Start with the work, not the technology
Inora’s first architecture starts with the work of care, not with a language model. It fixes three things: a closed space for each organisation, every document with its structure and authority, and, with every question, the role of the person asking it.
Written by
Founder and CEO
Essay6 min read
Anyone who expects an architecture for AI expects a model at the top: which language model, how large, from which vendor. That is not where the architecture I am sharing today starts. It starts with the work: with what a senior care assistant, a nurse or a team leader does on a shift, in what order, and at which moment they need to know something they do not have at their fingertips.
I will repeat one sentence from the document here, as I wrote it: “The real value comes from understanding nursing workflows deeply and building interfaces that feel like natural extensions of how nurses think and work.” The sentence before it is shorter. Technology is only a means.
Why not the model
A language model can be bought. Anyone who starts tomorrow can use the same model we do, and I expect today’s model to be overtaken within a year. Build a system around one model and you rebuild it the moment that model changes.
What cannot be bought is knowing how a question arises in care, what is needed at that moment, and what an answer has to carry to be of use on the ward. That is where the work is, and that is where the value is.
That is why the plan for the foundation, which I shared on 13 June, sets down that Inora is tied to no single language model. The model is a part we can replace. The same plan set down two more things: Inora runs in an environment we control ourselves, and every answer carries its source. Those three come before any feature. Today’s architecture builds on them.
How a question passes through Inora
In this architecture a question does not go straight to a model. First comes the context: who is asking, from which organisation, in what role. Then: what kind of question is this. Then: which of this organisation’s protocols belong to it. Only then does it follow the route that fits the task, and the answer comes back with its sources.
The model does its work inside that route. It does not choose the route.
That is what I mean by starting with the work. Senior care assistants who want to know how their organisation wants a handover done gain nothing from a general account of handovers. They need their own organisation’s answer, in the order of their own protocol, with the place where it is written. A system that starts with the model gives them the general account, fluently worded. A system that starts with the work gives them what their organisation has agreed.
Three things that are fixed
Three things follow from that starting point. They are fixed, whichever model we use and whatever we build on top later.
Each organisation a closed space of its own. A care organisation’s protocols are its own. A question asked in one organisation never gets an answer from another organisation’s documents. The separation is strict. At the same time, no organisation should have to reinvent the profession: knowledge that holds for everyone can be shared. What is shared is shared on purpose, never because it happens to sit in the same database.
Every document with its structure and its authority. A protocol is not a pile of loose sentences. It has chapters, steps in a fixed order, references to other documents and a version. And it has authority: a national guideline carries different weight from one ward’s work instruction, and that in turn from a leaflet.
Cut documents into loose fragments and retrieve them by how closely the words match, and you throw away exactly what a care worker needs: which step comes first, what it refers to and how much weight it carries. That is why Inora first turns every document into structured text that keeps its structure, its references, its version and its authority, before a single question is asked of it. The document calls this the knowledge translation layer. A source’s authority belongs in it on purpose, because being trusted by nurses is one of the aims of this architecture, and that trust begins with knowing what an answer rests on.
With every question, the role of the person asking. The same question is not the same question when someone else asks it. A senior care assistant, a nurse, a team leader and an administrator can ask the same thing word for word and still need something different: a different depth, a different part of the protocol, a different responsibility. So every question carries who is asking, from which organisation and in what role.
An answer that does not know the role is exactly right for no one.
What else the work asks for
Two smaller things in the architecture come from the same place.
In care, hands are rarely free. So the architecture takes speech as input, alongside typing.
And we measure what Inora returns in the work itself: how much time it saves, what kinds of question come in, and whether people keep using it. A demonstration proves little. What counts is what happens when people use Inora on an ordinary shift, and whether they come back the next day. We measure that from the start, not afterwards.
AI can make people’s work stronger, but only when it is used the right way. You cannot see that from a distance. Stay close to the work and you measure what improves, not how many people completed a training session.
Where the work ends and judgement begins
Starting with the work also means knowing what stays with the professional. Nurses who know what the protocol asks still decide for themselves. Giving knowledge is not taking a decision.
Inora does not diagnose, calculate a dose or decide urgency. It brings the source to the person who decides. The decision about a client is taken by the professional who is responsible for it. Software cannot answer for anything, so software does not take that decision.
A system that forgets that line does not fit the work. It tries to take it over.
What this means for what we build
What Inora has to become is a care organisation’s knowledge system: the place that holds what the organisation knows and has agreed, current and within reach of everyone who works there, from the care assistant at the bedside to the board. You do not build that by hanging a language model on a folder of PDFs. You build it by first understanding how the work is done, and shaping the technology to it.
The document runs to seven pages. We write down what we decide, with the date and the name of the person who wrote it, before we build it. Anyone who later wants to know why Inora does what it does can read it back here, with the day it was written.
We can replace a model. We cannot replace how the work is done.