Most AI reading of a public company is directionally right. A boardroom needs more than direction — it needs facts it can act on, and a way to check them. This page is how BoardBrain earns that trust: by construction, not by promise.
Every published campaign of the leading activist funds, reconstructed and priced against each company’s own market. Every campaign jumps on day one, whatever its quality. Twelve months later the top quartile stands +34%; the bottom quartile, −38%. A 72-point spread, invisible at the start. Method separated them, not aggression.
Precise, verifiable evidence of where value is trapped — a per-share number that survives diligence.
Management’s own record: promised versus delivered, dated and quoted. The one exhibit no board rebuts.
Asks a board can actually implement — and the market pays again, each time one is executed.
Assertion in place of proof — campaigns that only claim trailed their market by 22 points a year on.
Aggression as a substitute for a case: attacks, ultimatums, the proxy fight itself — zero premium.
The board-seat contest as the goal. The market pays for the action, never for the fight.
BoardBrain is trained on the reasoning of the campaigns that created value, not on activism in general. Rebuilt blind from only the facts available the day before each campaign became public, the engine reproduced the diagnosis, the lever and the ask of 82% of the value-creating campaigns, invented nothing, and declined every company that was never a target.
Leading activist investors, 2002–2026, market-adjusted. Portfolio-level results, not predictions for any security. Full methodology available under NDA.
Every fact, quote, number and name in a BoardBrain file is extracted and verified first, and locked. The language model is used only to interpret and narrate around those locked facts — it never generates them. A missing fact fails the build; it is never invented.
A general assistant uses the model far more broadly — to find, to summarise, to phrase — which lets it take shortcuts and produce fluent text that can be confidently wrong. We removed that exposure by design. That is why BoardBrain does not hallucinate on facts where a general assistant can: every figure is locked to a source before the model writes a word, so it cannot invent a number it was never given. Interpretation is judgement — and is reviewed and signed off by a person.
Extract. The fact is pulled from the filing, transcript or note — with its exact location kept.
Verify. Checked against the source, and against the company’s own record and its peers.
Lock. The verified fact becomes immutable. The model can reference it, never rewrite it.
Narrate. Only now does the model write — around the locked facts, citing each one.
Click any figure in the product and the source opens: the filing line, the transcript page, the note.
Behind each company file sit thousands of documents — the company and every peer, roughly a decade deep — assembled in three tiers. The sector is pre-digested before the company is read, so the file knows best-in-class economics and exactly where the company sits against them.
Filings across every exchange the company touches; earnings wires, transcripts and capital-markets-day material; the trade press.
Licensed sell-side research, premium press, industry and regulatory material, specialist pricing data — paid sources a general assistant cannot access at all.
The activist and precedent library, the management-commitment database, the theme-classification grid, the ownership and control map — built by us, per company.
That reading is paid for once, up front. A general model must keep the work behind each answer small, because every query is priced on demand — so it answers from a thin slice. BoardBrain answers every question from the full, verified asset, in seconds.
A European carmaker, from a live BoardBrain file, anonymised. Management’s story, against the record — and the disclosures the board never received.
No regional margin split — although the two closest premium peers disclose it.
No electric-vs-combustion profitability — although the pure-play leaders report it quarterly.
No capital-return policy, despite substantial and recurring free cash flow.
No platform economics and no return-on-investment for the autonomous-driving programme.
The director walks in knowing what was recast, what was missed, and what was never disclosed — and asks the question that follows. In product, every line above carries its source.
Worked example from a live file, anonymised for publication. Figures shown in product are source-traced.
We put identical board questions to a general assistant and to BoardBrain, and scored the answers on the dimensions a boardroom actually needs: fiduciary tension, accountability framing, governance depth, decision-grade questions. The gap is 3–6 points where it counts.
Our own benchmark, on worked examples — which is why we’d rather you ran the test yourself: name a company you know, and judge the file against your own knowledge of it.
Run the testThe method — and the files it produces — are reviewed by a senior circle: former investors, equity-research leadership from a leading investment bank, and practitioners from the strategy and activist worlds. Nothing ships unread: every file is challenged and signed off by a senior reviewer before a director sees it.
The system is deliberately constrained to what the evidence supports: a general model is built to please and to be fluent; this one is built to be checkable.
Nothing is asserted that cannot be shown. If a claim on this site reads confident, it is because the file behind it can be opened.
Can every number be traced to a source you can open?
Was the company read against its full peer set, or in isolation?
Are management’s past commitments scored against outcomes?
Does it end in questions a board can actually put to the CEO?
Who is accountable for it being right — a person, or a prompt?
Apply it to anything you are handed — including us.
Name it. Judge the file against twenty years of your own knowledge.
Request a briefing