Signups have compounded for eleven months. Support response times have crept from four hours to nine, then to two days. Both of those numbers get reported and only one of them gets discussed.
You are in one of four growth shapes and three of them are fine. The one that is not looks exactly like the others until the moment it does not.
Why these three models
The decision is whether to keep accelerating, and what would have to be true before you stopped. The features that fire are a quantity that accumulates, a limit that is somewhere ahead but not yet visible in the headline number, and a failure mode that does not arrive gradually.
Three lenses. Behaviour modes produce a cycle answer about the shape you are in. Thresholds produce a complex answer about why the limit will arrive as a step. Ruin sizing produces a random answer about which of the shapes you cannot afford to discover late. Naming the shape comes first, before any explanation, because every explanation you can construct will fit at least two of the shapes.
1. Name the shape before you explain the numbers
There are only a few characteristic trajectories a growing system produces, and each one implies a different underlying structure.1
Exponential growth: a reinforcing loop with nothing yet constraining it. Goal seeking: a balancing loop pulling toward a target, fast at first and slowing as it closes. S-shaped growth: reinforcing early, then a balancing loop takes over as a limit binds, and the curve flattens into a plateau. S-shaped with overshoot: the same, except the limit is perceived late, so the system passes it before turning back. And overshoot and collapse: the limit is passed, and passing it damages the very resource that was doing the limiting, so the ceiling falls and the system comes down with it.
The last two are separated from the first three by a single question, and it is the most useful question in this article: is the limiting resource erodible? A limit that simply binds produces a plateau. A limit that is degraded by being exceeded produces a collapse.
Work an example. Your support capacity is a limit. If you exceed it, tickets queue and customers wait, and when you add capacity the queue drains. That is a binding limit and it produces a plateau. Now change one thing: exceeding it long enough burns out the two experienced staff who hold the product knowledge, and they leave. Now the ceiling itself has dropped, the remaining team is slower, more customers wait, and more staff leave. Same limit, same growth, and the difference between a bad quarter and a spiral is whether the resource repairs or degrades.
So the diagnostic is not a chart. It is a list. Write down what actually constrains your growth right now, and against each one write whether exceeding it damages it. Staff goodwill, key-person knowledge, brand trust and a reputation with a small investor community are all erodible. Warehouse space, server capacity and cash are not, in the sense that overrunning them hurts without reducing them.
The list is worth more when it names people rather than functions. Support capacity is not erodible. The two people who answer the hard tickets are. Write each constraint as the smallest unit that can walk out of the building. A distributor who fronts you two weeks of stock is a constraint of that kind: exceed what he can carry and he does not simply stop, he starts allocating his working capital to a buyer who orders more predictably, and winning that allocation back costs more than it cost to earn the first time. Warehouse space does not hold a grudge. Anything that chooses whether to keep serving you does.
Then add a second column: repair time. Server headroom restores in a day. A distributor’s credit line restores after a quarter of clean payment. Product knowledge held by one person restores in six to nine months of somebody else doing the job badly first. Erodibility tells you whether the ceiling falls. Repair time tells you how long you would be trading under the lower one, and that is the number that decides whether you can survive the fall or only describe it accurately afterwards.
2. The limit arrives as a step
The second lens explains why you will not see it coming from the headline number.
Many constraints do not degrade smoothly as you approach them. They hold, and then they give way, because the system absorbs pressure through slack until the slack is gone.2 Support absorbs volume by working longer, cutting response quality and skipping the things nobody measures. Every one of those is invisible in the response-time average until the absorption capacity is exhausted, at which point the average moves all at once and everybody calls it sudden.
The metric was flat because someone was absorbing the difference personally, and that is not a metric you have.
Which gives the second instruction. Stop watching the outcome measure and start watching the absorption measures, because they move first. Overtime hours. Proportion of tickets handled by the one person who knows the answer. Time since the last person took leave. Escalation rate. These are leading and the customer-facing average is lagging, and by the time the lagging one moves the slack is already spent.
