AI is technology, not a feature
Many companies today are facing increasing pressure to adopt AI and launch AI-based features.
While AI is a transformative new technology, such pressure can be unhealthy and counter-productive. Faced with hyped-up external expectations — from customers, investors, or the general public — companies often feel the need to do something to show they're keeping up with the competition and are not being left behind.
This leads to two worrying trends:
- implementing internal AI workflows (in software development and elsewhere) without a clear strategy
- building and launching "AI features" without much consideration how they fit with the rest of their product
Challenges with internal AI adoption have been documented in detail, and I have previously written about how a realistic AI adoption strategy is critical for the transformation to be a success. In this article, I'm focusing on products and AI features.
Chatbots everywhere
The strategy is scarcely better on the product side. Companies rush to "add AI" and the first idea that pops to mind is a chatbot.
Chatbot was the original use case that popularized the new AI technologies. Before ChatGPT LLMs were a niche curiosity, getting better over time but without accompanying hype. After ChatGPT, and especially as the APIs were opened and competition arrived, everyone started building chatbots.
So it's no surprise that "chatbot" is the first answer that pops into mind when someone asks "what AI feature should we build?" Add a chat UI, connect some data to it, and you're ready to go! Many companies already provide turnkey solutions you can easily plug right into your data and product.
The first generation of chatbots were plain wrappers around the LLM APIs. The next added retrieval-augmented generation (RAG) — LLM paired with external data sources. The latest (as of 2026) are agentic assistants: LLMs that can use tools to execute some action.
They're still chat-oriented, with poor UX affordances ("what can you do with the chatbot?"), can be unreliable (hallucination, poor alignment), prone to misuse (prompt injection, jailbreaks), and tricky to get to a good standard of quality.
There are some good uses of chat-based agents: customer support, analyzing and explaining complex data in an user-friendly way, and other scenarios where a two-way discussion is a natural way to solve the problem.
But if your users have to think about the magic incantation they need to type into the chat box to get the result they want — chatbots are a poor fit.
Enhance with AI
The other popular trope for an AI feature is "Enhance with AI". A button that generates or transforms users' texts, images, videos or even math formulas, so that they're "better". What counts as "better" is context-dependent, but many implementations work so poorly that a "Make mediocre with AI" label would be more apt.
The problem is not that AI is bad as a technology. More often, it's a poor fit for the specific use case, is used in an inappropriate way, or just haphazardly slapped together so the product can check off the "has AI" checkbox.
Feature vs technology
Beyond checking the checkbox, there are few benefits to such approaches, and many drawbacks. The fundamental mistake here is to treat AI as a feature, while in fact it's a technology.
A feature is something the product does: it sends emails, or prevents spam, or lets you watch short videos. A technology is how the product does it: implements standard email protocols, does Bayesian (or AI) based spam filtering, or processes and distributes video uploads and does H.265 video decoding.
Good technology is invisible — you only notice it when something is wrong. You don't care about H.265, HLS, CDN, ELB, FFMPEG and other techno-babble when you want to watch Netflix. You don't care about SMTP, IMAP, DKIM or SPF when you want to send email.
AI is a technology: you don't care about LLMs when you want customer support or to proof-read your text. As a technology, AI may unlock many new and different possibilities, most of which we're not even yet aware of. It may improve existing features, by providing a different, complementary and better way of doing things. And where it isn't a good fit, it shouldn't be used.
Tech can be hype
As explored in Hype AI vs pragmatic AI, it is both a normal technology, and a marketing term. And indeed the pressure to adopt is purely hype-driven. As a company, you don't want to silently apply AI to great effect — you want to brag publicly about it!
To the extent such bragging is backed up by good implementation, there's nothing wrong with it. However, people are great at bending and stretching the hype for maximum effect, even if the real benefit is modest, or the implementation slapped together.
It's possible the pendulum might swing in the opposite direction. With so much AI slop sloshing around, it's increasingly associated with poor quality and tastelessness. Before long, companies may want to be more discreet about the AI tech they're using, and focus again on the features, and even more importantly, benefits.
Start with the outcome
This leads me to the better strategy for integrating AI into software products: as with any feature, start with outcomes, and work your way backwards.
What's the benefit, the desired outcome, for the user and for the company? What existing or new features of the product can support it? What's the best way to implement these features?
AI requires us to rethink what's possible, but shouldn't be shoehorned where it doesn't make sense.