Field Notes

Practical writing on AI in product development.

• The Boring ROI: Why AI Is Migrating to the Back Office — The AI deployments paying for themselves are not the customer-facing ones. They are internal, narrow, and boring enough to audit.
• Disposable Software: When the Interface Exists for Ninety Seconds — AI can now generate a single-use view for one task and discard it. That is genuinely useful, and it quietly breaks three assumptions your app depends on.
• From Chatbots to Agentic Workflows: The Verb Changes Everything — Moving from AI that suggests to AI that acts is not a bigger version of the same feature. It changes what you have to build underneath it.
• Inference Is the New Line Item: Budgeting for Agents That Never Stop — Spending has flipped from training models to running them. That turns AI from a project cost into an operating cost, and changes what you have to instrument.
• How Many Agents Is Too Many? Orchestration for Agile Teams — Splitting one agent into several buys precision and costs coordination. Here is how to tell which side of that trade you are on.
• The Abstention Advantage: When the Best AI Move Is No Move — We celebrate AI for what it does. The more useful behavior is often what it declines to do. Abstention is a feature, and it deserves to be designed on purpose.
• A Field Guide to Launch Readiness for AI-First Products — Launch-ready for an AI-first product means more than green tests. It means migration safety, scoped access, graceful degradation, and a real P0 list.
• Tracing a Single AI Decision: The correlationId Pattern — When an AI agent does something surprising, the first question is always \"what happened, exactly?\" One thin thread — a correlationId — is how you answer it.
• The Rubber-Stamp Trap: Beating Automation Bias in Human Review — A human-in-the-loop system is only as good as the human. When reviewers start approving everything on autopilot, the safety net becomes theater.
• The Warm-Professional Voice: Localizing AI Products for Global Teams — Your product voice has to survive translation and AI generation both. Here's how to keep a warm-professional tone consistent across languages.
• Designing an Approval Queue Reviewers Actually Use — The approval queue is where human judgment meets AI autonomy. Treat it like a product surface, not a to-do list, and the whole governance model comes alive.
• Feature Flags for AI: Shipping Autonomy Safely — Autonomy is a dial, not a switch. Feature flags let you ship AI capabilities incrementally and kill a misbehaving surface in seconds.
• Progressive Autonomy: A Maturity Model for AI Agents — Trust isn't granted on day one; it's earned in increments. The agents that end up doing the most are the ones that started by doing almost nothing.
• Measuring AI Product Impact: DORA, Cycle Time, and Beyond — AI usage numbers tell you activity, not impact. Pair them with DORA and cycle time to prove your agents actually move delivery forward.
• When to Let AI Close a Ticket (and When Not To) — Closing a ticket is the one AI action that is hard to undo. The right policy is not to forbid it, but to make autonomous closes earn extra scrutiny.
• Prompt Ops: Treating Prompts as Product Infrastructure — Prompts shape AI behavior as much as code does, yet most teams bury them as hardcoded strings. Treating them as versioned data changes who can improve them, and how fast.
• MCP and the Governed Tool Hub: Giving Agents Safe Superpowers — The Model Context Protocol lets agents use real tools. A registry, a gate pipeline, and workspace-minted tokens are what keep that power from becoming a liability.
• Cost Caps and Circuit Breakers: Keeping AI Spend Sane — AI spend fails open by default. A daily cap, a per-agent circuit breaker, and per-user metering turn an unbounded bill into a governed budget.
• OKRs Meet AI: Aligning Autonomous Work with Business Outcomes — An agent optimizing for ticket throughput can quietly work against your goals. Alignment means connecting autonomous work to the outcomes you actually want.
• Designing the Aha Moment for AI-Native Onboarding — AI onboarding fails when it explains instead of demonstrates. The aha lands when a new user watches governed AI do real work in minutes.
• Multi-Tenancy and AI: Keeping Customer Data Truly Isolated — AI agents make cross-tenant leaks easier to cause and harder to see. Isolation has to be enforced at the boundary, not requested in a prompt.
• The Agentic Mesh: Coordinating Multiple AI Agents on One Product — One monolithic agent buckles under a real product's variety. A mesh of specialists that delegate and hand off scales better, and stays legible.
• Confidence Scores: Teaching Product Teams to Trust (and Distrust) AI — The goal isn't blind trust in your AI or blanket suspicion. It's calibration, and a well-placed confidence score is how you teach it.
• Curating AI Actions: Why Fewer, Better Automations Win — The instinct to let AI do everything is exactly backwards. A curated allowlist of high-confidence actions beats a sprawling one every time.
• Building an Audit Trail for AI Decisions — When an AI agent makes a call, you need to reconstruct who, what, and why after the fact. A correlated decision ledger makes that possible.
• SLA Escalation When Your AI Agents Miss the Mark — AI agents will occasionally stall on the hard cases. A disciplined SLA escalation layer turns those silent misses into visible, accountable handoffs.
• Risk Scoring for AI Decisions: A Practical Formula — Not every AI decision deserves the same scrutiny. A simple, transparent risk formula turns \"should a human look at this?\" into a number you can act on.
• The Human-in-the-Loop Approval Pattern Every AI Product Team Needs — Human-in-the-loop only works if the queue respects human attention. The pattern that scales aggregates, filters, and prioritizes what a person sees.
• From Backlog to Breakthrough: Letting AI Triage Your Ticket Queue — Triage is the tax every team pays before real work begins. An AI agent that proposes labels, routing, and next actions can hand that time back.
• Why AI in Product Development Needs Guardrails, Not Just Horsepower — Raw model horsepower is the easy part. The teams that win with AI build guardrails that make agent decisions legible, risk-scored, and reversible.