We run a software delivery shop. This year we started tagging every ticket and commit across our AI engagements by what the work actually touched. Not what the project was called. What the engineer was doing.

Across 14 completed engagements (Jan through May), 73% of the engineering hours that got AI into production were data pipelines, integration layers, legacy system remediation, and human-in-the-loop tooling. The actual model work, the thing every board meeting focuses on, was 27%.

I mean, we knew infrastructure was heavy. We didn't expect the split to be that lopsided.

The part that really got me: when we compared budget allocations to actual hour distribution, it was almost perfectly inverted. Companies were budgeting 70/30 model-heavy. The projects that actually shipped to production ran 30/70 infrastructure-heavy. The ones that stalled had the original 70/30 allocation and ran out of infrastructure budget by month four.

Small sample, 14 projects, mostly mid-market (50 to 300 people). Take it with appropriate salt. But the pattern was consistent enough that we now run a data infrastructure audit before we even talk about models with a new client.

Amazon's Kiro situation earlier this year validated this at a completely different scale. They mandated 80% weekly usage of their AI coding tool before the safety infrastructure was ready. Tool worked fine. Foundation wasn't there. 6.3 million lost orders.

We wrote up the full breakdown with the budget patterns and a pre-mortem framework in our weekly briefing if anyone wants the longer version.

Anyone else tracking this kind of split on their AI work?

I want to know if 73% holds outside our sample or if we're skewed by client type.