DFS, the duty-free chain found in many airports, is one of several corporations that have entered a pilot scheme with Citi to use artificial intelligence to prevent fraudulent payments.
People familiar with the program say DFS began testing the software in mid-2018, and now all of its Asia Pacific payments are involved. The company hopes to replace manual checks, reduce headcount, and eventually outsource all of their internal controls to the bank, according to one executive in the pilot scheme.
Citi is said to be offering its solution for free as a value-added service, but even so, it’s a major commitment from the corporate side. “We need to put our time and resources into it,” said the executive, “We believe in its potential.”
The technology may be the work of Feedzai, a fintech in which Citi Ventures- the corporate venture arm of the bank-invested in 2016. This could not be corroborated by press time. Citi declined to comment.
Banks such as Citi have already poured resources into developing A.I. to raise red flags in areas such as anti-money laundering. Therefore it’s a relatively easy step for them to extend such surveillance tools to other functions – while the investment required by a given company to develop the same A.I. tools would be exorbitant.
Companies such as DFS had to first issue internal approvals to allow Citi to access their payments data. Citi is already DFS’s primary payments bank in Asia, so getting this permission was straightforward. But one treasurer at another Hong Kong-listed company told DigFin the company had to refuse, because HSBC is its primary payments bank.
“Citi approached us for this pilot but it was too complicated to make the switch,” this person said. They were also concerned about what could happen if something went wrong, noting that HSBC fields a large team, whereas more tech-reliant players such as Citi do not – so it’s harder to get someone responsible on the phone.
Training the machine
Such issues do not seem to concern companies in the pilot scheme. DFS and other companies worked with the bank to develop first of all high-level general rules for marking suspicious transactions.
When the computer flags irregular payments, DFS’s team needs to review whether or not it is fraudulent. In this way, the bank’s machine accumulates knowledge on the company’s payment patterns.
At this point, DFS is using Citi’s A.I. to flag potential problems in real time, but it is not stopping payments. This is called an offline mode, which is designed for the AI to be trained. People familiar with the test say companies are not changing their daily operations yet. Instead they are collecting warnings in batches, and then going through them all at once.
Once the number of false positives falls below 0.2%, companies will let the system halt suspicious transactions and send them to personnel to review. That is likely to take a full year of training the machine, to cover an entire payment cycle. Step by step, then, companies hope to eliminate internal controls until everything is automated.
For treasurers, using software to detect double payments or irregular payments is helpful, but what they really value is a tool to avoid payments involving counterparties on international sanctions lists.
If a payment goes to someone on a U.S. blacklist, the bank must immediately report this, and the cash will be frozen by the U.S. Treasury Department.
“With this solution,” said one treasurer, “we don’t have to manually update the sanctions list. We can avoid these payments before they go to the bank.”
Feedzai, the fintech reportedly behind this fraud screen, says it does not rely on rules-based detection. Rather it analyzes a given transaction against historical ones to find anomalies; its software gets better at this over time, with experience.
Citi announced in December 2018 that its treasury and trade solutions department entered a strategic partnership with Feedzai, and that the bank will integrate its transaction monitoring platform into its own, with enhanced control and risk management for payments transactions.
At the time, the bank said it expects to go live with Feedzai in mid 2019. If so, then DFS would likely be among the first corporate clients to use it.