On Productivity

One of the most common questions I get is: How do we measure AI’s impact on productivity?¹ It’s a fair question - if we’re investing in AI, how do we know it’s working?

Businesses often consider productivity as dollars saved or revenue generated - in other words doing more with less. It’s a two-sided coin². And if you can build a strong view of how AI impacts either side, you can start calibrating your investment.

Measuring productivity isn’t a new thing - we’ve been doing that since we invented the concept of measurement³ . We have many (DORA), many (SPACE), many (GTD), many (SMART), many (Kanban) frameworks to pick from. For each, you’ll have proponents and detractors. AI doesn’t change the fact that we’ll still have a bag of frameworks at our disposal.

Beyond any single productivity framework, what I’m interested in is how AI changes our deep-seated intuitions, and the structures built on them, specifically the belief that more input = more output.

Grounding us

Classical economic theory (everyone’s favorite start to a sentence) tells us that productivity is a measure of how efficiently inputs (originally just labor, then expanded to include capital and land) are turned into valuable outputs. This yields concepts like:

• Labor productivity = output per hour worked.

• Capital productivity = output per unit of capital.

• Total Factor Productivity (TFP)⁴ = growth not explained by labor or capital.

In modern companies, we think of inputs as time, headcount, or dollars invested, and outputs as features shipped, capital raised, or customers acquired.

Real-world complexity makes it too hard to create a strong link between inputs and outputs. For example, “two engineers worked on this for a week” (input) and “this generated $200k in revenue” (output) is rarely as simple as that. So we use proxies - subscriber growth, tickets closed, commits made - trusting they align with what matters.

Correlation breakdown

So what exactly is AI’s impact here?

AI is both collapsing the cost of production⁵ and accelerating production timelines, an order of magnitude further than software has.

While an oversimplification, consider some of these relationships:⁶

• Number of employees of a company related to the revenue generated

• Number of hours billed related to the contract size⁷

• 10,000 hours of deliberate practice to become an Olympian / master of your craft

These patterns hold in many domains, but now AI is challenging them.⁸ 

How AI shakes that intuition

We’ve long equated effort with value - AI challenges that belief.

• An individual can generate thousands of lines of code, dozens of designs, or contact millions of people in seconds. While accompanying this velocity is a decline in quality, the reality is companies expect people to operate at lightspeed.

• One person with an army of agents will outperform an entire team - not in every domain, but in many knowledge worker arenas.

• Serious people, now take seriously the idea of a one-person billion-dollar company.

That said, AI won’t fix your sink or help you shoot like Steph Curry - at least not yet. Some outputs still demand physical skill, practice, and good genetics. But when it comes to digital work, the old link between effort and output is quickly unraveling.

Squashing time

Ultimately we’re trying to understand where to focus our efforts. Do we hire more people? Invest in their training and tools? Do we open more stores? Do we raise money to fund larger marketing budgets? Will any of this lead to more revenue? Probably, but how much?

As noted above, the other intuition that AI challenges, is Time.

That means another one of our core input metrics, e.g., hours billed or days until X, needs to be reconsidered.

A problem that once took days to get to the next milestone (ideation to first draft, draft to PR, PR to deployment, etc) now takes minutes. This results in a shift of where we focus our time, from doing to managing (which spells trouble for middle managers!). This still takes us time (nothing is immune from Father Time), people and teams must adapt their skillsets to keep up.

It’s easier to measure the “effort” or “time required” for individual contributors when their focus is narrowly defined. It’s much harder to measure productivity for management-related activities, since the inputs are all mixed together. That leads us to rely on output-focused metrics, which have their own host of challenges.

Rethinking productivity with AI

AI is skewing the correspondence between inputs and outputs, how long things take, and what role we play in our work. As mentioned above, there used to be a correlation between, especially, labor, capital invested, land size, etc. and some target output. AI is changing all that.

In the old world, more effort led to more results. In the AI world, “working smarter” is ever more important. We need to consider leverage in a way we never did before.

There’s no single productivity metric, or even set of metrics, that everyone agrees is the right way to measure how efficient your team or company is. I don’t see that changing with AI. What will change is how we build teams and systems that can evolve to meet new challenges as the landscape shifts around us.

¹ This post doesn’t solve productivity in AI (whatever that means). There are lots of posts that go into greater detail on this subject. I have opinions and recommendations, and you can ping me to learn more!

² I often talk about a 3rd leg of the stool: Quality. The 3 dimensions (revenue, cost, quality) provide a nice tension to balance. A framework such as Balanced Scorecard gestures at this. But if we’re being real, most business are focused on revenue and cost.

³ TIL about the field of metrology (which is not the study of metronomes or metropolises, as I first imagined)

⁴ For example, consider two factories, with the same amount of capital invested and employees, that produce different amounts of cars. That delta is the TFP - how much more efficient one factory is over another.

⁵  Crazy salaries and untenable compute investments not withstanding.

⁶ To be clear, correlations, not cause and effect

⁷  Definitely going away with AI - the expectation now will be a bias towards using agents, not human labor. Hours per agent doesn’t work.

⁸ Though I’m skeptical AI will help much with the 3rd one…