Woven in, not bolted on: our philosophy of AI in healthcare

Health software is acquiring AI the same way office software did: a sparkle icon appears in the corner, a chat panel slides out, and the feature list grows by one. The product underneath is unchanged. The AI can talk about your work; it cannot do any of it.

We think that is the wrong shape for clinical software, and Hanah is built the other way around. There is one intelligence in the product — the same one that listens to your consult — and it is threaded through every surface: the brief before the visit, the note during it, the documents after it, the plan the patient takes home, and the follow-up in between. Not a chatbot beside the product. The connective tissue of the product.

An assistant with hands

The test of an assistant is not whether it can answer a question. It is whether it can finish the job.

Ask a bolted-on chatbot to write a referral and you get a wall of text to copy, paste, and reformat. Ask Hanah's assistant and you get the referral — drafted into a real document, on your letterhead, attached to the right patient, ready for your review. The assistant has the same hands the rest of the product has, because it is the rest of the product:

• Before the visit, it preps you: last visit's plan, open problems, recent results, the questions worth asking this time — pulled from the patient's actual history, including their practice-system record, not from whatever happens to be pasted into a chat.
• During the visit, it is the scribe.
• After the visit, it drafts the note, the referral, the insurer form, the exercise program, the plain-language summary — from one understanding of one consult, not five prompts into five tools.
• Between visits, it keeps the living documents alive: when a later session changes the plan, the care plan updates rather than going stale.
• Around all of it, it does the meta-work too — ask it to build you a new note template or redesign a document, and it works on the template the same way it works on a note.

None of that is possible for an AI that lives in a side panel, because a side panel has no context and no hands. Weaving the assistant in is not a design flourish. It is the difference between an AI that discusses work and an AI that does it.

It speaks product, not chatbot

Here is a small detail that carries the whole philosophy. Ask the assistant about "Sarah" when you treat three Sarahs, and it does not reply with a paragraph asking you to clarify. The same patient picker you use everywhere else in Hanah appears, you tap the right one, and the assistant carries on.

Small, but load-bearing: when the AI needs something from you, it reaches for the product's real controls — pickers, documents, programs — not more prose. Text is a terrible interface for clinical work. Interfaces are the interface. The AI should meet you in them.

The clinician signs. Always.

An AI woven this deep raises an obvious question: what stops it from quietly filing something wrong? The answer is structural, not aspirational — nothing the assistant produces lands anywhere without a clinician's sign-off. Every note, every document, every program is a draft until you finalise it. The assistant has hands, but you hold the pen.

That is also why we are conservative about the flip side: the AI never trains on your patient data, and your data never leaves your region.

Accountability is the licence to be everywhere

If the same intelligence is threaded through the whole product, then its error rate is not one feature's problem — it is the product question. Which is why we publish a standing accuracy and hallucination report: our methodology, our numbers, our worst results next to our best, updated as the system changes. We think any vendor weaving AI into clinical work owes you that. An AI you cannot audit is an AI you should not let near the record — however good the demo.

Three questions to ask any vendor

Including us. When someone shows you AI in a clinical product, ask:

• Can it see? Does it work from the patient's real record and history, or only from what you paste into a box?
• Can it act? Does it finish the job inside the product — the actual document, the actual program, the actual template — or does it hand you text to carry the last mile yourself?
• Who signs? Is every output a draft until a clinician approves it, and does the vendor publish error rates you can hold them to?

A sparkle icon can't answer those. A product built around the AI can.

That is the philosophy in one line: not AI as a feature of the software, but software as the body of the AI — with a clinician's hand on every decision that matters. It is why the scribe stays after the note, and why the assistant is everywhere rather than somewhere.