AI Product Design Patterns

The most expensive PM mistake in AI: automating the wrong thing

A team shipped an AI feature that automatically sent refund confirmation emails on behalf of support agents. No human review. The AI drafted it, the AI sent it. Efficient, right?

Then the AI approved a $4,800 refund for a customer who was only owed $48. By the time anyone noticed, the money was gone.

The mistake wasn't the AI. The mistake was automating something that needed human oversight. And it's the most common PM error in AI product design — because automation is exciting, and "the AI handles it" sounds great in a demo.

(Illustrative scenario. Incidents of AI systems processing incorrect refund or payment amounts due to insufficient human oversight have been documented across multiple industries.)

The automation spectrum: three modes

Every AI feature lives somewhere on this spectrum:

None of these is better than the others. Each fits a different context. The PM's job is to pick the right mode for each feature — and the decision comes down to two questions.

Beyond the automation spectrum, most AI product features fall into one of four underlying patterns — and knowing the pattern tells you what the core UX challenge will be:

Search & Retrieval (find): surface the right thing from a large corpus — main challenge is relevance. Generation (create): produce new content — main challenge is quality and voice. Classification (organise): sort inputs into categories — main challenge is accuracy and edge cases. Recommendation (discover): suggest next actions — main challenge is trust and explainability.

The two questions that decide everything

(In the previous module, you learned that hallucination is a structural property of LLMs — it can't be eliminated, only mitigated. That's exactly why the stakes-and-reversibility framework matters: some mistakes can be caught and fixed before anyone is harmed; others can't. The mode you choose determines how much risk you're building in.)

For every AI action in your spec, ask:

Question 1: What happens if the AI gets this wrong?

• "The ticket gets a wrong tag" → Low stakes (easily fixed)
• "The customer gets a wrong refund" → High stakes (money is gone)

Question 2: Can the mistake be undone?

• "An agent re-tags the ticket in two clicks" → Highly reversible
• "The refund email was already sent" → Not reversible

Now plot it:

The default: When in doubt, start with augmentation. It's the safe middle ground. Earn trust with data (high accuracy over time), then consider moving toward automation.

Real example: Maya's three workflows

Maya, a support ops PM at an e-commerce company, handles 4,000 tickets/week. Agents spend 40% of their time on tagging, drafting, and routing. She wants to reclaim that time — but not by introducing new errors.

Workflow 1: Auto-categorise tickets → Automation

Stakes: Low — a miscategorised ticket gets re-routed within minutes, customer never notices. Reversible? Yes — an agent fixes the tag in two clicks. Mode: Full automation. The AI tags every ticket without human review.

Workflow 2: Refund confirmation emails → Augmentation

Stakes: High — this is a financial transaction with legal implications. Reversible? No — an incorrect email cannot be unsent. Mode: Augmentation. The AI drafts the email. The agent reviews it and clicks Send.

Result: Average handle time dropped from 8 minutes to 3 minutes per ticket. Refund error rate: zero. The agent isn't writing from scratch anymore — they're reviewing and editing. Huge time savings, no new risk.

The turning point: An engineer flagged that auto-sending refund emails would mean the AI could commit the company to a payout without any human review. That single conversation changed the spec — exactly the kind of decision that belongs in a requirements doc, not a post-launch retro.

Workflow 3: Complex complaint responses → Inspiration

Stakes: Medium — a bad response damages the customer relationship. Reversible? Sort of — you can send a follow-up, but first impressions matter. Mode: Inspiration. The AI generates a list of 10 possible responses. The agent picks one, rephrases it in their own words, and sends it.

Why inspiration, not augmentation? The agent hasn't seen this complaint type before and needs ideas, not a draft. The AI contributes options; the human makes every decision.

Confidence-aware UI: show uncertainty when it matters

One more pattern PMs need to know: when the AI is less confident, the UI should reflect that.

The trap: Showing every response with the same confidence level. A 95%-confident answer and a 50%-confident answer should NOT look the same to the user. If they do, users either over-trust the uncertain answers or under-trust the confident ones. This is the core of what's called automation surprise — when a system produces a confidently wrong answer and the user has no signal that it failed. Automation surprise is prevented by explicitly specifying graceful degradation paths in your spec and never allowing silent substitution of lower-quality outputs.

Key takeaways

• Two questions decide the mode: "What happens if the AI is wrong?" and "Can the mistake be undone?" Low stakes + reversible = automation. Everything else = augmentation or inspiration.
• Start with augmentation. It's the safe default. Move toward automation only when you have data showing high accuracy, low error rate, and high reversibility.
• Financial, legal, or irreversible actions always need a human in the loop. No exceptions, no matter how good the model is.
• Match UI confidence to model confidence. Don't show uncertain answers the same way you show confident ones.