How Lead Forge's AI analysis reads a newly registered company

What the model is given, what it is told it does not know, how every sentence is labelled, and why the analysis is cached. A look inside the brief.

LF

Krapton Research

Lead Forge team · Product · · 3 min read

The most common question we get about Lead Forge is some version of 'how does the AI know that?' The honest answer is that it mostly does not know; it reads. This post explains exactly what the analysis is given, how it is asked to reason, how its output is labelled and where its limits are, so you can use it with the right amount of trust.

What the model is given

For each company we build a compact context from the public Companies House record: name, number, status, type, incorporation date and the company's age in days, jurisdiction, registered office and locality, SIC codes with their descriptions, the accounts and confirmation statement due dates, registry flags such as whether the company has charges or insolvency history, previous names, and up to twelve officers, twelve PSCs and twenty recent filings. Today's date is included so the model can reason about timing.

Just as important is what it is told it does not have. The context ends with an explicit list: no website, email or phone details; no financial figures; no headcount, funding or technology stack; no job postings; no social profiles. The model is instructed never to state or guess any of these. A brief that mentioned a website would be a bug, and the prompt is versioned so that briefs written under older instructions are regenerated rather than served.

What it writes

The output is a structured document rather than free text, validated against a schema before it is shown. It contains an executive brief, an activity read with a headline and detail, a why-now list, director intelligence covering footprint and ownership, a network insight, company signals, opportunity signals rated high, medium or low, an opportunity summary, a hiring signal with its confidence, potential needs, outreach angles and suggested questions. Each signal and need carries a basis.

The labels

LabelMeaningHow to treat it
PUBLIC DATATaken directly from the Companies House recordAs reliable as the filing itself
AI INTERPRETATIONInferred from the record by the modelA reasonable reading to verify on the first call
UNKNOWNThe record does not sayDo not assume; ask

The labels appear on the profile next to each statement, in the dark AI Analysis panel and in the exported brief. They are the whole point of the design: the analysis is useful because it is allowed to infer, and trustworthy because it says when it is inferring.

Ask AI

The same context powers the assistant on each profile. It answers questions about the company from the record only, keeps the conversation history for follow-ups, and is instructed to report anything the record does not hold as unavailable rather than guess. Every visitor can ask two questions per company; Lead Forge Plus removes the limit.

Caching and regeneration

A generated brief is stored against the company for up to 30 days. This keeps the analysis consistent for everyone who opens the profile, avoids running the model on every page view and keeps the product free. The Regenerate button re-runs the analysis on demand, and a brief is regenerated automatically if the prompt context has changed since it was written.

Limits, plainly

  • The model reads a filing, not a business. A SIC code is intent, not proof of trading, and the analysis says so.
  • Brand-new companies have thin records. Officer and PSC data can take a day to complete, so a brief written on day one may be regenerated on day two.
  • Interpretation can be wrong. That is why it is labelled. Verify before you rely on it, and tell us when it misleads.
  • The analysis is not advice. It is a research aid for deciding who to talk to and what to say first.

QUICK ANSWERS

Questions this article answers

Does the AI analysis use data from outside Companies House?
No. The model is given the public register data for the company and an explicit list of what is not available. It does not browse the web or receive website or contact details.
Why is a brief sometimes a few weeks old?
Briefs are cached for up to 30 days so every visitor sees the same analysis and the model is not re-run for every page view. Anyone can regenerate the brief from the profile.
  • ai analysis
  • product
  • gemini
  • new companies

Published · Updated

LF

WRITTEN BY

Krapton Research

Krapton is a software development and IT consulting company headquartered in New Delhi, India. We design, build and ship web platforms, mobile apps and AI products for founders and teams across the US, UK, UAE, India and beyond — and we build our own products too. Lead Forge is the team's own product for finding and understanding newly registered UK companies.

About Lead Forge →

PUT IT TO WORK

See the register the way this article describes it

Browse the companies incorporated this week, open one, and read its AI brief with every sentence labelled public data or interpretation. Free with a Lead Forge account.

MORE FROM LEAD FORGE