■ CRITICAL RISK ■ Finance
Yes — real-time authorization systems took this job so thoroughly that most people don't know it ever existed. Every card transaction is approved or declined by software in under a second, and the residual referral queue is now fraud-model territory.
“Approved or denied in milliseconds. Your 20-minute review was cute though.”
Our AI replacement risk score — how we score jobs
Once upon a time, a merchant handling a large card purchase phoned a bank, read out the card number, and a credit authorizer decided — checking the account's balance, limit, and payment history, sometimes flipping through printed bulletins of bad card numbers — whether to give the merchant an approval code. Authorizers also reviewed accounts flagged for unusual activity and decided whether to extend credit past a limit for a good customer. It was judgment work performed at telephone speed.
Electronic authorization networks deleted the phone call: the point-of-sale terminal queries the issuer's systems and gets an answer in milliseconds, evaluating balance, limit, and — via machine-learning fraud scores — the transaction's riskiness, all automatically. The volume is unanswerable by humans in principle: billions of card transactions flow daily through decisioning systems that also handle the judgment calls authorizers once made, like approving a slight over-limit purchase for a customer with a strong history, encoded as issuer strategy rules. What remains human is the referral queue — transactions the system declines to decide alone — plus fraud-alert callbacks, and modern fraud models keep shrinking exactly that queue, which is why this scores 95 rather than a nostalgic footnote.
The occupational descendants are real but different: fraud analysts investigating flagged patterns, credit-strategy analysts who design the rules the machines execute, and dispute-resolution staff. Merchant-side, humans still handle some high-risk manual reviews in e-commerce, though AI review tools are consolidating that too. The instructive part is the shape of the transition — the judgment wasn't eliminated, it was interviewed once, written into rules and models, and executed a billion times a day without further consultation.
Automatability: our editorial assessment of current and near-term AI capability
This replacement completed decades ago — electronic authorization networks made human transaction approval physically impossible at modern volumes, and the job title survives mainly in labor-statistics taxonomies. The active frontier now is the referral queue: ML fraud models keep narrowing the set of transactions needing human review, so even the analyst descendants of this role are seeing their routine caseload automated this decade, leaving investigation and strategy work.
Yes — before electronic authorization networks, merchants phoned the card issuer for purchases above a floor limit, and a human authorizer checked the account and issued an approval code, sometimes against printed lists of canceled card numbers. The entire interaction that now takes 300 milliseconds at a terminal took minutes and a person. It's a useful reminder of how completely infrastructure can absorb a job.
A residue, under different titles. Humans still work referral queues (transactions the automated system won't decide alone), verify suspected fraud with cardholders, and perform manual review on high-risk e-commerce orders. Labor statistics still track a small number of credit-authorizer positions, but they're exception-handlers around an automated core, and fraud models shrink their queue annually.
The judgment moved upstream and split: credit-strategy analysts design the approval rules, data scientists build the fraud models, fraud analysts investigate what the models flag, and dispute specialists handle contested transactions. Collectively these employ fewer people per transaction by orders of magnitude — but they're more analytical, better-paid roles. The transaction-by-transaction decision job itself has no modern equivalent.
Safer, with an asterisk. Complex fraud investigation — organized rings, identity theft, first-party fraud disputes — resists automation because adversaries adapt faster than static rules. But routine alert triage is being automated now by the same ML wave, so the safe part is genuine investigation and casework, not queue-clearing. Enter through fraud operations, then specialize toward financial-crime investigation.