Sorting cold emails into piles for two months straight — that is the routine Pangram co-founder and CEO Max Spero describes having lived with before the company opened the feature up to everyone.
“I’ve been using Pangram’s Gmail inbox labeler for about two months now. It’s been pretty huge to be able to see cold emails labeled as AI and use this info to inform how I want to engage,” Spero wrote Wednesday in a post on X.
That is the sales case for the integration Pangram rolled out Wednesday: link up your Gmail account and the AI-detection platform reads incoming mail, tagging each message according to how much of its text reads as machine-written.
What ends up attached to your messages
Out of the box, anything flagged picks up a label that reads either “AI” or “Mixed.” Labels for human-written mail can be switched on as well, should you prefer the complete picture instead of only the suspects.
The scan runs on arrival rather than backwards through your archive. “New messages are checked and labeled as they arrive. Pausing the integration will stop new scans. Existing labels stay in Gmail,” the company wrote on its website.
Once a label lands, in other words, it belongs to you. Hit pause and your inbox holds on to its history.
Individual senders can be left out of the scan entirely, a setting that carries more weight than it first appears to if you deal with anyone whose prose tends to get mistaken for a model’s. A dashboard rounds things out, showing how Pangram labeled your mail across the past 24 hours, seven days or 30 days.
Privacy was argued out inside the company
According to Spero, privacy prompted a good deal of internal debate, given just how sensitive email data is. Where the company landed was a zero data retention policy: nothing kept once processing finishes, nothing fed into training.
Pangram’s website pushes the point further, stating that the detector is never handed an email’s sender, recipient, subject line or any other metadata. Text is all the classifier ever sees.
The setting to be careful with is autopilot
Pangram can be instructed to archive AI-labeled mail on its own, or to send it directly to spam. For now, Spero sticks with the mild option himself: apply the label, let the email sit in the inbox. That might shift later, he said, “if AI emails ramp up.”
Weigh that recommendation properly before flipping on auto-spam. Detection has come a long way, yet false positives have not disappeared, and what one costs you is not a mislabeled message. It is a genuine email that never reaches your eyes.
The company puts its detector’s accuracy on AI-generated text at 99.98%, and an independent study that measured several AI-detection tools against one another rated Pangram the most effective of them. Even so, a high-traffic inbox pushes through enough messages to turn up a miss sooner or later.
No one at the company claims the arms race wraps up
Detection tools, by Spero’s own account, will forever be trailing generative models that keep improving.
“I don’t think we’ll ever get to one hundred percent accuracy,” said Spero. The alternative available to the company is to “accumulate more certainty using more data.”
It is the honest way to put it, and an argument for keeping this in label-only mode through the first few weeks. Pay attention to what gets tagged, look at the dashboard once seven days have passed, and judge whether the flags line up with what you would have written off as slop on your own. Once they do, archiving becomes worth discussing.