AI is moving into the plumbing

Deltek's latest research, picked up in the UK construction digest, says architecture, engineering and consultancy firms are starting to move from AI experimentation to operational use. The numbers are the useful bit. 55% describe themselves as advanced or mature in digital transformation. 29% say operationalising AI is now a strategic priority. Nearly half report productivity or cost improvements.

That is not a hype story. It is a sign that the work is becoming routine. Once AI starts showing up in forecasting, planning, reporting and resource management, the conversation changes. You stop asking whether it works at all. You start asking where it saves time without creating more risk.

Mikhail Parakhin's Shopify interview is not construction news, but it is relevant. His point was simple. The bottleneck is no longer just generating code. It is review, CI/CD and deployment stability. He said fewer agents, better critique loops and stronger PR review matter more than throwing lots of parallel agents at the problem.

That maps neatly onto construction. If you add more automation without tightening review and handoffs, you just move the mess around faster. The same applies to reporting, document control and any workflow where speed is useless if the output isn't trusted.

Google's Cloud Next announcements were big. TPU 8t and 8i split training and inference. Gemini Enterprise Agent Platform is now a proper enterprise surface, with agent studio, model choice and workflow tooling. Workspace Intelligence and other adjacent launches point in the same direction.

For construction teams, this matters because the enterprise AI layer is getting more coherent. If your project data lives in docs, sheets, mail and reports, this is the sort of stack that will shape how people search, summarise and act on it.

GPT-Image-2 is being used for slides, diagrams, mock-ups and other visual assets that need to be correct rather than pretty. That's the important shift. If a model can handle layout, text and reference-driven edits properly, it becomes useful in bids, presentations and early-stage client comms.

Kimi K2.6 and Qwen 3.6-Max-Preview both reinforced the same trend. Open and semi-open models are getting better at long tasks, tool use and coding workflows. That won't replace every commercial model, but it does change the economics for teams that want control, privacy or lower cost.

For construction, the practical takeaway is simple. If you want assistants for document search, takeoff support, reporting or internal Q&A, local and open models are getting good enough to matter. You still need the right workflow around them. But the model choice is less binary than it used to be.