# Don't Just Slap on a Chatbot: Building AI That Works Before You Ask Arthur Objartel, Evil Martians | AI Engineer Summit 2025 | 05:46 Source: https://www.youtube.com/watch?v=2cEGQEllBGc Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/dont-just-slap-on-a-chatbot-building-ai-that-works-before-you-ask Published: 2025-02-22 Tags: copilots, design, human-in-the-loop ## TL;DR - AI can watch work in progress and offer specific help without waiting for a chat request. - Proactive assistance should preserve user control, make recommendations, and stay inside the user's normal workflow. - Teams can find useful AI opportunities by studying friction points, repeated questions, and places where users get stuck. ## Summary Arthur Objartel argues that adding a chat interface is often the easiest way to add AI, but it does not always help users. While designing Tegon, an AI-powered issue tracker, his team built assistance that observes work as it happens. It can ask contextual questions while someone reports an issue, suggest how to split complex work into smaller issues, and consider existing data, timelines, and resources. Objartel calls these suggestion mode, action mode, and question plus action mode. The system stays inside the workflow, and users can revert its changes with one click. He gives three design rules: AI should support user agency, offer recommendations rather than force them, and avoid interrupting the workflow. He also suggests looking for friction points, repeated requests, and moments where users get stuck. The talk makes a practical case for experimenting with interfaces that do more than wait for prompts. ## Key ideas ### Chat is an easy default, but it may not help users [00:00](https://www.youtube.com/watch?v=2cEGQEllBGc&t=0s) Arthur Objartel opens by questioning the rush to add chatbots to products. A chat interface lets a team say it has AI, but users still have to stop their work, explain the situation, and ask for help. While working on Tegon, an AI issue tracker, he chose the opposite approach. The product would observe the task and help at the point where assistance was useful. His argument is about product behavior rather than the presence of a particular model or chat window. ### Clippy had the right interaction idea at the wrong time [00:46](https://www.youtube.com/watch?v=2cEGQEllBGc&t=46s) Objartel uses Clippy as an example of an old product idea that was badly executed and appeared before the technology was ready. Clippy did not wait for a user to open a help panel and formulate a request. It watched the work and offered assistance. Objartel thinks current systems can revisit that interaction with better context and timing. The useful part is the proactive behavior, while the intrusive execution is something current product teams need to avoid. ### Suggestion mode asks contextual questions while the user writes [01:49](https://www.youtube.com/watch?v=2cEGQEllBGc&t=109s) In Tegon, the AI watches someone report an issue in real time. When it sees a legality-related problem, it starts asking specific questions that move the report forward. These are not generic prompts such as "how can I help?" The product calls this suggestion mode. The assistant appears inside the existing work rather than opening a separate chat window. Its value comes from recognizing what the user is writing and choosing a relevant moment to ask for missing information. ### Action mode proposes a better issue structure [01:58](https://www.youtube.com/watch?v=2cEGQEllBGc&t=118s) When a user writes an issue that can be divided into sub-issues, Tegon identifies that complexity and suggests a different organization. The AI uses previous data to understand how similar work was handled, so it is not simply splitting text at random. Objartel calls this action mode. The assistant can also think about timelines and resources, extending the interaction from editing a description to helping manage the work around it. ### Proactive help can combine questions and actions inside the workflow [02:44](https://www.youtube.com/watch?v=2cEGQEllBGc&t=164s) Objartel describes a question plus action mode in which the AI asks for information and helps change the issue at the same time. The interaction happens in the natural flow of work, without a context switch, extra window, or chat interface. Users can revert the changes with one click. That control matters because the system is acting on the user's work, while the user still decides whether the proposed changes should remain. ### Useful proactive AI supports agency and makes recommendations [03:33](https://www.youtube.com/watch?v=2cEGQEllBGc&t=213s) Objartel gives three rules for this product pattern. AI should supplement user agency instead of replacing it. It should offer recommendations rather than force them. It should be part of the natural workflow rather than stopping it. He applies the idea to code editors that watch for common pitfalls, tools that suggest accessible design choices during creation, and communication tools that prepare meeting context or find a document mentioned on a call. ### Friction, repeated questions, and stuck points reveal where to add assistance [04:38](https://www.youtube.com/watch?v=2cEGQEllBGc&t=278s) For teams designing their own products, Objartel suggests starting with moments where users have to stop working to ask for help. Repeated questions and recurring behavior patterns can point to useful automation. Context matters as well. The places where users get stuck may be where AI can provide the most help. These observations give product teams a way to find opportunities before deciding that a chatbot is the interface. ### AI interface design still needs experiments beyond copied chat screens [04:59](https://www.youtube.com/watch?v=2cEGQEllBGc&t=299s) Objartel says AI interface design is still taking shape and does not yet have a fully formed playbook of best practices. He argues that copying and pasting chat interfaces into every product is not enough. Teams should experiment with unexpected user interfaces and challenge the assumption that every AI feature needs a conversation window. His examples point toward assistance that appears during work, at a moment when its context is available. ## Notable quotes - "What if AI didn't wait to be asked? What if it just knew what you needed?" (01:09) - "Specific contextual questions that actually move the work forward." (01:27) - "AI should supplement user agency, not replace it." (03:33) - "AI should offer recommendations, never force them." (03:33) - "AI should be part of natural workflow, not stop it." (03:33) ## Tools & references mentioned - Evil Martians - Tegon - Clippy - Microsoft Word ## Who should watch - Product and UX teams deciding whether a new AI feature really needs a chat interface. - Engineers building issue trackers, editors, design tools, or communication products that can observe work in context. - Teams looking for practical places to add AI without taking control away from users. ## Related talks - [Designing AI To Scale Human Thought](https://aietalks.com/talks/designing-ai-to-scale-human-thought) (Jun Yu Tan, Tusk, 12:24) - [Form factors for your new AI coworkers](https://aietalks.com/talks/form-factors-for-your-new-ai-coworkers) (Craig Wattrus, Flatfile, 15:35) - [Build Dynamic Products, and Stop the AI Sideshow](https://aietalks.com/talks/build-dynamic-products-and-stop-the-ai-sideshow) (Eliza Cabrera, Workday & Jeremy Silva, Freeplay, 18:10) - [Rise of the AI Architect](https://aietalks.com/talks/rise-of-the-ai-architect) (Clay Bavor, Sierra & Alessio Fanelli, Latent Space, 18:55) - [You Can't Prompt the Room: The Last Skill AI Won't Replace](https://aietalks.com/talks/you-cant-prompt-the-room-the-last-skill-ai-wont-replace) (Balázs Horváth, VisualLabs, 15:45)