry status alone. An LLM can organize the evidence supplied to it, but operating status normally requires a legal identifier, current registry status and a separate recent operating signal such as the company's own website or an attributable filing. Active incorporation is a legal status, not proof of trading.

What private-company data should an AI system cite?

Every field that can change a screening or valuation decision should retain its source and observation date. That includes legal identity, status, registered office, trading address, financial period, filed versus estimated figures, ownership evidence and the company's own activity description.

Should missing company data be estimated before it is given to an LLM?

Only when the use case explicitly permits modelling and the estimate remains visibly labelled with its method and uncertainty. For evidence-led research, an unsupported blank is safer than a plausible value because the model may repeat the estimate later as if it were a disclosed fact.

Sources

1. Artificial Intelligence Risk Management Framework — Generative Artificial Intelligence Profile — National Institute of Standards and Technology; accessed September 19, 2026
2. Companies House disclaimer — Government of the United Kingdom; accessed September 19, 2026
3. Searching the Companies House register — Government of the United Kingdom; accessed September 19, 2026
4. Principles for effective risk data aggregation and risk reporting — Basel Committee on Banking Supervision; accessed September 19, 2026