Predictability and variability: Lessons in balanced product development in the age of AI

Balanced product development in the age of AI means knowing where to automate, where to adapt, and where human judgment still creates the greatest value.

My first introduction to the world of product management was the ScriptRunner family of apps on the Atlassian Marketplace. In many ways, it was an ideal place to start: The product had a loyal customer base and a growing value proposition that I had grown passionate about in my previous work.

At its core, ScriptRunner was a platform; a suite of tools and guided pathways that allowed users to take an off-the-shelf product, such as Jira, and customise it until it fit ‘perfectly' into a company's way of working. Whether that meant changing how a product looked, integrating it with other enterprise tools, or automating manual work, ScriptRunner let users trade a small bit of upfront labour for ongoing automated labour. In practice, that meant a small IT team could scale its impact without increasing its cost. In a sense, you could exchange human labour for machine labour through automation.

There's a trade-off that many customers understand instinctively, and it still shapes how leaders approach the adoption of machine intelligence today. Machine labour is typically less expensive and more static. Human labour is typically more expensive and more adaptive.

When you are building anything, whether it's a product, a service, a workflow, or a piece of content, there are a number of different stages in the assembly line that combine together to deliver the final result. In the largest organisations we worked with, those chains of coordination can be astronomically complex.

Automation swaps out parts that chain, replacing an expensive, adaptive resource such as human effort with a cheaper, more predictable machine process. The organisations that did this well understood how work flowed through their systems and stayed focused on the value they wanted to deliver to customers. That made it easier to introduce machine labour in a balanced way. Balance is the key, as anyone who has felt the effects of over-optimisation or over-automation will tell you. But the trade-off is never free: in its classical, industrial sense, automation comes at a price because it cannot adapt. A Rube Goldberg machine is a useful metaphor for over-automation. What happens when one part breaks? And what happens when you want the system to do something slightly different from what it was originally designed to do?

The key point is that automation is not just a cost-reduction exercise; it is also a trade-off between predictability and variability.

This is not simply a cost-optimisation exercise, either. In a world where everyone automates work in the exact same way, with the same inputs and outputs, there is little room for competitive advantage. Everything starts to be the same.

An organisation's (your) ability to innovate depends on preserving some variability in its outputs: the freedom to research, test and develop new (and previously untested) approaches in search of a competitive edge. It is precisely this seek-and-exploit strategy that keeps a business successful in its industry and afloat when the conditions are less predictable and more difficult.

Large Language Models have accelerated a wave of machine intelligence embedding itself into the way everyday work gets done.

At their core, these systems introduce a new kind of machine labour: machine intelligence, rather than simple machine automation. They do this through variability, powered by complex probabilistic matrix math. A language model can take the same input that you might give a traditional automation, and, based on patterns learned during training, predict a useful output.

The picture becomes more complex because this intelligence can operate across many languages that power software and automation. And for many tasks, the combined cost of machine labour (automation and intelligence) can still come in below that of a human counterpart.

Know thy work labour.

You can think about these dimensions of labour in two intersecting axes:

Human <> machine
Automation <> intelligence

What people often misunderstand about AI systems is that they exist within an ecosystem: interconnected parts along an assembly line, supported by a mix of tools and intelligence at each stage of a workflow. After all, humans rarely build digital products in isolation, without tools. AI systems are no different: they are rarely deployed in isolation. What works in practice is usually a blend, and again, it comes back to balance.

The organisations succeeding with machine labour today are often the ones that have already learned to introduce automation thoughtfully and are now applying the same discipline to machine intelligence. In practice, it comes down to answering a few simple questions:

Where should this task or process stay static, and produce a predictable output?
Where should this task or process adapt, learn, or change over time?

There are plenty of sweeping claims about this technology, passed around at speed across feeds, forums, podcasts, and boardrooms. That is understandable. The topic is dense, and reducing the interplay between intelligence and action, humans and machines, is often the only way to make it discussable.

But the real story is far more interesting than that. New combinations of intelligence and automation, along with new CLIs, platforms, SDKs, APIs, MCPs, and UIs, are emerging almost as quickly as new AI models are trained and released. Staying on top of it requires a deep understanding of the nature of work, and of how different forms of labour create value in different ways.

For all the noise around AI, the lesson feels surprisingly familiar: know the work, know the trade-offs, and choose deliberately. The tools may have changed, but the principle remains the same. Great product development still depends on balance.