My question sounded like the title of a Marvel superhero movie. But the people who joined the conversation refused to pick a winner.

But let’s rewind for a moment. (Imagine that classic movie rewind effect.)

It actually all started with a completely different question.

How did we get here?

It started with something I felt was missing from AI: the ability to organize all my conversations into a structured, searchable database. For me, it was also about something I wanted to delegate. But , one of the readers, showed me just how much more there was behind this seemingly simple missing feature.

During our conversation, they actually suggested a way to solve my problem. Or at least how they would approach it:

Try exporting your threads to PDF and store in NotebookLM maybe.

And then came another option:

I could build an agent to do this task, but the design to minimise API calls would take time and interest.

And that’s where things got interesting for me.

A third option enters the game.

Historically, users have had two options.

But AI has introduced

And that’s where things get interesting. Now users can effectively build their own solutions, indirectly competing with existing SaaS products.

So I asked my subscribers a question:

Custom-built AI agents vs. digital products. Who wins?

And here’s what I learned.

Leave a comment

First, it’s not a choice between just two options.

The spreadsheet is still on the table. pointed out that people used to want the spreadsheet. Now they want the agent that fills it in.

And there are other options, too.

Like autonomy. Not allowing technocrats to control people. Well, the comment was actually much more specific about what exactly shouldn’t be allowed to control people. But my tone of voice won’t let me repeat it here, so I’ll leave you to read the original comment by .

But jokes aside.

Second, the ability to build your own AI agents opens up a whole new set of questions.

And this is where the conversation became particularly interesting.

The cost of custom-built agents -

nobody budgets for this: a product ships once, an agent needs a babysitter forever. most solopreneurs choosing “agent” are signing up for maintenance they never priced in

Reliability and specialization -

The product handles the boring, reliable part, and the agent takes care of whatever is specific to how your team works

The user’s workflow -

If the product already does the job well, why build an agent? But if you find yourself changing your workflow to fit the software, maybe the software isn’t the right fit anymore.

Use cases and the need for ongoing adjustments -

This raised another question for me as a product designer: how will users navigate between different agents if they decide to build their own, or use a multi-agent system built for them?

Flexibility -

Another factor in deciding whether to use an existing product or build an agent.

Sensitivity -

Whether an AI agent is even a reasonable alternative may depend on how sensitive the task is.

And each of these points opened up another conversation.

Share AI Adaptation Lab

More questions than answers

Are people ready to pay for ongoing agent maintenance?

And what happens if we show them the full cost upfront?

I’m still exploring what else I might have forgotten to include.

How do we account for the entire user workflow when building a product, rather than focusing on one specific task?

Oh, how I love this question.

From my experience as a product designer, this happens far too often. Products forget, or sometimes even ignore, the user’s actual workflow. And users end up either adapting their workflows to fit the software, cancelling their SaaS subscriptions, or building their own agents.

But the irony is that AI agents often have a similar problem. They might do one task really well, but struggle with what comes before and after it in the workflow.

Will managing multiple specialized agents become a UX challenge of its own?

Or will we simply have one interface coordinating everything behind the scenes?

Will custom-built AI agents compete with products in industries like fintech?

As Monika pointed out:

That raises another question about where existing products might have a competitive advantage.

So, who wins?

Perhaps it’s a good thing that “It depends” was the most popular answer.

Because this conversation raised more questions than it answered.

And now I have a whole new set of challenges to explore.

I’ll keep researching.

Stay tuned.

Thank you for taking part in this conversation:

, , , , , , , ,

Special thanks to for sparking this conversation.

This month, October 2026, we’re focusing on AI adoption metrics in the AI Adaptation Lab. A few days ago, I shared one signal about AI adoption metrics going up while product satisfaction goes down.

Just an observation, not a point I was trying to prove. But it became the starting point for an interesting discussion you wouldn’t want to miss.

The people in this conversation

Me with a background in product design and UX strategy, who started the conversation.

- Head of Engineering at Yahoo.

- Staff Software Engineer at Canva.

- an AI skeptic with around 28 years of experience in software.

What was on the table?

What if we’re measuring the wrong kind of work?

Haseeb brought up a metric I hadn’t considered: rework rate. And probably most importantly, he pointed out that almost nobody measures it.

Interesting, right?

But it’s kind of expected that users won’t get the right result from AI on the first try. I’d say it happens somewhere between never and very rarely, at least.

So I challenged his idea and got an important distinction:

Haseeb’s suggestion was to count only the second:

What I think is no less important in his comment is that dashboards rarely show this number.

Shouldn’t we be concerned about that?

Is reliability something we can fix?

The second perspective started with a concern about reliability.

But if I understood Timur Tibeyev, Staff Software Engineer at Canva, correctly, we ended up with three problems that, at this point in time, October 2026, still stand in the way:

• Context limitations.

• The cost of AI.

• The adoption of AI throughout the software engineering process.

And until we address these problems, reliability will remain a challenge.

How long will that take? Nobody knows.

AI adoption or greenwashing?

The third perspective came from Arun K S, an AI skeptic with around 28 years in software.

Greenwashing.

That’s what he compared AI adoption to, because, in his opinion, AI is being adopted in too many cases where it wasn’t actually needed.

What’s missing from these conversations?

Emojis, honestly.

What do I mean?

I mean that when you look at the metrics we’ve discussed with the professionals above, you don’t really feel the users’ pain. And if we want to make better decisions, shouldn’t we consider that too?

Maybe emojis could help a little with that. But jokes aside, here’s where I’m going with this.

First, I’d really like to continue this conversation. I think it started in an interesting way and became even more interesting as it developed.

So why not keep it going?

What do you think? Let me know in the comments.

Leave a comment

But since one of the goals of the AI Adaptation Lab is to propose experiments and observe AI in the wild to understand what survives and continues to grow, that’s exactly what I’ll be doing next.

I’ll be introducing practical experiments based on the questions we’re exploring here.

The practical part of the Lab launches this month.

Subscribe so you don’t miss it, and be among the first to experiment.

Adapt, not optimise.