# Why AI Delivery Is Not Free

**Podcast:** All Things Product with Teresa and Petra
**Published:** 2026-09-15

## Transcript

Hi, folks.
This is All Things Product with Petra Wille.
And Teresa Torres.
And we're so happy you're here.
Teresa, I know you will love what I say next.
Now with AI, delivery is basically free.
It's so cheap.
It's basically free.
It doesn't cost us a dime to develop software these days.
Shoot.
It's what we hear about these days in all our calls, right?
I mean, when I look at my AI bills, it's certainly not free.
No.
Yeah, let's unpack this.
So I'm seeing this everywhere.
Is delivery free?
Now that delivery is free, complete the sentence.
I don't think delivery is free.
I don't think delivery will ever be free.
I think delivery is getting cheaper, but not all delivery is getting cheaper.
So I think there's a couple of ways I want to unpack this.
Yeah.
The first thing is building a single thing is getting significantly cheaper.
I think that is true.
But if the way companies respond to building a single thing being cheaper means they build lots more things, then delivery is not going to stay cheaper.
Because let's just think this through.
We have a product.
It's a brand new zero to one product.
We launch.
MVP.
So the minimum product that we think will get us a customer.
Deliver value.
Yeah.
It has one feature.
Now delivery is free.
So we add five features in week one.
Now it has six features.
Now in week two, we add five more features.
Now it has 11 features.
Now in week three, we add five more features.
It has 16 features.
Here's what's happening in those three weeks.
We just did this very quickly.
What's happening in those three weeks is feature one was easy to build.
There's no code base.
It was super easy to build.
Feature 2 is a little harder to build, but it's still easy to build.
By the time you're getting to feature 7, feature 10, feature 12, feature 15, your data model looks like a Frankenstein strategy.
Because you're using coding agents to build it with almost no engineering insight because you expect engineering to be free, you...
have spaghetti code everywhere.
You've rewritten the same code 17 times, which means when you want to change something, you're changing it in 17 places.
Performance already down.
Yeah.
Your performance of your product is probably terrible from a speed and performance standpoint.
The maintainability of your code is terrible.
But sure, we got 15 features for free.
And I think this is like, if you don't understand engineering, it's really easy to naively think.
coding agents makes everything free.
Coding agents makes things faster.
That is certainly true.
And experimentation easier.
And it allows us to create throwaway code very easily.
But we can't mix those things up.
And actually, I really love the SVPG folks are starting to use this phrase build to earn versus build to learn.
And there's this idea of like, build to learn is we're going to write code to build an interactive prototype that we're going to put in front of people.
And that is dirt cheap today.
And good.
As long as we throw it away.
Yes.
Right?
As long as we throw it away.
When we build to earn, now we know what to build.
we're building so that a paying customer buys it, we still need to care about all those engineering principles we've always cared about, which is maintainability of our code, non-functional requirements, as we would call them.
Reusability of our code, the cost to maintain that code, the security of that code, how well that code scales.
And can AI help with us?
Yes.
Is it free?
No way.
Like, not even close.
And for that, you would still need a skilled engineering team observing and overlooking what AI does.
Absolutely.
And it takes time.
It takes time.
Like, AI can generate a lot of code really quickly.
We can get more sophisticated in our AI reviews of that code.
But it probably still requires a human to look at it and say, is this the architecture that's going to grow, given our company's vision of where we're going?
Right?
And like, maybe here's the challenge with AI episodes.
Like maybe in six months, I'm going to regret saying this because AI is going to get really good at understanding your product vision and creating the perfect code base that scales infinitely.
We never know.
Maybe.
But even if all of that happens, even if all of that happens, even if all these harder engineering problems just get solved by coding agents, we still have this core fundamental problem on any product.
The more bloat you add to the product, the harder it is to build on that product.
Yeah.
Plus, may I add another layer?
I just recently saw a chart basically that said like apps released versus apps used in the app store.
And so apps released is spiking, of course, but apps that people are actively using over time, that's very stable.
So it's not only that the code bases get worse if you use AI delivery for the stuff you want to earn money with.
Consumers have more options.
But they're more likely to be really sensitive when it comes to the quality of the tools and the apps they use because they will get very good in sensing like which products have been really thoughtfully built and designed and which ones are basically AI app slop.
So I bet that will be a skill that humans will develop in the next six months.
And then it's very hard to compete and earn with a product that is not well designed, thought through, tested, ruggedized, where all the non-functional requirements are not even taken into account, which is slow and where it's a lot of feature creep.
And yeah, so that's another layer of why it's not a good idea to think building is free.
I think there's another category to do this to explore, which is...
AI coding agents are really good at writing deterministic code.
And so if we're going to live in a world from five years ago and we do great discovery and we know exactly the right thing to build, sure, delivery got a lot cheaper.
We can build deterministic code features really easily with coding agents.
But we don't live in five years ago.
We live in 2026 where it turns out we can do amazing things with AI.
And if we're building AI products, which is probably what's going to be required to survive in the future.
At some point, yeah.
Coding agents don't make building AI features free.
No, it doesn't.
Right?
So I know I've been spending months building two AI features and it takes a lot of error analysis and creating evals that measure the right thing and iterating on prompts and orchestration and like...
Yes, there are some people out there that are starting to automate some of those pieces with AI, but I can tell you it's not very good at it.
No.
And so like we're stepping into this new world where the types of products we're building are changing.
And that delivery is definitely not free.
No.
Yeah, because everybody...
Because of the learning curves to some extent as well, right?
Because everybody has to experiment a lot, play a lot and learn a lot while they go because nobody has figured it out right now with all these AI products.
And yeah, we get this first, I always see like a river and now the water is kind of, the water is going down a bit and you see the stepping stones.
