Building a chatbot on FDA Guidance is easy. Scaling it to everything else you need isn't.
Internal Claude bots and “we uploaded a folder of approvals” projects are everywhere now, and the instinct is right: this work *should* be AI-assisted. Rhizome is what that project becomes when it has to be trusted on a >$100m decision: a focused, AI regulatory coworker on a maintained, exhaustive corpus, still pluggable into the general AI you already use.
How they differ
Why the build is harder than the demo. The day-two demo always works: load some PDFs, ask a question, get a decent answer. The reasons it degrades from there are structural, and they're the same reasons general AI struggles with this domain. The source databases — drugs@fda, guidance libraries, clinicaltrials.gov and its protocols, standards, enforcement records — aren't machine-searchable; much of the substance sits in image-only PDFs that need real OCR and understanding beyond a vector-store upload. New documents land every week across dozens of authorities. And nobody budgets for evaluation: how do you know your bot didn't miss the one predicate that matters?
That's not a weekend project. It's our entire company. We maintain >75 health authority databases across >12 markets (as of August 2026 — on pace for >150 databases and >25 markets by end of year), every line ingested, every image processed. >45 million documents, refreshed on an ongoing cadence, with the refresh being someone's actual job.
Coverage gaps you can't see. The dangerous thing about a snapshot corpus isn't what it's missing — it's that it answers confidently without what it's missing. Ask your internal bot about MHRA or PMDA activity since last quarter's upload and it will answer from last quarter's upload. An incomplete answer that looks complete is worse than no answer when the decision is big enough.
Nothing from memory, provably. Getting an agent to rely only on documents — and never quietly fall back to model memory — is an engineering and evaluation discipline of its own. We push our agents to substantiate nothing from memory and explicitly benchmark them for it, and every claim cites the exact passage so you can check. Internal builds vary from “pretty good” to “nobody has measured,” and with respect: if nobody has measured, it's not ready for a filing argument.
The real cost line. In-house feels free because the model API is cheap. The corpus maintenance, the OCR pipeline, the refresh jobs, the evaluation harness, the citation UX: that's engineering salaries, forever, for a system that's someone's side project. Rhizome is published software pricing. Do the buy-vs-build math on the run cost, not the demo.
Where building wins. If it's key to your differentiation as a company. An agent that assesses how likely a drug candidate is to work is core to a biotech's IP. So are your fine-tuned models for drug design. Very few companies aim to differentiate on regulatory, so it rarely makes sense to invest key R&D AI staff here.