#919 1988 · American Express (with Inference Corporation) · Financial services / credit
American Express scaled its rarest judgment call with software instead of headcount
问题
Every unusual charge needed a human authorizer's split-second judgment, but training one to veteran quality took years
背景
An authorization decision looked instantaneous from the outside, but a veteran was really weighing a handful of signals — spending pattern, account history, merchant type — learned correctly only after years of seeing how past calls turned out. Hiring ahead of volume meant a pipeline of new authorizers making worse calls than veterans until they caught up, for as long as that catch-up took.
American Express partnered with Inference Corporation to see whether that judgment could be captured directly: interviewing its most experienced authorizers about the exact signals and thresholds behind their calls, rather than assuming the skill could only be transmitted by years on the floor.
换别人会怎么做
The standard response to authorization volume outgrowing trained staff was to hire and train more authorizers, accepting years of ramp-up per hire and worse calls from the newest hires until they caught up.
他们看到了什么
Inference and American Express treated a veteran's split-second yes-or-no as a learnable rule set drawn from account signals, not unique intuition, and built a system junior staff could consult on demand.
那一手
The two companies built the Authorizer's Assistant, encoding more than 3,000 core business rules, drawing on up to 13 internal databases of transaction and account history, and deployed it across US Personal and Gold cards in the fourth quarter of 1988 to reproduce the judgment calls that used to require a veteran authorizer.
为什么管用
An authorization decision looks intuitive from the outside, but it is really a fast lookup across a few signals a veteran has learned to weigh correctly through years of seeing outcomes. Once those weighting rules were captured from veterans and encoded as software logic pulling from the same account databases a human would check, a newly hired authorizer working alongside the system could produce a veteran-quality decision on their first day, because the system supplied the pattern-matching a veteran would otherwise need years to build.
值了多少
The system ran over a decade without rewriting its rule base, maintained by just five people versus the 30–40 typical for a system its size.
什么时候会失灵
A rule-encoded system like this stays accurate only while the underlying fraud and spending patterns it was trained on don't shift too fast; new fraud tactics can silently outrun a static rule base until someone notices the miss rate creeping up. It also depends on the original interviews having captured veterans' real reasoning rather than what veterans believed they did.
后来呢
Authorizer's Assistant became one of the most cited commercial successes of the 1980s expert-systems boom in financial services, credited with letting American Express scale authorization judgment with card volume without scaling headcount at the same rate.
资料来源
- [1]American Express Authorizer's AssistantBizRules.info (Business Rules Group), 2001bizrules.info
- [2]The Authorizer's Assistant: A Knowledge-Based Credit Authorization System for American ExpressProceedings of the First Conference on Innovative Applications of Artificial Intelligence (IAAI-89), AAAI, 1989cdn.aaai.org