What Are You Optimising For?
Ask my favourite question in a meeting or boardroom, and you will get a range of answers within seconds. The most common are growth, margin, shareholder value, customer satisfaction or efficiency. The chair will eventually say “all of the above”, and whilst everyone nods, the meeting will move on. Question addressed …… if only.
Any question answered that fast has not been understood.
In my view, “What are you optimising for?” is not one question; it is at least 6, layered on top of each other, and they all matter because they unpack complex/wicked problems, and, like all problems, the way you frame it determines what you are allowed to see.
There are two ways in. The first is what happens when people optimise a complex system. The second is what engineering already knows about optimising anything at all. They arrive at the same place from opposite directions; pick the one that you prefer.
One Direction. Optimising a Complex System
Nobody optimises for the organisation
Ask a founder what they are optimising for, and they will tell you growth, or customers, or revenue. Dig one layer down, and something more primitive appears: the need to prove that their original/current vision is right.
Founders are optimising for business success whilst also optimising for psychological safety, because they do not want to admit that the thing they invested everything in might need to die. This is not a founder pathology; it is a human strength or flaw.
In any organisation of any size, nobody optimises for the organisation as it does not exist without the people. We optimise for the things that determine whether we are considered good at our job. Which is a shadow or proxy, usually as a number, for something real.
We know that the moment that number becomes a target, Goodhart’s Law does what it always does. The call centre cuts handling time and destroys the customer relationship. The development team hits the date by borrowing against its own future. The hospital improves reported outcomes by declining the difficult cases. Nobody cheats. Everybody responds rationally to the structure they are inside.
The stack of proxies
Every time something becomes abundant, we add a layer of abstraction, and that layer becomes the new scarce thing. Reality, then data, then information, then metric, then dashboard, then incentive, then behaviour. Each layer is less coupled to physical reality than the one before it — we are living this.
I believe that we have crossed a threshold where the symbolic layer — the constructed world of metrics, algorithms, brand and status — is more functional, more responsive and more rewarding than physical reality for an expanding range of purposes. Corporations proved this decades ago, optimising abstract numbers while externalising physical costs. People are now doing the same thing to themselves.
So when someone in your organisation optimises, they are not optimising the system. They are optimising their model of the system. And the model is 6 layers of abstraction away from anything you could drop on your foot, (unless you are hands-on the tools)
This is the same question I have asked before as who is supervising your thinking, and then, more uncomfortably, who is supervising your data’s thinking. Optimisation is what happens after that supervision has already been handed over.
The things you cannot optimise for at all
Trust. Resilience. Judgement. Culture. Immunity. Capability. Loyalty. These are the properties every board says it wants, and not one of them can be built. They arise, or they do not, from conditions. This is what I have been calling consurgence: from con- (together) and surgere (to rise). Things that rise together, entangled, from conditions you tend rather than outputs you construct.
You cannot put trust in an objective function as you can only maintain the conditions from which it might arise. The conditions that allow these properties to consurge are, almost without exception, the things that optimisation frameworks are designed to remove. Friction. Slack. Slowness. Redundancy. Uncertainty. Presence. Repeated unnecessary contact. Time that produces nothing measurable.
Every one of those reads as waste on every dashboard ever built.
So the honest statement of the human problem is this: the emergent properties critical to organisational survival depend on conditions that optimisation frameworks systematically reward eliminating.
That is not a trade-off to balance, but it is a failure of the board to navigate.
The blindness, and then the recursion
Success creates blindness. When you have optimised for one model long enough, you hire people who think in that model, build processes that assume it, reward behaviours that defend it, and lose the organisational memory of anything else. Alternatives do not look wrong, they sound like gibberish as you now lack the vocabulary to process them. This is the argument I made in Why Are We Blind, Even Though We Can See?, The conditions you never had, or deliberately removed, make the obvious invisible.
The second is recursive: your capacity to judge which conditions matter may itself depend on conditions you are already optimising away. Judgement is a consurgent property, it has conditions. If those conditions are already gone, then the faculty you would use to evaluate the trade-off has been degraded by the very process you are trying to evaluate.
You are being asked to assess the damage using an instrument the damage has already reached.
Second Direction. What Engineering Knows
Engineering has been optimising systems for two centuries and has arrived somewhere with maths, but not with an answer, only as a discipline. And the discipline is the interesting bit, because it is precisely what organisations have discarded (optimised away) while keeping the words, narrative, story, vocabulary.
You cannot optimise until you know what you are optimising for!
Formal optimisation will not start until you understand 3 things:
- the objective function — the thing being maximised or minimised, stated explicitly
- the decision variables — what you are actually allowed to change
- the constraints — the boundaries within which any solution must sit
There is an unspoken 4th, which is the system boundary. What is in scope, and what has been declared outside it. An engineer cannot begin optimising until someone has committed, in writing, to what “better” means.
Get Tony Fish’s stories in your inbox
Join Medium for free to get updates from this writer.
Organisations inherited the maths and dropped the declaration. We optimise continuously without anyone ever writing down the objective function. Which means the objective function is implicit, unexamined, and usually inherited from whoever set the incentive scheme four years ago and has since left.