Most companies cannot instrument any of that this month. The cheap partial version is one question put to the constrained function every Friday: what did you not get to this week? The answer is an absorption measure. It depends on the person being willing to say so, it captures nothing systematically, and it still moves weeks ahead of the response-time average. Keep the answers in one place so the fourth week reads as a trend rather than as a complaint.
Expect the opposite of candour at first, because the person absorbing the difference is usually the one proudest of absorbing it, and the honest answer sounds like an admission of not coping. Two things fix that. Ask for the list of work skipped rather than for a judgement about load, since a list is reportable without conceding anything. And never answer the first honest reply by adding to it. Respond to “I stopped doing the weekly reconciliation” with anything other than removing something else, and the next four Fridays come back empty.
There is a variety check hiding in here that is worth running separately. Count the distinguishable situations arriving at the constrained function, and count the distinguishable responses it has available. If the responses are fewer than the situations, the function cannot be regulated no matter how hard anyone works, and adding people with the same response set changes neither count.4 A shortfall of variety is not fixed by a shortfall of people. The two available repairs are to reduce the incoming variety, by standardising or restricting what arrives, or to increase the response set, by delegating or automating a whole class of case. Hiring is neither of those unless the new person brings a response nobody else had.
3. Only one of the shapes can end you
The third lens ranks the four shapes by consequence rather than by likelihood, and that ordering is what should drive the decision.
A plateau is a disappointment. An overshoot with a recovery is an expensive quarter. A collapse in an erodible resource can take the company, because the ceiling comes down faster than you can rebuild it and the rebuilding depends on the very thing that was destroyed. When one branch can end the game, the average across branches stops being the relevant quantity and the size of the worst branch becomes the only one.3
This inverts the usual growth conversation. The question is not what growth rate can we achieve. It is what growth rate can this erodible resource sustain without degrading, and that number is usually knowable and usually lower than the one in the plan. Growing faster than it is not ambition; it is spending a resource that does not restock on demand.
The cap belongs in place before the opportunity arrives, because in the month a large customer appears nobody will agree to a limit. Write it now, in the form of a specific observable: we do not take on new accounts in a month where escalation rate exceeded a stated level, or where the single-point-of-knowledge ratio is above a stated share. A limit set in advance is a policy. The same limit proposed during the opportunity is an argument you will lose.
The cap has a failure mode of its own. Written too tightly it fires in a month when the erodible resource is fine and the escalation rate moved for an unrelated reason, and the second time you override it the policy is dead. So set the trigger where you would genuinely stop rather than where you would prefer to look disciplined, and write the restart condition into the same sentence: new accounts pause until the named measure sits under the line for two consecutive weeks, and only the founder restarts them. A cap with a stated restart survives being triggered. A cap without one gets argued away the first time it costs money.
Decide now what stopping means, because refusing revenue outright is rarely the version you will accept. The usable form is to keep selling and delay onboarding: contract signed, start date set four to six weeks out, deposit taken. The cash arrives on the old schedule, the load does not, and the constrained function gets the gap. It is a smaller action than a hiring freeze and it is available to a company that cannot afford to turn anything down.
What the three say together
- List your live constraints and mark each one erodible or not. That single column separates a plateau from a collapse.
- Move your attention from the outcome measure to the absorption measures, which move first.
- Run the variety count on the constrained function before approving a hire.
- Set the growth cap on the erodible resource now, as an observable trigger, while nothing is at stake.
Where they disagree
The mode analysis and the ruin analysis conflict about how much slack to carry.
Reading the modes says that most limits are not erodible and most overshoots recover, so carrying large buffers against a limit that will simply bind is expensive caution that a competitor will not be paying. The ruin argument says that you cannot tell in advance which resource turns out to be erodible, that the erodible ones are usually the human and reputational ones that are hardest to measure, and that surviving matters more than the buffer’s carrying cost.
