Symptoms and diagnosis

← Portfolio · ← Collaboration 2.0

1.1 · Symptoms and diagnosis

How a list of ordinary complaints turned into a causal graph, why the graph would not terminate, and what had to be done to it before anything could be designed against it.

Breakdown of symptoms and their causal relations

1.1.1 · The starting suspicion

The symptoms were unremarkable. Slipped schedules. Duplicated work. Friction at every integration point. Products that reached the market without anyone in the building able to say who the customer was. Anyone who has spent time in a technical organisation has seen all four, and each of them has a standard remedy that any competent manager can name.

What made them worth a second look was not their severity but their distribution. The same patterns kept recurring across teams that shared no members, no manager, and no codebase, and across projects that had nothing in common but the building they were run in. That is the observation the whole diagnosis rests on, so it is worth stating precisely:

Independent causes do not produce repeating patterns. A shared mechanism does.

If four teams with nothing in common fail in the same way, the explanation cannot be four independent local failures — that would be a coincidence of implausible size. Something upstream of all four is generating the pattern. And if that is true, then treating each symptom on its own terms is not merely inefficient; it is a category error. You would be repairing outputs while the thing producing them continues to run.

This also explains a fact that was otherwise puzzling: every obvious remedy had already been tried somewhere in the division, and had already failed. Reorganisations, new reporting lines, tooling mandates, values statements. A remedy aimed at a symptom is absorbed by whatever is generating the symptom. The failures were not evidence that the people applying the remedies were careless. They were evidence about the shape of the problem.

1.1.2 · A proof that the problem was not capability

Before any structural argument could be made, one explanation had to be eliminated — the one the organisation reached for first, and the one that was about to be acted on: the engineers were not good enough.

This could not be settled with internal evidence. Every internal instrument for judging capability — performance reviews, management opinion, project post-mortems — is produced by the same system under examination, and in a politically charged environment the honest answer to any of them is career-relevant. Asking the organisation to assess itself would have measured the atmosphere, not the capability.

So the assessment was outsourced, in effect, to the international labour market. For everyone on the technical staff over a window of a few years, roughly six to eight years ago, I recorded where they are now, and classified each destination against international academic and industrial standards. Every placement in that classification represents a judgement by an independent third party — an admissions committee, a hiring manager, a research group spending its own funding — none of whom had any stake in this division’s internal politics.

Roughly half of the technical staff from that period now hold internationally competitive positions. The classification was set conservatively on purpose: several placements that most readers would count as distinguished were recorded as ordinary exits, so the figure is a floor rather than an estimate, and it survives an unfriendly reader re-coding the marginal cases.

Full method, coding rules and destination inventory: 4.1 — the intellectual-assets analysis.

Note carefully what this establishes and what it does not. It establishes that the capability was present. It does not, on its own, establish that the organisation was dysfunctional — departure rates from this labour market in this period were high generally, and I have no comparison group that would let me claim otherwise. The attrition data is not my evidence of dysfunction. It is my instrument for measuring capability, and it is the only uncorrupted instrument that was available to me.

What follows from it is a reframing of the question. The division was not short of talent. It was short of a mechanism for converting talent into products anyone had asked for. So the problem must be stated as a property of the system:

The system does not reliably turn its intellectual assets into products.

1.1.3 · The method: a regress of causes

With the problem stated that way, I went looking for the mechanism. The procedure was borrowed, with some deliberate violence, from Aristotle’s regress of ends. In the Nicomachean Ethics he asks of each thing we pursue, for the sake of what? — wealth for security, security for a good life — and argues that the chain cannot regress forever, so it must terminate in something wanted for its own sake.

I applied the same interrogation downward, to symptoms rather than ends. For each observed behaviour, drawn from direct observation and from conversations with people at every level:

• What makes individuals behave this way?
• Why do they not choose one of the available alternatives?
• What condition makes this behaviour possible, or rational?
• What does this behaviour then make possible or necessary elsewhere?

And then, the two questions that did the most work:

• Is there any team or project where people face the same situation and respond differently?
• What is different there?

That last pair matters more than it looks. A condition that produces the same behaviour everywhere is a candidate mechanism. A condition that produces different behaviour in one team is a local effect wearing a mechanism’s clothes. The variation is the test.

1.1.4 · What emerged — from complex to complicated

The procedure did not produce what Aristotle’s does. His chain terminates. Mine did not.

Answers began arriving at nodes I had already written down from an entirely different starting symptom. Follow “products reach market with no identifiable customer” far enough and you arrive at the same place you arrive at from “engineers stop raising objections.” The diagram stopped being a tree and became a graph — and the graph contained closed circuits, in which each element sustained the conditions for the next, and the last sustained the first (Figure 2).

