So I was reviewing a client's new "customer sentiment" pipeline built in Make.com. They were bragging about how it automatically routes support tickets based on positive/negative sentiment from transcribed calls.
My immediate thought: sentiment analysis is famously squishy. A score from some third-party AI module is now a business logic trigger? That's a recipe for hilarious misfires.
They showed me the setup. It's actually pretty straightforward, which is almost the problem.
* You get a sentiment score (often -1 to +1) from an AI service like OpenAI or Google's NLP.
* Make lets you set up a router with numeric filters. Score > 0.3? "Positive" branch. Score < -0.3? "Negative" branch. Everything else is "Neutral."
* These branches then trigger different workflows: send to a senior support agent, notify a CSM, add to a "churn risk" spreadsheet.
The concerning part is everyone treats the score as gospel. I asked the obvious questions:
* What's the confidence threshold? A score of 0.31 is barely different from 0.29, but it'll take a different path.
* How are you validating the accuracy? False positives (angry customer tagged as positive) are worse than no automation at all.
* Where's the human oversight loop? I didn't see one.
It's a powerful feature, I'll give them that. But the implementation I saw was brittle. They're using it for priority routing without any sanity checks. Sentiment is useful as a *signal*, not a *switch*. You'd want to combine it with other triggers—like ticket keywords, customer tier, or issue type—before you automate a workflow that could piss someone off.
Anyone else seen this in the wild? What safeguards did you build in?