I remember first thinking seriously about entropy while watching Brian Cox’s BBC programmes about the universe and life. What stayed with me was the apparent tension: nature and chemistry building intricate forms, while the wider universe moves towards greater entropy.
Those processes belong to the same physical story. Life can build and sustain local order by drawing on a flow of energy, while total entropy increases. Complexity can emerge, but it needs conditions that support it.
Cox also introduced me to Darwin’s phrase, “endless forms most beautiful”. I wrote a song with that title. It never made it out of the practice room, but the mixdown is still on SoundCloud.
Darwin’s account also describes how divergence allows forms of life to occupy different places in nature. Specialisation matters: different characteristics can confer an advantage in different conditions. There is a useful business parallel here. A business can find its place by becoming particularly good at something that a particular group of customers needs.
Years later, I find myself returning to that tension when thinking about businesses. People build useful processes, relationships and ways of working. As the business changes, those arrangements need attention if they are to keep making sense together.
I use business entropy as an analogy, not a law of physics. What interests me is the work needed to sustain coherence while allowing something new to grow.
A small example from the shop floor
At Biketrax, I used to buy one of each size of a garment. If a particular size sold and I did not restock it, its recorded sales stayed at one. The figure was perfectly accurate. What it could tell me about demand was much more limited.
Perhaps only one person wanted that size. Perhaps several more would have bought it if it had remained available. The sales record could not distinguish between those possibilities. Once the item was gone, there was no further opportunity for demand to appear as another sale.
The awkward part was the circularity. My purchasing decisions determined what was available. Availability limited what could be sold. The resulting sales history could then inform the next purchasing decision. I risked treating the consequences of an earlier choice as independent evidence for what to do next.
Nor was the answer simply to replenish everything. Stock committed cash, ranges changed, and leftover sizes and unsuccessful experiments accumulated. Buying more might reveal demand, but it could also leave more money sitting on a shelf.
Nothing had to be wrong with the till, the stock records or the report for the business to struggle with the decision. The problem sat in the relationships between them.
How the gaps accumulate
This is what I mean by thinking in systems: looking beyond the immediate task to understand what produces its inputs and what its results will affect.
By business entropy, I mean what happens when those relationships become harder to follow. A process changes, a responsibility moves, a new service is introduced. People adapt to get the work done, but the rest of the organisation does not necessarily adapt with them.
Over time, useful work can require more reconstruction. Someone has to explain why an exception exists, recover the reasoning behind an old decision or reconcile two accounts of what was promised. The parts are still working, but making them work together takes increasing effort.
The garment example exposed how easily a connection could go missing: the report retained the sale, but the decision needed an understanding of the conditions under which sales were possible. A record was being asked to stand in for more than it contained.
A better report is not always the answer
A more polished chart would not, on its own, have resolved that uncertainty. Before asking what to order next, I needed to distinguish what had sold from what had actually had a chance to sell.
That points towards a different investigation. How long was each size available? Did customers ask for something we could not supply? How quickly could we replenish it, and how much cash could we sensibly commit while learning? None of those questions provides a perfect demand forecast. They establish what the decision needs and where judgement remains necessary.
That is a change to how the work is organised, not merely how it is reported. Sales observations need a way to reach purchasing. Availability needs to be considered alongside transactions. Whoever makes the buying decision needs to understand the uncertainty as well as the numbers.
I do not need every possible relationship mapped before acting. I need the ones that could materially change the decision. Otherwise, “everything is connected” becomes an excuse for making a problem larger rather than making it more manageable.
Divide the work without losing the understanding
Custom bicycle design gave me another version of the problem. The intention was that salespeople would design and sell the bicycles, while mechanics would build them. It was a reasonable division of labour. Sometimes it worked well; sometimes it generated difficulties.
Compatibility questions did not wait until the bicycle reached the workshop. They were already present while discussing what the rider wanted and choosing components. Technical understanding was also part of earning the customer’s trust. The person making the recommendation needed to be credible about what they were recommending.
We could divide the tasks more neatly than we could divide the understanding needed to perform them.
The answer was not necessarily to have mechanics conduct every sale or salespeople build every bicycle. It was to examine where their knowledge needed to overlap: which choices could be made confidently, which needed checking, and when the workshop needed to enter the conversation before a promise was made.
I recognise the same problem in ERP work. A commercial commitment can look reasonable until someone examines what delivering it requires. Handing over the paperwork does not resolve an assumption that should have been tested before the agreement.
That is why I pay attention to the boundaries between roles. A clear job description tells someone what they own. It also needs to leave them a dependable way to resolve decisions that cross into someone else’s expertise or authority.
Enough context to act
I think of this as creating little pockets of context: enough understanding to do a particular piece of work properly, with a route back to the detail when it is needed.
