From an engineering and product perspective, I use generative AI every damn day, so this is not a critique of the technology or any particular application of it. But as someone who's been knee-deep in building AI-native products for 4+ years, when I look at the messaging today in the product and dev ecosystem, it almost seems like a pre-requisite that products have some sort of generative AI capability to be relevant.
The obvious statement is that not every product benefits from having generative AI. But even for the ones that do, it's far from being all upside; there are considerable costs, risks, and even stark product-level trade-offs, which I don't see as widely discussed, but are worth serious consideration.
Oftentimes, as users, we seek products to help us navigate ambiguity within a problem space. We're often looking for systems that help make sense of something that's difficult, complex, or tedious for us to accomplish, understand, or maintain, rather than something that delivers a clear-cut solution to a clear-cut problem (that we may not be able to even articulate).
While extremely powerful when leveraged correctly, generative AI doesn't replace a strong conceptual model, nor does it provide an intuitive experience that delivers a clear, differentiated value proposition. That's the job of a good product, and while generative AI can fit into that as part of a larger strategy, it should always be considered a tool that strategically helps deliver on a core value prop.
This is a clear-eyed take on integrating user-facing agentic AI into products in particular, from a product perspective (and with the assumption that we're talking about integrated external generative AI models, not ones we create and own ourselves, in which case that is the product).
I have a lot (9+ years) of experience building no-code products from both the engineering and product side; the high-level problem space was always at least somewhat clear to me: given user x, what is the most compelling and intuitive experience to allow them to design and build y?
Regardless of whether it was b2b, b2c, or b2b2c enterprise platforms, that problem space loosely framed how I thought big-picture. With the introduction of generative AI, and my own experience over these past years pivoting no-code platforms to agentic AI, as well as building them from scratch, I've seen first-hand the profound implications it has not only on problem spaces, but value propositions, user relationships and expectations, financial modeling, stability and quality, maintenance, and more.
Responsibility and Ownership Shifts
In no-code platforms, my experience is that we're generally presenting what we hope is a compelling, intuitive conceptual model and experience for a user to build something that either:
- they don't have the expertise to create themselves
- would be cumbersome, impractical, or expensive to learn the requisite skills or offload onto a third party (other than you, of course!)
Ideally, this problem space is examined and iterated on holistically by a product team—the mental model and experience has to be intuitive and compelling for users, and that entails some level of meaningful, cross-cutting abstractions over system design, UI/UX, content management, and (at every company I've worked for) the inevitable escape hatch for when a more technical user needs to duck your abstractions and actually write code.
How this manifests in reality has been utterly different for every product I've worked on. What has never varied is the concept of responsibility and ownership: we provide a platform that enables the user to create content with. This breaks down into relationships between:
- the platform and user
- the user and their generated content
There's always contention on where exactly to draw boundaries here, especially user-authored code in escape hatches. But with agentic AI, these boundaries become even fuzzier, opening a liminal zone of responsibility and ownership.
I think that's why it makes for a great case study in the risks and benefits of integrating agentic AI into a platform.
The Product Relationships Are Changing
When an agent is added to a platform, the product-user relationships becomes meaningfully altered. Before, it was between:
- the platform and user
- the user and their generated content
Now, there's:
- the platform and user
- the platform and agent
- the agent and user
- the agent and user co-generated content
For each of these relationships, there's potential for product quality risks and expectation misalignments. However, agentic AI can magnify the cost, complexity, and time spent on these issues; and while the agent acts as a service provider to both the platform and the user, making autonomous decisions on behalf of both, accountability for its actions and behavior still lies with the platform.
The new greater dynamic is that both the product and user become engaged in a potentially expensive steering exercise from both sides, trying to extract the highest quality output out of a now probabilistic system. In an ideal world, both the product and the user have the same goals in mind, but there can sometimes be a tension, where the user wants something the platform can't provide, and the agent in the middle is responsible for navigating that tension on behalf of the product. This dynamic is a fundamental change and risk to a product that didn't have it before, and needs to be handled very intentionally to maintain trust between the user and the product.
Another more subtle change over time to a product is on the conceptual model. It's tempting to think that it's possible to integrate an agent neatly into an existing or pre-planned conceptual model, but it's rarely that simple in practice. It's almost guaranteed that you'll end up frequently tweaking or loosening your conceptual model to make affordances so you don't "fight the model", which is...as frustrating as it sounds.