The resolution is to separate the two classes rather than pick a philosophy. Carry minimal slack on the limits you can verify are non-erodible, since those genuinely recover and buffers there are pure cost. Carry deliberate slack only on the erodible ones, and accept that this is insurance you will look foolish for buying in every quarter where nothing happens. Founders tend to do the reverse, holding inventory and server headroom while running key people at the edge, because the first two appear on a balance sheet and the third does not.
What none of them contain
None of the three tells you when. Every model here gives you a shape and a structure, and none of them gives a date or a magnitude. That limit is real and it is not a defect of this particular set: take the shape from a systems model and never the number, because the numbers such models produce are artefacts of their assumptions rather than measurements of your business.
None of them handles the resource that repairs slowly rather than either recovering or degrading. Team trust after a hard period is like this, and it fits neither the binding nor the eroding category cleanly, which means the four-shape scheme is a simplification at exactly the place founders most need precision.
And one property the ensemble will not produce that you should expect anyway: constraints interact. Support strain drives engineering interruptions, which slow the fixes that would reduce support volume. Each of these models treats one limit at a time, and the thing that actually ends companies is two erodible resources feeding each other, which no single-limit analysis will show you.
The one action that survives the ignorance: before the end of this week, write your three live constraints and mark each erodible or not. If any erodible one is currently being exceeded, that is the only growth conversation worth having this quarter, and the decision is a cap rather than a plan.
Who has to move
The founder sets the cap and nobody else can, because every functional leader is measured on the growth that the cap would restrict. The instinct when the absorption measures move is to hire, which is correct only if the new person adds a response category the team did not have. The cheapest first test costs an hour: count the distinguishable situations arriving at your most strained function and the distinguishable responses available to it. If responses are fewer, you have found a variety shortfall, and the fix is standardisation or delegation rather than headcount, which also means it is available immediately rather than in a hiring cycle.
Sources and notes
- Donella H. Meadows, Thinking in Systems: A Primer, Chelsea Green Publishing, and John D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World, McGraw-Hill. The characteristic behaviour modes of dynamic systems, including exponential growth, goal seeking, S-shaped growth, overshoot with oscillation and overshoot with collapse, together with the feedback structures that generate each, are set out in both. The condition distinguishing collapse from plateau, that the limiting resource is itself degraded by being exceeded, is Meadows’ treatment of renewable versus non-renewable constraints. Used throughout section 1.
- Stefan Thurner, Rudolf Hanel and Peter Klimek, Introduction to the Theory of Complex Systems, Oxford University Press. Threshold behaviour, in which a system absorbs change with little visible response until a critical point and then shifts abruptly, is treated in the chapters on critical phenomena and phase transitions. Used in section 2 for the claim that the limit arrives as a step, applied here by analogy to an operational constraint rather than as a measured physical result.
- Michael J. Mauboussin, The Success Equation: Untangling Skill and Luck in Business, Sports, and Investing, Harvard Business Review Press, on variance and the asymmetry of outcomes, together with the standard treatment of absorbing barriers. The principle used in section 3 is that where one branch removes you from the game, the expected value across branches is not the decision-relevant quantity. No probability is assigned to collapse here, because assigning one would require the very forecast this article says these models cannot produce.
- Stafford Beer, Cybernetics and Management, English Universities Press, 1959. The law of requisite variety, attributed to Ashby, is adopted in the chapter on the Black Box as an axiom in the process of controlling complex systems, with the condition stated as sufficient permutative variety to provide a one-to-one transformation from the controlled system into the control box. The two repairs of attenuating incoming variety and amplifying the regulator’s are named as such in Brain of the Firm, Allen Lane, 1972. Used for the variety count in section 2 and the adoption test.
A note on a number this article does not give. There is no sustainable growth rate that transfers between companies. The rate an erodible resource tolerates depends on what it is, how fast it repairs, and how much of it you started with. Take the shape and the erodibility test, and get the rate from your own absorption measures over the next two months.
Joshua Agonya Pi’Rwot, Founder.