The person buying clothing does not need every transaction permanently in front of them. They do need to know when apparently weak sales might reflect limited availability. A salesperson does not need to become a mechanic. They do need to recognise a compatibility question before turning it into a customer commitment.
A useful pocket of context is therefore different from a silo. It narrows attention without cutting the work off from the relationships that matter. It should also make it possible to say, “I don’t have enough information to decide this.”
Giving everybody access to everything is not the same as giving them that understanding. Nor is requiring the owner to explain the whole situation every time a decision crosses a boundary. The work is to make the relevant context available where it can influence what happens next.
Applying this to my own practice
I am exploring the same principle in OrchestratorOS, the system I am developing to organise consulting work at Polyphony. The aim is to preserve the connections between what a client needs, what we agree to do, what gets delivered and what the client accepts.
Part of the design involves smaller AI models assigned to specific tasks, alongside people and ordinary software. One task might draw out unresolved questions from workshop notes. Another might prepare instructions from behaviour that has been tested and accepted. Each needs a different view of the work, but those views must connect to the same decisions.
Consider a hypothetical change to purchasing approvals. The software is updated so that a deputy can approve an order, but the instructions still tell staff to wait for the original approver. The configuration may work exactly as requested. The instructions may accurately describe the old process. Together, they leave the original delay unresolved.
Completing the individual tasks is not enough. The system needs to identify what the changed decision affects, and the combined result needs checking against the business purpose.
The garment example provides another test. An AI asked to recommend stock from sales alone would still be missing the availability problem. Several models agreeing about those same figures would not supply the missing evidence. The useful response might be to identify what is unknown, rather than produce a more confident recommendation.
This is work I am developing and testing, not a claim that consulting can now run autonomously. I would choose a model by how well it performs the task, not because smallness is a virtue. People remain responsible for consequential decisions.
The principle does not depend on AI: give each participant enough context to contribute properly, preserve the relevant connections, and test whether the contributions work together. The pieces can become smaller while the whole becomes less fragmented.
Find the next move, then see what changes
When I think I have recognised a pattern, I test it. What does the explanation predict? What else should I expect to find? What would make me change my mind?
Much of that reasoning begins with experience and instinct. I rarely see the whole solution at once. I start with the most foundational problem I can usefully address, or the immediate constraint that needs attention before other work can progress.
It is close to how I used to construct a song. I would develop a part, listen to what it changed and let the next part grow from that relationship. Something could sound good on its own and be wrong for the arrangement. I needed to return to the whole, not just keep improving the individual tracks.
In business, that means examining where the effects of an intervention appear. Has a quicker sales process genuinely helped, or left operations doing the checking later? Has a simpler report made a decision clearer, or hidden the uncertainty that mattered?
It also means resisting the urge to make everything conventional. Björk is a useful musical reference for me: an unfamiliar combination can make sense on its own terms. An unusual business process deserves examination, not automatic removal. I want to know what it contributes before deciding that its awkwardness is a defect.
Sometimes that specialisation is exactly what gives a business its edge. A particular service, depth of expertise or way of handling a customer’s needs can be the reason that business has a place in the market. We should never fight that specialisation simply because it makes the business harder to fit into a standard process. We need to understand what sustains it and organise the surrounding work so it can succeed.
I explored this in From realistic unicorns to testable requirements: an unusual way of working may be part of the reason customers choose a business. The question is what value it creates and what would be lost by changing it.
Coherence should give those differences room to work. A business can become easier to understand and coordinate while preserving the specialisation that makes it worth choosing.
Organise for understanding
The lasting improvement I am looking for is not that I become better at joining everything up personally. It is that other people become better equipped to recognise the connections relevant to their work.
That is why mentoring matters to me. I want colleagues to develop their reasoning and instincts, not memorise my answers. What else depends on this choice? What are we assuming? What evidence would change our recommendation? When does someone with different expertise need to be involved?
Those questions need somewhere to go. People need access to the relevant information, authority appropriate to their responsibility, and a way to challenge an arrangement that no longer fits. Asking someone to take ownership without those things leaves them responsible for a problem they may be unable to resolve.
I would judge progress in practical terms: a promise checked before it creates rework, a buying decision that acknowledges what the sales figures leave out, or a changed rule reflected in both the system and the instructions people use. Less time spent reconstructing yesterday’s reasoning, and more ability to make today’s decision.
The business will keep changing, so this is not a one-off exercise in tidying up. Thinking in systems means noticing when the relationships change. Organising in systems means giving people a way to respond.
That brings me back to what caught my attention in Cox’s programmes. The capacity to build something coherent is worth sustaining. In a business, that means putting effort into the relationships that allow people to learn, adapt and create, as well as into the individual tasks.
At Biketrax, “one sold” was a correct answer to a narrow question. The business needed to understand what had made that answer possible, what it left out and what to do next.
That is the work I want Polyphony to do: make those connections clear enough that people can reason and act together, without needing everyone to hold the whole business in their head.