We've spent our careers building the systems that keep automation off commercial websites: device intelligence, behavioral biometrics, bot defense, and risk decisioning. Much of that work has followed one rule: if it's automated, treat it as adversarial.
For twenty years, that rule was mostly right. Credential stuffing, scraping, carding, and inventory hoarding were all reasons to block automation, and doing so was a clear win.
Then automation started arriving with money.
When the false positive is a customer
A shopping agent sent by a loyal customer can look almost exactly like the thing a bot-management layer was tuned to stop. It drives a headless browser, navigates by script, and moves with inhuman timing. There's no mouse movement or pause before scrolling, none of the behavioral traces defenses learned to expect from a person. Some agents identify themselves through published user-agents or verified-bot programs, and the cooperative ones may eventually be allowlisted. Most runtimes, though, announce nothing. Even when an agent identifies itself, that doesn't tell you what it came to do. The signals still say to block it, challenge it, or quietly serve it a degraded page.
The defense worked. The sale died. And nobody on either side of it filed a report.
What bothers us is how quiet that failure can be. A blocked agent doesn't complain to support. It moves on to a competitor whose site it can use, and it may remember what happened. The customer's assistant can learn that your business "doesn't work." Your funnel shows nothing because the visit was never counted as a session. A false positive used to cost a retry. Now it may cost a sale and the customer relationship behind it.
Start by finding one number
The wrong conclusion would be to loosen the defenses. Adversarial automation hasn't gone away; if anything, AI may have made it cheaper and more convincing. Those defenses still matter. The decision just needs another dimension. For twenty years, "automated versus human" was a serviceable proxy for "adversarial versus welcome." It isn't anymore. What matters now is who sent the visitor and what it came to do. A scraper collecting your catalog and an assistant completing a purchase for a known customer both fail the Turing test, but they shouldn't necessarily get the same response.
A practical place to begin costs nothing but attention: find out what share of the traffic your defenses reject is agentic, and how much of it may have been there to transact. Most teams haven't seen that number, and most tools can't produce it, because upstream systems put everything "automated" into one bucket before anyone considers intent.
Once you can see that number, the trend stops being abstract. Revenue may already be reaching your site and getting turned away, because the stack did what it was told.