ChatGPT captured the world’s imagination, but it may have also trapped it. The chatbot interface—with its familiar conversational format—made AI accessible to millions, demonstrating the remarkable capabilities of large language models (LLMs) in a package that felt natural and inviting. Yet this very success has created a misconception: that AI equals chatbots, and that every application needs a chat window to be AI-powered.
The reality is more nuanced. ChatGPT succeeded not just because of its underlying technology, but because it brilliantly matched interface to capability. By packaging AI in a conversational format, OpenAI created a product where errors were acceptable—even expected. Users could correct misunderstandings, refine prompts, and iterate toward better answers. The chatbot became the perfect vehicle for technology that was inherently probabilistic and occasionally wrong.
But what works for general-purpose exploration doesn’t translate to domain-specific business applications. When companies rush to add chatbots to their products simply to appear AI-forward, they often create more problems than they solve. The impulse is understandable: executives want to demonstrate AI adoption, and chatbots seem like the fastest path. Technically, implementation can be straightforward—connect to an API, add a chat interface, and declare victory. But this approach typically delivers minimal value while expanding risk exponentially.
A chatbot embedded in a business analytics platform suddenly needs to handle not just data queries but also random tangents that may have nothing to do with the core product. If the integrated LLM can deliver accurate answers only 80% of time, the surface area for errors explodes. Information doesn’t arrive the moment users need it. The interface becomes a distraction rather than an enhancement, satisfying executives while frustrating actual users.
The User Experience Revolution
The real opportunity lies in rethinking how AI integrates into workflows rather than bolting on generic chat interfaces. Working to realize this opportunity will require classic product discipline: understanding jobs to be done, making sense of complex data, and presenting information alongside relevant actions at precisely the right moment. AI should make these experiences better, not worse. The interaction surface should become narrower and more focused, not broader.
Consider the evolution of AI coding assistants. While LLMs have become somewhat commoditized, the winners in this space distinguish themselves through superior user experience. They have embedded AI directly into developers’ existing workflows—providing real-time suggestions while typing code, allowing developers to guide the AI with simple configuration files, and integrating seamlessly with familiar tools. Chat elements exist, but they’re not the only interaction mode.
The massive opportunity lies in taking existing LLM capabilities and integrating them into domain-specific workflows in narrow, targeted ways rather than widely deploying generic chatbots horizontally.
The Agentic Evolution
The next phase—agentic AI—amplifies this need for thoughtful UX design. Agents can reason through complex tasks by breaking them into smaller components and can use tools to act on users’ behalf. An agent might research options, make travel booking arrangements, or complete transactions autonomously, escalating to humans only when guidance is needed.
But agentic capabilities don’t dictate a single interface paradigm. The tools they integrate with, the information they present, and the interaction modes they employ will vary dramatically based on domain-specific requirements and user needs. Consider an AI agent designed to assist with travel bookings vs. an agent designed to assist with enterprise information security. Even though both leverage generative AI, the travel agent is likely to present information in ways that look very much like popular travel websites. Imagine a highly visual interface presenting you with a prompt to “please select from the three hotels that met your price criteria and itinerary.” Whereas the infosec agent is likely to convey data-intensive communications on incidents or indicators of compromise much the way today’s enterprise IT security platforms do: “here’s your sev2 security breach incident report.”
Why Narrow Solutions Win
The path to adoption favors narrow, vertical-specific AI applications over broad, horizontal platforms. For Enterprise, benefiting from AI isn’t primarily a technology challenge—it’s a change management challenge. Enterprise AI adoption stumbles, in part, because the technology is probabilistic and sometimes inaccurate rather than the deterministic and precise technology we are used to adopting. When an AI system is 90% accurate, extracting value requires careful process design and gradual integration alongside human co-pilots. Organizations struggle to redesign workflows across departments, especially when those workflows have been optimized for human workers over decades.
Adoption challenges are exacerbated as efficiency gains typically result in each employee doing more “thinking” work not less: Programmers adopting AI tools often comment how much more exhausted they are because the routine work that allows their minds to relax between deep thought no longer exists (it has been automated away). The best method for adopting “Human in the loop” AI solutions that empower rather than exhaust is still very much a work in progress.
Customer support provides a telling example. AI could handle 80% of repetitive inquiries, but the remaining 20% would require human expertise and therefore carries high error costs. Simply replacing an entire team isn’t viable. The change management challenge becomes insurmountable without careful UX design that supports hybrid human-AI workflows.
Narrow solutions succeed because they’re easier to adopt. A focused sales assistant agent has a clear user, a distinct role, a defined path for escalation to humans, and measurable impact. Getting local adoption within a specific function proves far more achievable than top-down enterprise-wide AI initiatives.
Building for the Real Future
The companies that will win the next wave of AI applications won’t be those with the best models or the most parameters. They’ll be those that build exceptional user experiences tailored to specific domains and workflows. This means:
- Deep integration with existing tools and systems rather than standalone interfaces
- Information and actions presented in context, at the moment of need
- Workflows designed around AI’s probabilistic nature rather than fighting it
- Domain-specific features that solve real problems rather than generic capabilities
The narrow approach puts you into position to very quickly establish a user feedback / data flywheel that is essential to creating ever more seamless experiences, and the opportunity to lock in loyalty. It also means building beyond the core AI functionality—handling middleware, compliance, permissions, security, and pricing models that make expensive AI technology economically viable.
The future of AI isn’t about chat windows. It is about invisible intelligence woven seamlessly into how people work, making complex tasks simpler and tedious work disappear. That future requires rethinking user experience from the ground up, not retrofitting chatbots onto existing products. The winners will be those who recognize this distinction and design accordingly.
As we move into a multi-modal future the need to re-think and innovate in human computer interaction models will only become greater. Most examples of this technology today feel clunky or gimmicky – but there is no doubt in my mind that we are on the path to ubiquitous compute, and the interaction models invented and adopted over the coming years will shape the human experience for decades to come.
