Meeting prep with a verification layer

Before a first meeting with a company, someone senior wants one page: who they are, what's moving, what to ask. The easy version — ask an AI to summarise the company — produces something worse than nothing: a confident narrative in which a two-year-old disposal reads as last week's news and a single unconfirmed press rumour reads as a done deal. The reader can't see which sentence to trust, so they either trust all of it or none of it.

What it does

The workflow takes a company (its website as the anchor), the meeting context, and the angle the meeting is being taken from. It resolves the company to a single canonical identity, gathers from public sources, and then does the part that matters: every claim goes through a registry before it can appear in the brief, labelled on two separate axes.

The first axis is is it true — verified (primary source, or two independent sources agree), partly verified, inferred (with the reasoning shown), or unverified. The second axis is what it means for us — evidence-backed, hypothesis, or open question. The two never collapse into one score, because a fact can be fully verified and still mean nothing, and a rumour can be unverifiable and still be the most important thing to ask about in the room.

The output is a one-to-two-page brief: bottom line, verified signals, triggers and opportunities with confidence attached, the questions worth asking, a source appendix — and a "what we could not find" section, because absence of evidence is a finding too.

The demo: a FTSE-listed company, July 2026

To show the layer working against a live public record, I ran it on British Land plc — public sources only, deliberately at company level (the workflow's privacy rule keeps runs involving named individuals off any public surface). Three moments in that run are the whole argument:

• A "stock drops on stake sale" story surfaced undated in current search results. The layer chased the date: it was the Meadowhall disposal, completed July 2024 — registered as historical and excluded from the current picture, where a plain summary would have reported it as news.
• A reported £700m development sale was surfaced by exactly one trade-press article, with no company confirmation. It stayed unverified — and became the meeting's best open question instead of a fake fact.
• The company's own "some of the UK's best and most sustainable real estate" stayed labelled as marketing self-description. A company's website is a primary source for what it is, and no source at all for how good it is.

The full brief — claim registry, labels, sources, and the "could not find" list — is the demonstration artifact.

Honest scope

This is an internal preparation workflow, not a product: it drafts, a human reviews, nothing is sent or exported autonomously. It reads public sources only — no logins, no paywalled personal data, and the fetched pages are treated as untrusted data, never as instructions. On earlier real-input runs the same layer caught a misspelt company resolving to the wrong firm, a three-way name ambiguity, and two wrong job titles — and refused to invent a person it could not resolve. That refusal path is the point: the tool is only useful because of what it declines to say.