Archives

Engineering teams have shown how to build company knowledge bases. Finance needs a different animal. Seven organs, what each one is for, and how I built them.

The entire pitch for what I built, asterisk included: why the board pack questions that cost an afternoon stop getting asked, what changed this year, and how I'd test any AI that says it's fixed that, including mine.

Is it safe to put company data in ChatGPT? A working protocol for finance teams: what to paste where, how to anonymise a question properly, and how to get IT to yes.

I published an exam for finance AI: 34 realistic documents, 25 questions, a deterministic grader, answer key included. The frontier models got most answers right. Proving them right was a different story.

It's the question every finance leader asks, and it deserves a straight answer, including the cases where a chatbot genuinely is the right tool.

The four ways an AI agent can search your files (keyword, vector, file system, graph), each mapped to how you already find things, and why spreadsheets break the most popular one.

The word finance people hear in every vendor call, explained through their own day: how actuals move between NetSuite, Anaplan, SAP and Workday, what a key is, and why IT's questions are your own posting-access discipline.

What happens to our runway if the Phase 3 runs a year long? How Grounded's advisory mode answers a board-level question in three tiers: facts, assumptions, analysis.

Two of your own documents disagree about R&D spend by $2.1 million. Which one is right, and how long would it take your team to find out? A walkthrough on a synthetic biotech reporting pack.

Variance analysis has four steps. Only one of them is analysis. A look at where an FP&A analyst's time actually goes, and which steps a machine should be doing.

Every finance team knows the out-of-nowhere question that lands at the worst moment. It isn't an interruption. It's a tax. Here's what it actually costs, and why you'll never get headcount to fix it.

Provenance and honest abstention aren't features you bolt on at the end. Build them in from the start, or you've built AI a controller can't defend.

Grounded is an audit-grade retrieval layer for financial documents: version-correct, sourced to the exact figure, and honest when it can't answer. Here's why we built it the way we did.