Where there is no “best”, there is choice
Optimise for one variable, and there is a single answer. Optimise for two and there is no single answer, but there is a Pareto frontier, a set of solutions where you cannot improve one objective without degrading another. Lighter or stronger. Faster or cheaper. Efficient or resilient. The maths cannot select a point on that frontier; only weightings can, and weightings are a value judgement (who decides)
So when a board says it is optimising for growth and sustainability and culture and returns, engineering has a precise response: you have not stated an objective, you have stated four, and you have not disclosed the weights, but the weights are the priority (strategy.)
Getting better requires getting worse first
Any hill-climbing algorithm finds the nearest peak and stops. It has no way of knowing whether there is a mountain range behind it. To reach a higher peak you must first go downhill — accept a period of deliberately worse performance with no guarantee of arrival.
Every serious optimisation method has a mechanism for this. Simulated annealing keeps randomness in the system on purpose. Evolutionary algorithms maintain a diverse population rather than breeding only from the leader.
Can you name a quarterly reporting system that permits this? How about the incentive scheme that funds a period of intentional decline in pursuit of an unproven peak? This is why local optima are the natural resting state of any organisation with a board calendar, and why the SaaS model reaching its ceiling is not a failure of execution. It is a peak. You cannot iterate your way off it.
Optimal control versus robust control
Optimal control delivers peak performance against a nominal model of the system. Robust control deliberately sacrifices peak performance in order to keep working when the model turns out to be wrong. Control engineers know which one to put in an aircraft. They accept a measurably worse machine in exchange for one that survives conditions its designers did not anticipate.
That is the resilience argument, already solved, already standard practice, in a discipline where being wrong kills people. An optimised system is, by construction, the system least tolerant of the difference between the model and reality.
Margin is deliberate waste, and we keep attacking it
Structural engineers apply safety factors. A component might be designed to two and a half times its expected peak load. On a spreadsheet, that is 150% waste, material bought and never used, cost with no corresponding output.
No engineer would defend removing it, because everyone understands that the factor is not covering the loads we know about. Now look at what an efficiency programme does to an organisation. It finds the slack, the redundancy, the doubled-up capability, the person who is not fully utilised, the meeting that produces nothing, and it removes them. It is a systematic assault on margin, conducted by people who would never dream of thinning a load-bearing beam.
We optimised away our margin and then wondered why we had no immunity left. Immunity is a distributed, adaptive, learning conversation between a system and its threats, and it is made almost entirely of the things a dashboard reads as slack.
Overfitting, and everything outside the boundary
Overfitting: optimise a model hard enough against its training data, and it will perform beautifully on the past and fail in the world today. An organisation optimised precisely to last year’s conditions is an overfitted model. Its accuracy about yesterday is the direct cause of its fragility about tomorrow. This is why ancient errors carry an advantage — error is the only thing that stops a system fitting itself to a world that has already moved.
Boundaries: everything outside the declared system boundary is free. It has no term in the objective function, so the optimiser cannot see it and has no reason to preserve it. Emissions, groundwater, burnout, community trust, the supplier’s supplier, your grandchildren.
We cannot afford to leave behind a burnt Earth just because we wanted to reach the Northstar at any cost. But no optimiser will ever tell you that, because the Earth was never in the objective function. Externalities are not a market failure. They are a boundary declaration, made by someone, at some point, usually without minutes.
So both arguments reach the same point from opposite ends and arrive at the same finding. Optimisation is not a way of discovering what is best. It is a way of obeying an instruction.
Six Questions
When I ask “what are you optimising for?”, here is what I am actually asking:
- What is the objective function — stated explicitly, in one sentence, with no “and”?
- Whose is it? Who chose it, who benefits, and who was in the room?
- When — over what time horizon? Optimal at three years and optimal at thirty are frequently opposites.
- Where is the boundary? What have we declared out of scope, and therefore free?
- What are we holding constant that is in fact variable? Every constraint is an assumption, and unexamined constraints are where strategy goes to die.
- What conditions does this consume? What is being spent that appears on no ledger, and could not be rebuilt if we wanted to?
And then the seventh, which is a governance question rather than an optimisation one: who is permitted to ask?
Why This Is Wicked
Rittel and Webber’s wicked problems have no definitive formulation, no stopping rule, and no solutions that are true or false, only better or worse. Every attempt changes the terrain. The choice of explanation determines the resolution. Every one of those applies here.
There is no definitive formulation, because how you frame the objective is the problem, not a step towards solving it. There is no stopping rule, because optimisation has no natural terminus; it runs until something breaks. There is no true or false, because the Pareto frontier is a set of defensible positions, not a right answer. Every attempt counts, because conditions consumed are not recoverable on demand — you cannot restore trust by scheduling it. And the choice of explanation determines the resolution, because whoever writes the objective function has already decided what the answer is going to be.
This is the same shape as the problem I described when AI learns to lie: an asymmetry that no amount of effort inside the existing frame will resolve, because the frame is the thing generating it.
It is also why measuring the success of AI is so hard, and why most AI dashboards measure the wrong things. We are answering the optimisation question before anyone has agreed what is being optimised. And when we hand the decision itself over to the machine, the objective function goes with it — which is the sovereignty question in its most concrete form. An AI is an optimiser. It will do exactly what you asked, against exactly the boundary you declared, at exactly the cost you did not price.
If optimisation is an answer, we need to make sure someone in the room has agreed the question.
