Or: The future of software is less software.

1. The mistake I made, and why I’m glad I made it

When generative AI and the frontier LLMs started to become practical, I pushed every team in my portfolio to build an AI feature. I wasn't stressed about what they were building, but I pushed hard to get them all to build something. This wasn't because I thought adding on an AI feature would change the business or even make us money. In fact, we never charged customers for any of them as an added feature. I did this for the sole purpose of forcing people to engage with this new technology.

Just like any other time something new has come along. There were the early adopters who were already using the tool when I learnt about it, and then there were the mainstream people who were keen, and then there were the people who were a little bit nervous or were too busy or, for whatever reason, were not willing to immediately engage. The way to solve that problem and to get people engaged was to get something in production.

Once we got there, the interesting thing happened. The more we built and the more people showed each other, the more obvious it became that AI wasn't behaving like a feature at all. It was actually behaving like a capability. It was a capability that could touch every workflow in a product, and the more people used it and the more people built, the more they understood what this new technology could enable.

2. AI is no more a feature than a database is a feature

Nobody talks about “database features”.

Databases underpin software. AI is heading in the same direction. It's not the sparkle that gets put on to your logo, and it's not a sidebar. It's a foundational capability that changes how work gets done.

The question isn't "Where do we sprinkle the magic AI fairy dust?" The question should be "What work should exist at all?"

3. Customers don’t buy software. They buy outcomes.

Customers don't buy our software because they love CRMs or they love property management systems. They buy our software because they have a need. That need was never a need for forms, screens, web pages, workflows, menus, or dashboards. These all exist because they were previously necessary. What customers want is the outcome. The software is simply the mechanism that they're engaging with to get there.

Now that we've gone past using generative AI as a text generator, we're seeing that AI changes this equation because software can increasingly perform parts of the work itself.

The goal now isn't to make users more efficient inside of a workflow. The goal is to eliminate the workflow altogether, where possible.

4. The future is not humans using software. It’s humans directing software.

We often hear that the future will be agents talking to agents. I think that’s probably true. Agents will negotiate with other agents, complete transactions, exchange information, and trigger workflows without a human ever being aware it happened. The more interesting question is who they’re working for?

Humans will still define the outcomes and set the goals. Humans will still be sitting at the top of the value chain, determining acceptable risk, and humans will still be making the judgement calls. Whenever there's a non-reversible decision, that will be the human in the loop. The role of the human shifts from being the role of an operator to being the role of an orchestrator.

Every knowledge worker ultimately becomes the manager of a team of agents performing work on their behalf.

5. Context is the real moat

If agents are going to do meaningful work on our behalf, they need more than intelligence. They need understanding. When we talk about context, I don't mean prompt context. It's not a matter of injecting more information. There's operational context - questions that need to be answered, like:

What job is being done?
Where are we in the workflow?
What risks exist?
What guardrails apply?
Does this action require approval?
Can it execute autonomously?
Is it reversible?

An agent without context is just a chatbot, but an agent that has this context can perform meaningful work safely in a regulated environment. The companies that win won't necessarily need to have access to better models. They will have a deep understanding of the work itself.

6. Why dashboards survive

One prediction I hear constantly is that chat interfaces will replace software. I’m not convinced. Some CRUD screens will disappear, and some forms will disappear. Many workflows will collapse entirely, but the visibility becomes more and more important, as agents perform more work. Humans need confidence that things remain correct and under control.

There's evident behaviour emerging already of a concept called Cognitive surrender.

In Thinking, Fast and Slow, Daniel Kahneman describes how humans naturally conserve mental effort. When a shortcut is available, we tend to take it. Cognitive surrender is a natural extension of that behaviour.

When we see agents performing a job successfully over and over, we begin to trust the process and we tend to stop thinking. We start hitting the OK button and assuming everything is fine. That is ultimately what we call cognitive surrender.

To avoid this dystopian future of humans mindlessly hitting the OK button like Homer Simpson's drinking bird in the nuclear power plant, there are a few things that the future requires us to bake into the next generation of software:

governance
transparency
oversight
auditability

Not because AI fails. Because humans do.

7. The tapestry from the Atacama Desert

I wrote a post a while ago about a fragment of tapestry that I saw in a museum in Chile, which came from the Atacama Desert. This fragment was thousands of years old, yet it was still something to marvel at in terms of the skill that it took to produce it. The quality of the fabric, the vividness of the colours, and the sheer artistry to produce it. I stood in front of this thing and looked at it, and I was just struck by how, even today, that artwork is impressive. Extraordinary craftsmanship that was thousands of years old, but today there are very few people who could actually create that.

Automation didn't destroy textiles, though. There's more fabric produced today, more textiles produced today, than at any point in human history, but the skill necessary to create that has been delegated to the machines. There's also more humans who work in the industry today than at any point in history. What happened is that the skill itself became specialised, but the output became abundant. The same thing is probably going to happen with knowledge work.

The job to be done remained, but the work to get there changed.

The value moved, and the humans moved up a layer of abstraction, but ultimately it's still a human process delivering value to a human society.

8. What product leaders should do now

As we roll on into this world of change that is increasing every day, I think the most important thing to do is to stay grounded. Talk to customers frequently and find out where they're spending time, where work is getting stuck, whether passing between people or in laborious flows inside of your systems. Find out where the friction points are, where things are slow and where the effort challenges the value.

Then ask yourself, "Why does this work exist?" and focus on what we can do to:

eliminate work
unlock capacity
accelerate outcomes

and not focus on features shipped, prompts created, or models integrated.

Build capabilities.

Build context.

Build orchestration.

Build trust.

The companies that win this next phase won’t be the ones with the most AI features.

They’ll be the ones that understand their customers’ work deeply enough to remove it.

Your customers didn’t buy your software to use your software.

They bought it to get a job done.