AI is changing how people use technology. For product designers, that raises the question of whether to build for how people use technology today or for how they might use it as AI improves. This essay considers both.
Building for the present
People bring expectations shaped by earlier generations of software. Building for the present means working with those conventions while using AI to improve existing workflows. Copilot in Word is one example: much of the working world still reviews documents there, and probably will for a while. Microsoft’s decision to build it there made sense.
Building from first principles
The alternative is to reason from first principles and ask what people will need in the future, independent of transient conventions they’ve learned. This is sometimes called “building for the future.” So how can we know what people will need in the future? We struggle to predict where technology is headed, and the pace of improvement makes that even harder.
Invariants
We may not know where technology is headed, but we can identify what stays constant. Software changes monthly; our cognition and biology, not so much. That makes the human side of the interface a more reliable thing to build on than predictions about where the technology is going. These are the invariants to build around. Whatever the future looks like, it still has to fit the human using it.
Intent may be the first invariant. However capable AI becomes, you still have to tell it what you want. As it gets better at carrying out tasks, you spend more time specifying the requirements in enough detail that it can execute them well. Clear instructions matter more when the AI can do more with them.
Speech is the most direct way to communicate intent. It’s roughly three times faster than typing and takes almost no physical effort. As agents become more capable and widespread, what matters most is how quickly you can tell them what to do, which makes that speed advantage more valuable. Wispr Flow built essentially a single feature on that speed in 2023. At the time, getting useful work from AI still meant supervising it closely. Better models make it easier to hand off the doing, which makes the ability to communicate your intent more valuable. That may help explain why a transcription tool has become so useful for talking to agents. The obvious destination is something like Neuralink… For now, though, the mouth is probably the best interface.
Attention is another invariant. Agents take time to work, so waiting for one feels wasteful. Running several at once makes better use of that time, but it forces you to track what each is doing. I switch between agents and forget what I asked the first one to do. A tool that helps manage that limit is another example of building around an invariant.
Learning from History
We don’t have to theorise about these limits in the abstract. The structures we’ve built over centuries already show how we’ve dealt with them. Where close supervision matters, companies don’t give one manager thirty direct reports, and org charts make the limit on attention clear. Résumés, references and probation address another limit, which is that trust takes time to build, and delegation isn’t safe without it (for AI, the closest equivalent might be benchmarks, though people’s faith in them is seemingly fading…). These are old structures solving a human problem, that is, getting work done when people are the limiting component. So if AI tools take a similar shape, say, an orchestrator agent supervising a handful of subordinates, I doubt the resemblance is a coincidence. Instead, both reflect the same constraint - a human with finite attention at the centre. This is one example, and I’m sure there are others.
That gives us two ways to find invariants. We can derive them from first principles (for example, any interface between a person and a machine must carry intent from the human’s mind into the machine), or we can read them in the structures we’ve built; what survives centuries of organisational trial and error is usually an accommodation to a human limit. Current software conventions are accommodations too, but mostly to the technical limitations of their era. Conventions age and invariants don’t.
In defence of satisficing
Building for the present is just as valid. Most people haven’t used the newest AI tools, so there is still plenty of near-term opportunity in making AI useful within the software they know. Building too far ahead is probably the bigger commercial risk: a product for users who don’t exist yet is not a business. The way to handle this is to work with the conventions people use now while keeping in mind how those conventions may change. The mistake, in my view, is to bolt AI onto existing systems without considering how it is changing the way people use technology. The two modes answer different questions. If the question is what to build now, build for the present. If it’s what using technology will feel like in ten years, invariants are seemingly steady enough to build on.
The full set of invariants
What is the full set of invariants, and what does the technology of the future look like when built on them? That is a more ambitious question. So far, I have identified intent and attention. I suspect trust and verification belong in the set as well. I want to work through the list properly next. Hopefully I will write about it soon - we’ll see.