The average said build. The odds said sell.

The idea came out of Covid. Anxiety was everywhere, and a friend of mine who specialises in helping people manage it was overwhelmed. When we talked about it, she explained how big the problem was: generalised anxiety disorder, PTSD and complex PTSD are major issues worldwide, and there simply aren't enough therapists. Self-help techniques had been researched and shown to work, but no single product could deliver them.

So I founded a company to develop a smart watch that could. The idea was to help someone at the moment they needed it, wherever they happened to be: driving, sitting in a meeting, lying awake in bed. I built the prototype myself. I worked with psychologists who specialise in generalised anxiety disorder and PTSD to develop a range of methods the watch could deliver. Academics validated the ideas, and the medical efficacy was mostly proven.

Then I had to decide what to do with it. I could build the company in the UK and take the watch into production, or sell the IP. I sold it to a US company.

I didn't model that decision. Getting the watch into production would have needed about $10m, along with tooling, techniques and contacts in China. That meant years as a constantly travelling chief executive, which I'd done before and didn't want to do again. And this was post-Brexit, post-Covid Britain, where investors had very little appetite for risk. My reasoning, if I'm honest, came down to one line: it wasn't worth the hassle.

Modelling decisions like this is part of what I do. So I've finally done it for my own, to find out whether my gut was right.

How the model works

The usual way to make a decision like this is with single numbers. We'll need $10m. We'll own 40% after the raise. The company will be worth $50m. Multiply them together and the answer looks precise. It isn't, because every one of those numbers is a guess.

A better way is to give each guess a range: a low figure, a likely figure and a high figure. Then let a computer play the decision out many times, picking a value from each range on every run. This is known as Monte Carlo simulation, but the idea is simple. Instead of one answer, you get the spread of answers, and you can see how often each option wins.

The inputs I used are below. Apart from the $10m, which is my own figure, they're illustrative: my best judgement now, not records from the time.

The inputs aren't independent, and that matters. My share of the company depends on how much I raise and what the company is worth before I raise it. Raise more and I own less. So the model calculates my share rather than guessing it: the value before the raise divided by the value after it. In most runs that put my share between about a third and three-fifths. A real model has dozens of links like that one. They're where most of the insight lives.

What 200,000 runs said

On average, building won. Across all the runs, building the company was worth about $4.3m to me. In one run in ten, it was worth more than $20m.

But in the typical run, building paid me nothing. The company only got funded in about one run in three, and only succeeded in about one run in eight. In the other seven, years of work ended with nothing to sell.

That's the whole decision in two lines. The average says build. The odds say sell.

The average is pulled up by a handful of spectacular outcomes. If you could make this decision a hundred times, building would be the right call, and that's how investors think: they make many bets, and a few big wins pay for the rest. A founder makes this bet once, with their own years. Seen that way, a seven-in-eight chance of nothing looks very different.

The number I didn't need

I'm not going to say what I sold the IP for, and it turns out I don't need to. Whatever sale price I tried, from $100,000 to $10m, selling came out ahead of building in nearly nine runs out of ten.

That surprised me. The sale price was the number I'd have been most anxious about at the time, and it barely moves the answer. Two other numbers did almost all the work: the chance of raising the money, and the chance of success once it was raised. Multiply a 30% chance by a 40% chance and you're down to about 12% before anything else comes into it.

This is one of the most useful things a model can do. It doesn't just give you an answer. It tells you which uncertainty to go and find out about before you commit. Here, the right next step wouldn't have been haggling over the sale price. It would have been testing how much appetite investors really had.

What the model can't price

None of this includes the reason I actually gave. Seven years or so of constant travel, building supplier relationships in China, and returning to a role I'd already decided I didn't want are real costs. The model doesn't price them, because I didn't put them in.

That's not a flaw in the method. It's a reminder of what the method is for. A model informs a decision; it doesn't make it. The honest thing is to name the costs you're weighing that aren't in the numbers, rather than pretend the numbers are the whole story, or quietly bend the ranges until they agree with what you want.

My gut was right, but not quite for the reason I'd have given at the time. "Not worth the hassle" sounds like a lack of appetite. The model says something more precise: the odds were poor, the upside was a gamble only an investor with many bets can afford to take, and the cost in years was mine alone.

The catch

I should be honest about the weakness in all this. I built this model knowing what I'd decided and how it turned out. That's the easiest way to fool yourself: you can nudge the ranges until the model agrees with you, without ever quite noticing you're doing it. I've tried not to, but you only have my word for it. I've written elsewhere about how generous memory is to the people doing the remembering, and I'm not immune.

The time to build a model is before the decision, when the ranges are your genuine uncertainty rather than your hindsight.

It works on engineering decisions too

Business decisions aren't the only place this helps. Many technical problems have a structure we understand well: the system follows known mathematical rules, but the parameters that feed them are uncertain. Component tolerances vary. Cells age. Temperatures, loads and usage patterns spread across a range. Suppliers quote a figure that turns out to be a best case.

That's exactly where a simulation earns its keep. Put the equations in, give each uncertain parameter a range, and run it thousands of times. Instead of a single “it should work”, you get the chance that the design meets its specification. You see which parameters drive most of the risk, which tells you where to spend testing and measurement before you commit to tooling. You can compare design options on the same footing. And as real measurements come in, you narrow the ranges and run it again, so the model improves as the project does.

For a wearable like the watch, the questions would have been things like battery life across real usage, or whether a sensor reading stays within tolerance across temperature and manufacturing spread. The method is the same one I used above, with physics in place of funding rounds.

What a real analysis looks like

This example is deliberately small: six inputs, one equation, two options. Real decisions in engineering and product businesses are bigger. A real model typically has many variables linked by the equations that connect them: cost to volume, volume to price, price to adoption, adoption to cash. It tests specific hypotheses, compares several options, and shows which uncertainties are worth paying to reduce before anyone commits. The principle is the same: ranges instead of single numbers, the odds as well as the average, and an honest list of what the model leaves out.

If you have a decision like this in front of you, it's part of the decision analysis I do.

---

Related: Decision Mining · The older I get, the better I remember I was · After the idea · Decision analysis · all writing