Figure 2 — the symptom graph, before any grouping. Eighteen observed symptoms and their connections, drawn without clusters. Solid edges are transcribed from the original interconnection analysis. Dotted purple edges are closures I inferred from what the nodes mean; they are analysis, not transcription, and they carry most of the loop structure. The figure is deliberately undifferentiated: this is the object the regress produced, and the point is that it cannot be read.

This is a stronger and considerably less comfortable finding than a root cause would have been. A root cause is a gift: it exists at the end of a chain, it can be located, and it can be removed. A self-reinforcing circuit offers none of that. It has no beginning to find. It sustains itself without anyone maintaining it, and it repairs itself around any single point you attack — which is precisely why every obvious remedy had already failed, and why the failures told us nothing about the people who attempted them.

It also produces a specific institutional pathology, which is worth naming because it explains why the problem had survived so long: the circuit had no owner. Each symptom, taken alone, belonged to a different department. Schedules to project control. Unclear products to market. Interpersonal friction to HR. Motivation to management. Every one of them was, from its owner’s desk, an ordinary problem of ordinary size that someone else’s carelessness had caused. Nobody’s job description contained the loop.

Further analysis of the graph

With “Aristotle’s regress of ends” in mind, you do not expect to end up with a graph of closed loops, where the nodes cause themselves through a process — a phenomenon sometimes called self-reinforcement, or a perpetuation cycle.

When you are dealing with dependency relations shaped like a tree, you only have to trace a cause backwards, and reaching the root is almost easy. In that case, careful identification of the causal relations is what matters. Once you get that right, the trace-back is the easiest part.

On the contrary, when you are facing a graph, tracing back from a node throws you into any number of non-terminating loops. It is pretty much what English speakers call a rabbit hole. In the next two sections you will see how I tried to find a way out of it.

1.1.5 · Clustering

I had to stop asking which node was the root, and impose a new frame on the graph so that it could lead me further.

Reviewing the graph over and over, I realized that there are sets of mutually reachable nodes in it. Inside such a set the elements sit very close to each other, with frequent causal relations between them, and they also belong to the same realm epistemologically — in other words, they are conceptually entangled. The immediate consequence was the idea of contracting each of these sets into a single node.

Clustering was the new frame I was looking for. By distilling the problem into one at a higher level of abstraction, it becomes much easier to think through the relations, and safer to make general statements about the situation and the dynamics of the issues.

What the groups are

The grouping I came up with is not a subject-matter taxonomy — the people problems, the process problems, the business problems. It is based on the mutual entanglement of the many different factors that were named during a violent brainstorming process.

Applying these gave three groups: hierarchy and structure, culture and norms, and strategy and market (Figure 3).

Figure 3 — the same graph, clustered. Same nodes, same edges as Figure 2; the only addition is that the three groups are drawn — hierarchy and structure, culture and norms, strategy and market. The inferred loop-closing edges are labelled N1–N10 and justified in the table below. The crossings between clusters become visible only in this view, and they are what the intervention is later aimed at.

Some loops these produce

1. Information loop — bottleneck → no participation → political behavior → opacity → bottleneck. Now literally drawn. Requires N1 and N2 only.
2. Toxicity loop — confrontation → negativity → confrontation (with the biases branch as a longer turn). Requires N3 alone; the source text’s word “contagiously” already asserts it.
3. Attrition loop — knowledge unused → motivation suppressed → the capable leave → knowledge unused. Requires the new node.
4. Withdrawal loop — no participation → political behavior → no channel → confrontation → negativity → suppressed motivation → less participation. The long culture-to-motivation circuit.
5. Market blindness loop — undefined goals → no feedback function → nothing against which to define goals.
6. Grand circuit — one circuit through all three clusters: seven nodes, four transcribed edges and three inferred ones.

Nothing has been added between Figure 2 and Figure 3. Same symptoms, same connections. The only thing that changed is that the groups are drawn — in other words, a new framing is imposed, and it lets us think about the graph at a higher level of abstraction.

Once the groups are drawn, a small set of connections becomes visible as a category of its own: the ones that cross between clusters. There are only a handful, and each one can be named:

What made me more confident that I was going in the right direction is that the finding is compatible with — and explains — a lot of failure stories from bigger companies attempting the same move toward the digital world.

These are the main relations (causal effects) — the pathways by which a failure in one cluster becomes a failure in another, and therefore the pathways that make three local problems into one circuit. They are the answer to the question the first figure could not be asked. And they have a direct design consequence, which the intervention document takes up: an intervention aimed at a cluster is absorbed, because the cluster is sustained from outside itself. An intervention aimed at a route is not.

Next: 1.2 · Clues — where the regress terminates, or jump straight to 2.1 · A new lens — the organization as a complex system.

⟵ Up: Collaboration 2.0 · Next: 1.2 · Clues — where the regress terminates →