The question every clinical-stage board eventually asks: what happens to our runway if the Phase 3 runs twelve months longer than planned?
It’s not a lookup. It’s analysis. And analysis is where AI tools usually get away with murder, because a fluent paragraph with a number in it looks like an answer whether or not the number is defensible.
We think the fix is structural. Grounded’s advisory mode answers questions like this in three tiers that never blur into each other, and I’ll show it on Super Fancy Bio, our fictional synthetic biotech (all 24 documents built by us, planted problems included).
Tier one: facts. Every figure cited to its source document and cell. For the runway question: $717.2 million of cash at the end of June, average net burn of $15.8 million a month, baseline runway of roughly 45 months. Click any of them and the source opens.
Tier two: assumptions. Listed explicitly, with their basis. Which programme’s run-rate drives the scenario. What was held constant. What was estimated because the documents don’t state it. This is the tier most analysis hides, and it’s the tier that decides whether the conclusion survives contact with a sceptical board member.
Tier three: analysis. Clearly framed as opinion derived only from tiers one and two. In this scenario, a twelve-month delay to the Phase 3 compresses runway to roughly 31 months.
Why the ceremony? Because of what happened in one of our test runs. The scenario’s burn assumption was derived from two monthly reports, one of which carried a planted $2.1 million accrual error. The assumption came out at $19.2 million a month when the true figure was $17.1 million. Reading the files faithfully meant inheriting their mistake.
With the assumption buried, that’s an invisible error compounding into a board decision. With the assumption listed in tier two, it’s a line someone can point at and say “check that one.” Same underlying data, completely different risk profile.
That’s the standard I think finance teams should hold any AI analysis to: not “is it smart,” but “can I see exactly what it assumed, and check every fact it built on?” If the answer is no, it’s not analysis. It’s an opinion with good formatting.
Super Fancy Bio is fictional and the numbers are synthetic, but the failure mode is not. If you’ve seen an assumption error survive all the way into a board deck, you know exactly why tier two exists.