Authorizer's Assistant: an expert system decides on credit

In November 1988, according to its authors' paper at IAAI-89, American Express put the Authorizer's Assistant into full operation in all US centres: an expert system of 890 rules in ART from Inference Corporation prepares card authorization decisions for authorizers and approves millions of dollars of credit daily without human intervention. By the same paper the pilot ran at one centre from March-April 1987; other sources date the pilot and the full deployment differently.

The system complements CAS, a program on an IBM mainframe that approves most requests itself and refers the rest to authorizers. It runs on Symbolics workstations, each serving up to nine authorizers, and applies its rules to thousands of transactions an hour, 24 hours a day. The paper's timeline: a request for proposals in July 1985, the contract with Inference in October 1985, knowledge development by about 4.5 knowledge engineers over 13 months from November 1985 to November 1986; the decision to deploy in September 1987; a beta test in July-October 1988. By the paper, productivity savings alone would pay for the system in under two years, and the potential savings on losses are more than five times larger. Kingston's independent review (ES 2003) says an authorizer makes such a decision in about 90 seconds and that the system saved more money through better decisions than through time. The 1988 book The Rise of the Expert Company allows seventy seconds for a decision (pp. 92-93). Volumes. The paper gives only a rate: thousands of transactions an hour, and millions of dollars of credit approved daily without human intervention (p. 168). No source read gives a count of all authorizations: the company presents its findings in relative terms (the 1988 book, p. 110), as does Release 1.0 of 21 September 1987. Only the human workload is known: by the book, the computer first analyses a charge statistically and can deal with about 80 percent of charges automatically (before the Authorizer's Assistant), passes the rest to people, and 300 authorizers companywide handle over a million such transactions a month. Computerworld of 23 November 1987 speaks of more than 300 authorizers in four US centres, Release 1.0 of 200-odd; Computerworld in 1993 (an American Express director) and Newquist in 1994 say the system does the work of 700 authorizers. Effect. Every figure comes from American Express itself; no source read gives an independent measurement. Computerworld of 23 November 1987 passes on its data: accuracy of 96.5 percent against 85 percent for an average authorizer and 76 percent fewer bad decisions in a static evaluation; in April (1987 by context), on 2,500 cases handled by the system and by an authorizer, a third fewer refusals and half as many questionable approvals. By the paper, in experiments with the system as an adviser productivity rose 20 percent. The cut in time is more an expectation: the company's report (Release 1.0, the book) foresees 20 percent less time on transactions that go to authorizers, and Robert Flast expects 25 percent or more of authorizers' time to disappear (p. 104). The 45-67 percent internal rate of return on productivity alone is a line of the company's report in Release 1.0 and the outcome of its return-on-investment analysis in Computerworld in 1987, while the book credits the estimate to Computerworld. In money the figures differ: Davenport and Short (1990) give $7 million a year less in credit losses, a 25 percent shorter authorization and 30 percent fewer improper denials, with no source in the passage read and numbers unlike the paper's; the book's authors estimate $27 million a year themselves, since the company will not say; a company director (1993) speaks of millions a year, Newquist (1994) of tens of millions. A discrepancy: a paper by the same first author at IAAI-92 dates US-wide production to November 1989; the record follows the 1989 paper, written closer to the event. The abstract of a 1996 paper by M. Stark of American Express agrees with it: the company deployed the system in 1988, after a pilot that gave the business case and a rollout that made it fault-tolerant for round-the-clock service. So does a 2001 interview with Marilyn Stark, who leads the system at American Express: by the journal author's account, production began in the fourth quarter of 1988. Against November 1988 stands The New York Times of 31 July 1988, which says, naming no source, that the system was fully installed in June; the sources do not say whether that means the paper's full operation in all centres. Kingston's article of 1989 only expects full deployment during 1988. The pilot's start differs too: the paper has experimental use at one centre in March-April 1987; Computerworld of 13 April 1987 says a pilot version debuted the previous November; the 1988 book describes a demonstration on live transactions on 22 December and the new system installed in Florida in January (no year printed). What the record does not claim: an independent measurement of the effect, since every figure comes from American Express itself and the method is described only in outline; a count of all authorizations, which no source read gives; one settled date of full deployment or of the pilot's start. Unread: the full text of Stark's 1996 paper (paywall), the Forbes article of 2 November 1987, and the footnote of Davenport and Short naming their source.