What does AI productivity look like in production?

AI, Engineering, Productivity, Teams

Human engineer and AI agent reviewing production code with reported throughput and cost savings

Most AI engineering productivity claims focus on how quickly code can be generated. This case is more interesting because it measures something closer to the whole delivery system.

Odyssey Logistics needed to move business logic from legacy Microsoft Access and VBA applications into its new cloud-native platform. A seven-person Cognizant engineering team used Devin to support the conversion — producing code, tests, deployment pipelines and documentation.

The reported result: delivery throughput increased by roughly one-third and Odyssey recorded a 37% net cost saving compared with the conventional approach.

Importantly, this wasn't fully autonomous delivery. Every change was reviewed and approved by a Cognizant technical lead or architect before merge. Human review remained part of the system.

There are limitations. Cognizant and Odyssey haven't published the underlying throughput calculation, baseline period, work units or detailed cost model behind the 37% saving. These are reported production results, not an independently validated study.

Still, that's what makes the case interesting. The useful question isn't “how much faster can AI write code?” It is whether AI can improve the economics of getting production software delivered while keeping engineering governance in place.

Practical takeaway: measure AI engineering at the delivery-system level — throughput, cost, lead time, quality and human review effort — not just generated code or developer activity.

Source:

Cognizant

Read full article here:

Cognizant and Cognition put autonomous AI engineering into production at Odyssey Logistics, with a 37 percent net cost saving

My take:

What stands out to me is that humans remain part of the delivery system, reviewing and validating the outcome.

With agents, it can be difficult to understand what they are actually doing and whether that activity is moving us toward the result we want. That's why I think we need to measure AI at the system level: what value does it add, what does each activity cost, and does it actually improve the final outcome?

Agent activity itself isn't productivity.