The latest articles on Forem by Marzun (@marzun9620). https://forem.com/marzun9620 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3850020%2Fc039a311-e909-4540-8713-93dd899267bc.png https://forem.com/marzun9620 en Marzun Sun, 29 Mar 2026 20:38:48 +0000 https://forem.com/marzun9620/building-an-ai-assistant-that-doesnt-take-action-automatically-32md https://forem.com/marzun9620/building-an-ai-assistant-that-doesnt-take-action-automatically-32md <p>Most AI apps today try to <strong>do things for you automatically</strong>.<br> But when it comes to personal data — like money or health — that can feel risky.<br> So I explored a different pattern:</p> <blockquote> <p>What if AI <em>suggests</em> actions, but never executes without approval?</p> </blockquote> <h2> 🧠 The Idea: Approval-Gated AI </h2> <p>Instead of this:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight plaintext"><code>User → AI → Action happens </code></pre> </div> <p>I built around this:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight plaintext"><code>User → AI → Draft → User approves → Action happens </code></pre> </div> <h2> 💬 Example </h2> <p>You type:</p> <blockquote> <p>"Add coffee for 40"</p> </blockquote> <p>The AI:</p> <ul> <li>Parses the intent</li> <li>Prepares the expense entry</li> <li>Shows it to you</li> </ul> <p>You:</p> <ul> <li>✅ Approve → saved</li> <li>❌ Reject → nothing happens</li> </ul> <h2> ⚖️ Why This Matters </h2> <p>In domains like:</p> <ul> <li>Finance</li> <li>Health / nutrition</li> <li>Personal logs</li> </ul> <p>Users care about <strong>control and accuracy</strong>.</p> <h3> Problems with full automation </h3> <ul> <li>Silent mistakes</li> <li>Wrong assumptions</li> <li>Loss of trust</li> </ul> <h3> Problems with manual apps </h3> <ul> <li>Too slow</li> <li>Too much friction</li> </ul> <h2> 🔄 This Pattern Sits in Between </h2> <p><strong>Approval-gated AI is:</strong></p> <ul> <li>Faster than forms</li> <li>Safer than automation</li> </ul> <h2> 🧪 Where I Applied It </h2> <p>I've been prototyping this idea in a small app where AI helps with:</p> <ul> <li>Expense and income tracking</li> <li>Budget alerts</li> <li>Nutrition logging (calories and macros from descriptions)</li> <li>Receipt parsing</li> </ul> <p>Everything goes through <strong>user approval first</strong>.</p> <h2> 🤔 Trade-offs I Observed </h2> <h3> ✅ Pros </h3> <ul> <li>Builds trust quickly</li> <li>Prevents unexpected changes</li> <li>Makes AI behavior transparent</li> </ul> <h3> ❌ Cons </h3> <ul> <li>Adds an extra step</li> <li>Can feel repetitive for simple actions</li> </ul> <h2> 🧩 Open UX Questions </h2> <p>This is where I'd love input:</p> <ul> <li>Should small actions auto-approve?</li> <li>Should users be able to set confidence thresholds?</li> <li>When does confirmation become annoying instead of helpful?</li> </ul> <h2> 🚀 Implementation Stack </h2> <ul> <li>React 19</li> <li>Hono</li> <li>Cloudflare Workers + D1 (SQLite)</li> <li>Drizzle ORM</li> <li>Tailwind</li> <li>Radix UI</li> </ul> <h2> 🔗 Demo &amp; Project </h2> <ul> <li>🎥 Demo: <a href="https://www.youtube.com/watch?v=V09mr6b7RXU" rel="noopener noreferrer">https://www.youtube.com/watch?v=V09mr6b7RXU</a> </li> <li>🌐 App: <a href="https://keepmylog.com" rel="noopener noreferrer">https://keepmylog.com</a> </li> </ul> <h2> 💬 Final Thought </h2> <p>AI doesn't always need to be autonomous.<br> Sometimes the best UX is:</p> <blockquote> <p>AI that assists decisively — but lets humans stay in control.</p> </blockquote> <p>Curious how others are thinking about this pattern.</p> ai webdev saas productdesign