ChatGPT calculates Net Promoter Score correctly about 70% of the time in my BYU classroom. The other 30%, it either miscategorizes responses or just makes up a number that sounds right.

The dangerous part? Half the time it fails, it still lands on the correct answer through completely wrong reasoning. It’ll count 7s and 8s as promoters, miscategorize a few responses, then arrive at the right NPS anyway because the errors cancel out.

You catch and fix a wrong answer. A right answer from flawed reasoning? You trust it. You build on it. You make decisions with it.

This isn’t just about NPS. It’s about how we should and shouldn’t use AI in systems where getting it wrong costs money and customers.

Why NPS Breaks Language Models

The calculation is dead simple. Subtract the percentage of detractors (0-6) from the percentage of promoters (9-10). Passives (7-8) don’t count toward the score, though they tell you plenty about how fast customers will jump ship if someone better shows up.

Something about those category boundaries confuses models when they’re generating text instead of running code. Modern frontier models mostly solve this by switching to code execution when they spot math. But older models, smaller models, and chat-only setups still slip. Even the best models occasionally hallucinate plausible numbers when they’re in the wrong mode.

The fix: make your AI compute with code, not generate the answer as text. When I build automated feedback systems, I calculate NPS twice. Once through AI, once through direct code. Any discrepancy triggers a flag.

Trust but verify isn’t paranoia. It’s how you keep a system from lying to you at scale.

What Customer Feedback Actually Tells You

Most companies treat NPS as a satisfaction metric. It’s not. It’s a leading indicator.

Declining NPS shows up in feedback months before it shows up in revenue. By the time finance sees the churn line bend, customer success often had the signal in detractor comments ninety days earlier. By the time sales feels pricing pressure, NPS for price-sensitive segments typically drifted downward two quarters before.

This is what makes feedback systems strategic infrastructure instead of feel-good hygiene. When LexisNexis acquired my previous company, NPS was part of the valuation. Sophisticated acquirers pay for feedback systems that produce high scores because they represent more than happy customers. They represent an early-warning system and documented proof you can listen and respond.

But only if the system works. Only if the numbers are real and the feedback connects to decisions that change things.

AI transforms the economics of every stage. Survey design, multi-channel collection, sentiment analysis, theme extraction, routing to the right teams. What used to require a dedicated analyst and manual compilation every week now runs continuously without supervision. Every response gets processed, categorized, stored. Urgent detractor feedback triggers alerts. Leadership gets synthesized intelligence every Monday morning.

The system I describe in the book uses Google Forms or Qualtrics for collection, n8n for automation, AI for analysis, Supabase for storage. The full workflows and prompts are at my companion resources. The specific tools matter less than the pattern: systematic collection, intelligent analysis, appropriate distribution, decisive action, closed loops with customers who gave feedback.

Add Verification Today

If you’re using AI to analyze customer feedback right now, add verification. For any numerical calculation, make the AI show its work step by step. For NPS specifically, make it count each category before calculating percentages.

Better yet, calculate the same number twice through different methods and flag discrepancies. The extra thirty seconds of processing time beats trusting plausible-looking numbers from flawed reasoning.

The principle applies beyond NPS. AI handles volume. Humans ensure accuracy. The pilot doesn’t let autopilot land without watching the instruments.

Customer feedback systems are your nervous system. They tell you if your strategy works, where competitors gain ground, which customers are at risk, what product changes matter most. Getting the infrastructure right turns customer voice into strategic advantage.

The full framework is in Chapter 18 of AI Strategy for Business Leaders. The five-stage feedback flywheel, competitive intelligence tactics, how to connect feedback to actual decisions. You can read more about my background building and selling companies that treated feedback as infrastructure, not theater.

Start simple. A basic NPS survey with manual analysis beats no system. Then automate, systematize, scale. Just make sure the math is right.