By Aram & the Algorithms.
The Smallest Model That Doesn’t Lie
Organizations adopting artificial intelligence are understandably reaching for Occam’s razor. Strip away the theatrical claims and the speculative futures, and ask what remains: which task improves, by how much, at what cost, and who is accountable for the result.
This discipline is valuable. The trouble begins when the razor is applied not only to the explanation, but to the world being explained.
“AI adoption” sounds like a single phenomenon until the same technology enters a pharmacy, a law firm, a classroom, a design studio, a hospital, an electrical grid. The model may be general-purpose; the work is not. Each setting organizes knowledge, judgment, liability, apprenticeship, trust, and consequence differently. A simplified business case can be perfectly legible while describing the wrong unit of reality.
The question is not whether to simplify. The question is how far we can simplify before the model begins to lie.
Borrowing from Ashby, Carefully
We take our name from the British cybernetician W. Ross Ashby, whose 1956 An Introduction to Cybernetics set out the Law of Requisite Variety: a regulator responsible for a complex environment must have enough possible responses to meet the meaningful variety of disturbances it faces. In Ashby’s phrase, only variety can destroy variety.
We should be precise about what we’re taking and what we’re not. Ashby’s law describes a controller’s repertoire of action — how many distinct responses it can produce. Our accordion is a discipline for a different problem: how many distinct categories of reality a model must preserve before it stops predicting the right intervention. These aren’t the same claim, though they rhyme. Requisite variety says the thermostat needs more than one setting if the room faces more than one kind of temperature swing. Our accordion says the business case needs more than one description of “AI adoption” if the hospital contains more than one kind of work. We use Ashby’s name because his instinct — that a model can be simple and still be catastrophically undersized for what it governs — is the instinct we’re borrowing. The formalism is his; the application here is looser than the original theorem, and we’d rather say that than let the citation imply a proof it isn’t carrying.
Opening and Closing the Frame
Ashby’s Accordion is our name for the practical instrument that follows from that instinct. It expands and contracts the frame of inquiry until we arrive at the smallest actionable model that still contains the differences capable of changing the intervention.
We might begin with an industry: healthcare. That frame is too broad to determine what responsible adoption looks like, so the accordion opens — from sector to institution, institution to function, function to workflow, workflow to role, role to task, task to consequence.
The point is not to keep expanding until every situation appears infinitely unique. Context can become its own form of avoidance — a fog thick enough that no decision can be made because everything depends on everything else. The accordion has to close as well as open, and closing is the harder discipline, because nothing forces it. You can keep adding nuance forever; only judgment tells you when the nuance has stopped changing what you’d do.
Where It Closes: A Worked Pass
Take one hospital. Scheduling, diagnosis, pharmaceutical quality assurance, and bedside care all sit inside it, and “AI adoption” cannot mean the same thing in all four.
Scheduling is close to codified production: known inputs, known constraints, a correct answer that doesn’t depend on who’s asking. An AI system that optimizes it is accountable the way a spreadsheet is accountable — you can audit the output against the constraints and know if it’s wrong. The accordion can stay closed here; almost nothing about the rest of the hospital changes what a good scheduling intervention looks like.
Diagnosis is different in a way that matters for the decision, not just in a way that’s interesting. Diagnostic judgment is built partly through repetition — a resident who has seen a hundred ambiguous chest X-rays recognizes the hundred-and-first differently than a resident who has seen ten. If an AI system absorbs the routine cases, it may make today’s diagnoses faster while quietly removing the repetitions through which tomorrow’s diagnosticians learn to see. That is a fact about apprenticeship, not about model accuracy, and a business case scoped at “diagnostic support tool” will never surface it. Here the accordion has to open past the workflow and into the training pipeline underneath it, because that’s where the actual cost shows up.
Bedside care opens further still, because a meaningful share of its value is presence itself — the fact that a person came, stayed, and paid attention — which doesn’t show up as a task at all and can’t be error-corrected the way a scheduling mistake can.
So the frame doesn’t close at one consistent depth across the hospital. It closes at the depth where adding more detail stops changing which intervention you’d choose — and that depth is different for scheduling, diagnosis, and bedside care, which is the whole reason “hospital AI adoption” was too coarse a unit to begin with.
What the Accordion Protects
This movement between expansion and contraction runs through our broader process. Forensic Systems Cartography opens the accordion by reconstructing how the present arrangement actually works — tracing where authority sits, where labor goes unrecorded, and which of the institution’s histories are still shaping what happens now. The idea of a task ecology helps locate the meaningful unit of work inside that larger reconstruction. Intervention Design Thinking then closes the accordion around a bounded move: something that can be tried, observed, corrected, or stopped.
The instrument earns its keep on workforce questions specifically, because those are where a coarse frame does the most damage. A task that looks inefficient may be producing institutional memory. A junior assignment that looks automatable may be one of the repetitions through which someone becomes senior. A drafting tool that saves an hour may spend forty minutes of someone else’s time on review and repair — time that doesn’t appear in the productivity calculation because nobody scoped the frame wide enough to include the person absorbing it.
Seen this way, the governing question stops being what will AI do to the workforce and becomes what does this intervention do to this work, here — what it compresses, what it transfers, what it removes, what new thing it requires someone to now be good at. A human “kept in the loop” is not the same as a human who has the evidence, the standing, and the time to actually change the outcome; the accordion is what lets you tell the two apart before the difference gets treated as friction and designed away.
Occam’s razor still belongs on the table. It removes unnecessary assumptions and exposes hollow claims. Ashby’s Accordion is doing something else: making sure that in the process of simplifying the story, we haven’t also simplified away the part of the world the story was supposed to be about.
Leave a Reply