The tension between the product and agent becomes giving the agent enough autonomy to deliver value, but within enough guardrails to minimize the myriad of issues that come along with that process, such as tool hallucination or misuse, security, bad or strange behavior, context management, drift issues, rolling back bad output, etc. These are non-trivial challenges from both a system design and UX perspective, and for many of them, there's no perfect solution.
The other trick is ensuring that the agent delivers value to the user consistently. The best way to ensure that today is with evals and user feedback, so in case alignment with the user wasn't a priority before (which of course it should be!), it's absolutely critical now for your evals. This will take a lot of time to tune and maintain them, as the user changes, the models change, and the product changes (they're also not free, so running evals should be factored into costs as well).
Value Propositions, Differentiation, and Margins
The unspoken bet on the value of agentic AI is...somewhat aligned for products and users, being that either now or at some point in the future, the agent will be capable of making expedient, mostly correct decisions and generating valuable content, insights, or performing actions that the user couldn't or wouldn't want to do themselves.
But what is "correct" and "valuable"? In a perfect world, users would know exactly what they're looking for and how to ask for it. We already know this is not reality, as users ourselves, as well as engineers and product people.
I think a good heuristic is looking at how the users interact with an agent. If users are spending time chatting with an agent in the product to resolve ambiguity instead of using the agent to drive the product, it's an expensive signal that the product may be leaning too hard on the agent to express the core value prop, and that there may be tightening to be done with the conceptual model, agent, and/or UX.
This goes back to the aforementioned tension between the platform and the agent. It's a delicate balance to present the user with a consistent, useful conceptual model that helps them navigate ambiguity while also having an agent work within its constraints, and without offloading too much of that ambiguity navigation to the agent. It's a constant exercise in tightening, loosening, and tweaking the product as you experiment with that, and it'll change over time.
I will say that watching user sessions as users navigate ambiguity with an agent is incredibly interesting and useful! However, these conversations cost money; nothing is free...
Forgiveness Threshold
If you watch users interact with generative AI (at least in 2026), you'll see some interesting behaviors in the tolerance they have for a mis-behaving AI. At the end of the day though, every time there is a bug in the system or an agent mis-behaves or underperforms, it is a reflection on the product. While educating a user on prompting, giving them tools and abstractions to help steer agents, or the ability to rate and provide feedback on output are all industry-standard practices right now to help users get better results from an agent and improve the product, they still signal to a user that they need to have some level of patience with your product.
That's fine, again at least in 2026, to ask for some level of patience from the user in regards to AI—but not in regards to the product's capability to deliver on its value prop. It's another delicate balancing act, and requires a lot of honest examination around what the actual value the platform, outside of the agent, is providing.
It's also good to be mindful about this, because it's very unlikely this patience will last. My hypothesis is that fascination, ai insecurity, and experimentation are greatly skewing user's "forgiveness threshold" for AI. But that's likely temporary, and it's worth examining:
- how much of a core value proposition can be realized today without forgiveness from a user
- how heavily a product leans on forgiveness to operate as designed
Beyond the Hype
There are also just some annoying and concrete downsides with external model providers and generative models that are inevitable if you're building an AI-native product:
- cost
- latency, latency, latency
- downtime
- security
- product misrepresentation, quality fluctuations, tool misuse, and potential hallucinations across all three
- limited debugging ability
- level of control beyond a certain threshold
With agentic AI particularly, these can also balloon into time-consuming, expensive agentic garbage loops that your user has to sit through, and even potentially pay for (the product will definitely pay for it, either directly or by the user bouncing if it's particularly egregious).
The simplest takeaway from all of this might be:
AI is just another technology in the arsenal to better deliver on a product's value proposition.
My rule of thumb is, before reaching for generative AI, I always want to be confident the problem genuinely calls for it, and that I have an experience in mind already that would deliver value, but requires this specific technology to do it. Like any other tool I would use. That's a very different approach than first thinking about all the amazing things the technology can do, and then building experiences around that. Though to be clear, there's nothing wrong with experimenting or making a short-term play; but neither form a long-term coherent product strategy, which is mainly what I'm focused on here.
I don't write any of this cynically; I write it because agentic AI in particular carries real costs, risks, and complexities that are easy to underestimate in the very real struggle to remain a relevant product in 2026.
I'm curious on how this plays out in terms of product and engineering quality in the long run, and how users perceive and measure that. Maybe it'll change our expectations for products altogether; maybe it already has. Though, while I am certainly fascinated by AI, my forgiveness threshold isn't as high as it used to be—I sure hope the AI products I use are accounting for that...