First, the disclosure: Super Fancy Bio is fictional. We built a full synthetic biotech reporting pack, 24 documents, to test Grounded on the problems real packs have, and we planted the errors ourselves. Everything below is reproducible and none of it is a client’s data.
Here’s the scenario we planted, and every biotech finance person will recognise it.
Super Fancy Bio’s May monthly P&L shows year-to-date R&D expense of $236.5 million. The R&D spend by programme report, the one that goes in the board pack, totals $234.4 million for the same period. They are $2.1 million apart. Both documents came from the same finance team in the same close. Neither one announces the disagreement.
The cause is mundane, which is what makes it dangerous. A late CRO accrual for the Phase 3 programme landed in time for the P&L close but never made it into the programme report. No one did anything wrong, exactly. The two documents just stopped agreeing, quietly, in one month out of six.
Asked whether the May report ties to the P&L, Grounded says no, shows $236.5 million against $234.4 million, names the programme where the difference sits, and cites both cells so you can check it in seconds. Asked the same question about June, where the accrual has posted and the files genuinely tie at $284.9 million, it says they tie.
That second answer matters as much as the first. A tool that finds gaps is only useful if it doesn’t invent them. The failure mode of generic AI on this question isn’t missing the $2.1 million, it’s confidently reporting a discrepancy that doesn’t exist, and sending your team on a two-day hunt for nothing.
There’s a sting in the tail. Weeks later, a runway scenario built on this same pack derived the Phase 3 programme’s monthly run-rate from the May and June reports, and inherited the error: $19.2 million a month instead of the true $17.1 million. One quiet accrual, first a reconciliation miss, then a corrupted forecast assumption. Errors in reporting estates don’t stay where they were born.
How long would this take your team to find by hand? That’s a genuine question, and I’d like to hear the answer.