Stay informed with free updates

Simply sign up to the Technology myFT Digest – delivered directly to your inbox.

A lot of critical financial and government infrastructure runs on Cobol. The more-than-60-year-old mainframe coding language is embedded into payments and transaction rails, even though there are very few Cobol-literate coders available to maintain them.

The big argument in favour of sticking with Cobol systems is that they work. The catch is that, whenever they stop working, it is difficult to figure out why. That’s not good in a crisis, which is exactly when they’re most likely to break.

Covid-19 put a lot of strain the US state benefit systems. The ones that used Cobol for processing unemployment claims failed spectacularly, according to a new working paper from The Atlanta Fed:

States that used an antiquated [unemployment insurance]-benefit system experienced a 2.8 percentage point decline in total credit and debit card consumption relative to card consumption in states with more modern UI benefit systems. [ . . . ]

Using this estimate in a back-of-the-envelope calculation, I find that the lack of investment in updating UI-benefit systems in COBOL states was associated with a reduction in real GDP of at least $40 billion (in 2019 dollars) lower during this [March 13 2020 to year-end] period

The paper uses Cobol as a proxy for old and inefficient IT, not the direct cause of failure. Claimants faced much longer delays in the 28 states that still used Cobol in 2020, both because of the unprecedented volume of claims and the difficulty updating systems with new eligibility rules, author Michael Navarrete finds.

Untangling decades of spaghetti code in a hurry often means bringing back retirees at consultant rates, begging for volunteers, or calling on the Cobol Cowboys, a group of veteran coders who bill themselves as IT first responders. Ageing Cobol systems also tend to lack automation so, even when they work, delays lengthen when the caseload rises. Wisconsin, for example, was reportedly taking at least two months to process unemployment claims filed in March 2020.

The later a claimant is paid, the more likely they will save the money rather than spend. Each dollar’s economic benefit therefore remained relatively weaker in Cobol-powered states even after delayed payments had been made.

Having legacy systems in place also helped convince the government to keep things simple and offer a flat-rate extra of $600 a week in all states. Means-testing, or using a percentage of each state’s regular unemployment insurance, would have been too complicated to apply quickly. The result was “the median UI recipient receiving more from UI benefits than from their previous employer”, Navarrete says.

As an aside, one oddity of the data is that Republican-controlled states were more likely to have replaced old IT systems, even though their standard unemployment insurance payments are lower on average. Why? Absolutely no idea, but here are the maps:

2020 Cobol map

2020 electoral map

And, once adjusted for state politics, here’s the key finding:

All of which makes an irresistible case for upgrading public systems before they break rather than afterwards. But of course, financial system mainframes should be in better shape than government systems, having been stress tested in 2020, 2015, 2011, 2010, 2008 and 2001.

The bigger worry, as Dan Davies recently wrote, is critical bank infrastructure moving to the cloud, where borders are not clearly drawn and everyone is reliant on the same handful of suppliers. And it may or may not be a source of comfort to know that Morgan Stanley has been asking AI to figure out what its old Cobol code does.

Further reading: — Banking’s critical functions are vanishing into the cloud (FTAV)