And I think we already see some stepping stones AI wise.
So these are the things that are more stable.
We can step on that stone.
That is something we need to take care of.
of context rod and context engineering and everybody starts to have terminology to talk about certain things, your eviles topic.
But still, there's so much influx.
There's so much water still.
And still, we are maneuvering this kind of early beginnings of AI and how it works.
It's like command line has been invented and now everybody is figuring out how to use the internet.
Yeah, I think we...
are just scratching the surface on this stuff.
I mean, I can give a really personal example.
We are AI generated snapshots and opportunity solution trees with Vistily are going really well.
We're in closed beta.
We're actually seeing our beta customers convert to paid, which is amazing.
We're getting ready to go to GA and offer it to everybody.
It's really fun to sit on a customer call and watch a customer, like look at it and go, wow, this is better than what we're doing on our own.
It looks really successful, right?
Awesome.
I look at it and I go, oh, that opportunity is poorly framed.
Oh, that opportunity is closely worded with the key moment.
And it looks like a duplicate.
But we had that in another episode.
If you're an expert.
Yeah.
So I see quality issues.
And so what do I do?
I capture all those quality issues.
I'm going to go do my error analysis.
How often do they show up in my product?
I'm going to try to create an eval to measure it.
But it's not this simple.
Like we can all read about how do we do evals.
Okay, well, I got to do an LLM as judge to detect it.
But if in that single part of the tree, there's actually four competing errors, I can't measure one without solving for the others.
This is like a very complicated data science problem.
Yes.
And I'm using Fable to help me.
So we're talking at the time that we're recording the latest frontier model.
That apparently was so great that it was a national security risk to help me.
But it's back in Europe.
Just if you're wondering, currently it's back in Europe.
It's back in the US as well.
It does really shallow analysis.
It mixes up the failure modes.
I'm not going to let it just solve this problem for me.
It's not there yet.
So I can tell you from personal experience, my delivery is absolutely not free.
I'm still spending days looking at my data.
trying to like distinguish between this error and this error, trying to run experiments to figure out how to make those errors go away.
And this is for a product that works reasonably well for customers.
Yeah.
But I don't want to build a product that works reasonably well for customers.
I want to build a product that is amazing.
Yeah.
But what it is really helpful with, so yesterday I was sitting on, on a demo where basically, um, It is an entire new marketing claim, new logo, new design, new design systems and design for basically print and leaflets and digital and social media.
And so it's an entire new design system, right?
And the designer was so creative in demoing it because he really created with the help of AI, obviously.
super interesting exciting that spikes curiosity tool that helps the people to explore the extremes of the design system for example so and is this built to earn not but it's built to learn and it's not even built to learn maybe it's just like built to convince the internal stakeholders over or something like that.
But still, because it being cheaper to build such a thing, he was able to build it and he was able to put it in the hands of the people that can now get curious about a new design system and explore it.
And then they can have way more informed discussions afterwards about the, yeah.
the extremes and guardrails that the design system maybe should have.
And they would have never been able to have these in-depth conversations without 10 or 15 people playing with it and see it, what it can do, right?
Without the tool, the design person would have not have the substantial conversation.
So for these kinds of use cases, I think it's amazing that we can spun up such a tool in such a short amount of time.
I completely agree, but I would put all of that in the build to learn category.
Yes, of course.
And I would put all of that in the discovery bucket, right?
Yeah.
I think when we talk about delivery is free.
And it's free enough for that bucket.
Yes, we're talking about production quality, scalable, maintainable code.
Yeah, yeah.
And I don't even think we're there with deterministic code.
I can see on the horizon we might get there with deterministic code, but I don't see on the short-term horizon that we're going to get there with AI products.
And I think even with AI products, it seems like we're there because I can create a prototype of an AI-generated OST, and it's going to look pretty darn good.
And you're going to go, wow, this is an amazing product.
In fact, I hear from product teams every day that are doing this themselves.
They don't understand why they would buy a product to do this for them because they can get Claude to generate an opportunity solution tree for them.
Okay, great.
But...
You're at the prototype level.
You're building to learn.
You're not at a production quality product that you can trust and rely on.
And I'm not saying this just because of what I'm building.
I think this is true.
Like in every just now possible episode that I record, what the products that are doing really well, what's really clear is that team has immense domain expertise about the problem they're solving.
And they're putting in a lot of hours of hard work.
to train the AI to apply that domain expertise expertly.
And the first 60 to 70% easy.
It's a prototype.
It looks reasonably good.
Closing that last 30% is months to years of work.
And it's not free.
It's not even remotely free.
It's like what's missing in the blog post, the last 10% of words so that they really land well.
And I think if we look at like, I think in our previous episode, we talked about how many apps are being added to the app store, but they're not being used.
This is the same thing, that 60 to 70% prototype.
That's great.
You can ship something really quickly.
But I think customers are going to require that last getting to 95% before it's really useful.
Yeah.
So the important thing is we don't subscribe to the delivery is free narrative and we don't want to reemphasize it because we think it is not, right?
Yeah, and I'll say I have every incentive to tell you that delivery is free.
Because when people tell you deliver is free, what's the back half of that sentence?
They say delivery is free.
They either say one of two things.
Delivery is free.
Taste is all that matters.
Or they say delivery is free.
Discovery is all that matters.
Okay, taste is baloney.
We've already talked about that in the past.
I would love to tell you that discovery is all that matters.
It's not true.
Delivery has not gone to zero.
Like they both matter.
They're both going to continue to matter forever.
Yes, discovery might be coming more and more.
I'd love to tell you that.
But delivery is not free.
Yeah, it's not free people.
It's not free.
Thanks, Teresa.
Thanks, Petra.