Extend, the virtual card and spend management platform, has put AI to work around the clock — especially the hours when its engineers aren’t. 

Extend has been running a “night shift” project for the past six months, in which AI agents write, test and fix code during off-hours, effectively turning idle server time into productive engineering capacity, Chief Executive Andrew Jamison, told FinAi News. 

“You’ve got 24 hours in a day. Our employees are busy for the best part of eight or 10 of them,” Jamison said. “What do you do with the other 14 hours?” 

The model has reshaped how Extend staffs its engineering teams, he said. 

Projects that once required eight people — covering front-end, back-end and product functions — now need three, while output has roughly tripled, according to Jamison.

“You’re getting, like, basically 2X to 3X productivity out of your technology teams, and we’re seeing this firsthand,” he said. 

AI tokens, usage 

As it deploys off-hours AI, Extend is keeping tabs on the bill for token usage, he said. 

“Our CFO raised this question last year, that ‘What will you do when AI becomes more expensive than humans,’” Jamison said. “So far, with the night shift projects, we have found that we are spending $100 a month per head for [each of] our engineers.” 

Other FIs like Happen Bank and JPMorgan are using different ways to clamp down on AI spending, according to FinAi News’ prior reporting. Happen Bank for one is putting in caps on the token burn rate of its employees, while JPMorgan is asking employees to be cognizant of what model they are deploying for what task. 

The agents are also integrated into Extend’s Slack, Jamison said. 

When a client reports a bug, the system identifies the problem, locates the relevant code, and suggests a fix — without waiting for a human to triage the ticket, he said, but a person makes the final decision on fixing the bug. 

“The agent says, ‘Here’s where the bug is, and here’s where the fix is. Here’s the code for you to go and deploy,'” Jamison said. “We always maintain human oversight on all code that is getting shipped out.” 

The New York-based company uses a multimodel strategy for internal and external operations, Jamison said. The company uses Anthropic’s Claude the most, followed by OpenAI’s ChatGPT. 

Finance team agent coming 

Extend is using the night shift project model to create a program for its finance function and plans to offer the result to customers, Jamison said. 

The company is building a finance agent it expects to ship by the end of the third quarter, designed to automate the month-end close, he said. 

“We have two and a half people [a part-time employee] in accounting that do everything — AP, AR, payroll, everything — and the most time-consuming time is month end,” he said. 

The agent is being built to ingest bank statements, reconcile them against internal transaction records and flag discrepancies for human review, he said. 

For example, the agent would identify a late fee and an unmatched charge and ask the user whether to create corresponding transactions for export to QuickBooks, collapsing what can be tens of hours of work into a single conversation. 

“It’s going to be saving 10 hours of reconciliation,” he said. 

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas.