AI Bolted On

AI bolted on is the default 2024 shipping pattern for any enterprise team that wants AI in the launch deck but not in the actual architecture. The product team leaves every screen exactly as it existed before LLMs existed. Then they add a chat panel docked to the right edge that users must open manually. Or they drop a sparkle icon next to every input field that promises to rewrite or summarize the content. The model operates in its own silo with limited access to the rich state of the main interface. Users try the AI once or twice. They notice the responses lack the context that would make them valuable. They close the panel and return to the old flows that still work perfectly. The AI becomes a costly afterthought that drives no meaningful engagement. The pattern repeats across CRMs, project management tools, help desks, and analytics platforms. Each one ships the same sidebar. Each one watches the usage metric flatline. The deletion test catches it every time. Delete the model and the product remains fully operational which is the entire indictment.

This is not what AI-native products do. AI-native products make the model the primary surface so the UI exists only to feed it input and display its output in accountable ways. Cursor is AI-native because the editor surface is meaningless without the model driving autocomplete, chat, and agentic edits. Remove those and nothing usable remains. AI bolted on survives deletion without a scratch because the original UI was never touched. The principles expose the difference in sharp relief. Bolted-on products default to side panel instead of core surface so the model is the last thing users reach for. They keep form-as-input dominant instead of replacing forms with prompt-as-input that lets users speak naturally. They choose permission by default with constant confirmation dialogs instead of agency by default that acts first and provides undo paths. They ship magic outputs without transparency surfaces that show context and sources. They expose model pickers on consumer screens instead of hiding details by default. They treat latency as an engineering bug instead of designing the entire rhythm around streaming and skeleton states from the first prototype. Linear threads the needle on some surfaces by embedding AI in the command bar power users already love. Most other tools fail every principle and wonder why their AI features collect dust.

Take the CRM that shall not be named but shipped a prominent AI sidebar in summer 2024 as your concrete example. The main lead management dashboard stayed identical to its 2022 version. The sidebar promised context-aware assistance but could not reliably pull the full customer history or understand the stage of the sales pipeline. Salespeople opened it during onboarding then abandoned it for the traditional note-taking fields that felt more reliable. The same pattern appeared in the knowledge base tool that added AI article generation as a side panel. Writers ignored it because the generated drafts lacked the brand voice and internal links that the regular editor made easy. The analytics platform that bolted summarization onto every chart delivered generic insights that ignored custom filters and metrics users had already set. These failures stand next to the native successes. Cursor weaves the model into every interaction so the product feels broken without it. Granola uses the model to augment meeting notes silently then displays the raw transcript next to the augmented version for instant verification. Perplexity builds the entire search experience as a model surface with citations embedded in every answer and fast streaming responses. Arc Search turns the browser into an AI agent that delivers synthesized pages instead of link lists. Linear embeds AI in the command palette so it feels like a natural extension rather than a bolted panel. Each native example fails the deletion test in the best way possible. The bolted-on examples pass it and expose their own shallowness.

Deploy AI bolted on when you need to satisfy a checkbox on a requirements document or when you want impressive looking screenshots without doing the hard work of redesign. It fits when the AI is truly secondary and the main value lives in the pre-existing workflows. Never deploy it when you need users to adopt the AI as part of their daily habits or when your competitive positioning depends on being perceived as AI-first. The approach guarantees the feature will be ignored past the first session. It wastes compute costs on a panel nobody opens. It trains users to treat your AI as noise. Use the native approach instead when you have time to run the deletion test on every screen and replace forms with prompts where it makes sense. The teams that commit early to the six principles ship products that feel alive. The teams that bolt on AI spend the next year running experiments to drive engagement on features that were doomed from the architecture phase.

AI bolted on turns the most powerful technology of the decade into a sparkle icon that users learn to ignore within a single afternoon.