The Data Problem Nobody Wants to Admit -- and the Week It Became Unavoidable

This week had a story that most construction AI coverage skipped over. It shouldn't have.

Siemens published something unusually direct for a technology company: an acknowledgement that the biggest problem in industrial AI isn't the tools, it's the data. Gartner backs them up. 60% of AI projects unsupported by AI-ready data will be abandoned by the end of 2026. In manufacturing and construction specifically, the failure rate is already 76.4%.

Read that number again. Not 30%. Not half. Three quarters of industrial AI projects fail -- and the culprits are OT/IT integration gaps and data quality problems. Not the AI itself.

This is the context that the rest of this week's news needs to be read in.

NavLive won Best Use of AI at the Digital Construction Awards this week. Their LiDAR scanner generates RICS-grade surveys and full BIM models directly on site, in under 30 minutes. AWW Architects has adopted it specifically for Building Safety Act compliance. Procore went agentic after acquiring Datagrid AI in January, with an Agent Builder now in open beta and a pre-built RFI agent that reportedly cuts resolution time from days to seconds. The UK government named AI infrastructure a formal national security priority in public procurement guidance. These are all real, positive developments.

But they all have the same dependency. They need data to work.

NavLive produces structured spatial data from a scan. That's useful if you have a project environment that can receive, store, and act on it. Procore's agentic RFI tool validates an RFI against the latest drawings, checks the spec, and drafts a response -- but only if the drawings are current, the spec is properly uploaded, and the revision history is consistent. AI that navigates your document store is only as good as the document store it navigates.

The Building Safety Act is accelerating AI adoption faster than almost any other force in UK construction right now. It's creating demand for audit trails, documentation quality, and compliance evidence that paper-based and fragmented digital processes simply can't deliver. Firms are reaching for AI tools because the regulatory pressure is real. That's a reasonable response. But if the underlying data isn't clean and consistent, those tools will either fail quietly or produce exactly the hallucinated outputs that Construction Dive warned about earlier this week -- confident-sounding, wrong, and legally problematic.

There's a pattern I keep seeing in construction AI adoption stories. A firm deploys a tool, reports impressive early results, and then the case study goes quiet. What often happens next is that the tool runs into the data problem. It surfaces inconsistencies. It fails on edge cases that weren't in the pilot. The ROI calculation gets murkier. And the tool ends up in a corner of the stack, used by a few people who know how to work around its limitations, rather than embedded as a genuine workflow.

The firms that don't hit this wall are the ones that fixed their data before they deployed the tool. Not perfectly -- you don't need a pristine data environment to get value from AI. But you need enough consistency, enough structure, and enough quality control that the AI has something reliable to work with.

What does that actually mean in practice?

For most construction firms, it starts with the document register. Are drawings properly named, versioned, and accessible in a single location? Is the spec linked to the relevant drawings? Are RFIs indexed against the documents they reference? These sound like basic hygiene questions, and they are -- but they're also the direct enablers of every AI document tool that construction software vendors are selling right now.

The next layer is process consistency. AI tools work best on repeatable tasks. If your RFI process is different on every project because each project manager runs things their own way, an AI agent that learns one version of the workflow will struggle on the next project. Standardisation isn't just about efficiency -- it's the foundation that makes AI tools generalisable.

And the third layer is feedback. The practitioners who are getting real value from AI tools are the ones who have built evaluation into their workflows. Not just generating AI outputs and moving on, but checking them, logging errors, and feeding that information back into how the tool is configured and where it's trusted. The over-agentic problem flagged in technical coverage this week -- AI that looks productive but isn't -- is fundamentally a feedback problem. Without evaluation, you can't tell the difference between an AI tool that's working and one that's producing confident-sounding noise.

None of this is a reason to hold off on AI adoption. Quite the opposite.

The firms that take data quality seriously now are building a compounding advantage. Every AI tool they deploy will work better because the underlying data is cleaner. Every new tool they add will integrate more easily because their processes are consistent. And the failure rate that's killing three quarters of industrial AI projects will fall dramatically, because the data infrastructure that keeps tripping firms up won't trip them up.

The UK construction industry is at a genuine inflection point. The awards, the agentic platforms, the government procurement signals, the net zero standards -- they're all pointing in the same direction. But the firms that will still be talking about these tools in two years, with numbers they can stand behind, are the ones that answered the data question first.

Tools second. Data first. The order matters more than most people are admitting.

NavLive, developed from University of Oxford research, won Best Use of AI at the Digital Construction Awards 2026. Their handheld LiDAR scanner generates RICS-grade 1:100 surveys, 2D floorplans and 3D point cloud models on site -- no cloud upload required. BW Workplace Experts scanned an eight-storey London bank in under 30 minutes and had a full BIM model within hours.

What makes this more than an award story: AWW Architects has adopted NavLive specifically to meet Building Safety Act compliance requirements. That's a named firm, on live projects, driven by regulatory obligation. NavLive's GBP 4m funding round confirms serious commercial interest beyond the accolade.

The scan-to-BIM market has been building quietly for several years. This award, this adoption case, and this funding are the signals that it's crossed from specialist-contractor territory into mainstream practice.

After acquiring Datagrid AI in January, Procore has moved fast. Agent Builder is now live in open beta, letting construction teams build custom AI agents via natural language prompts. The pre-built RFI Creation Agent reportedly reduces resolution time from days to seconds. The agents can search across fragmented data sources, validate RFIs, run scope checks, handle prequalification, conduct audits and generate reports -- all autonomously.

This is the platform-level shift the industry has been watching for. Not AI features bolted onto existing software, but agentic workflows embedded into the project management layer. The distinction matters in practice. A chatbot that answers questions about your project data is useful. An agent that validates an RFI against the latest drawings, checks the spec, and drafts a response is a fundamentally different proposition.

UK contractors on Procore should be exploring the Agent Builder now. This is the first mainstream construction platform to ship agentic AI as a configurable, user-facing product rather than a back-end feature.

New procurement guidance published this week names AI infrastructure alongside steel, shipbuilding and energy as a critical national security sector. Government departments must now prioritise British businesses for AI infrastructure contracts or formally justify overseas sourcing.

This is the first time AI infrastructure has been enshrined in public procurement guidance as security-critical. The practical signal for the construction and infrastructure market is clear: significant public-sector AI build-out is coming, and procurement processes will favour UK-based contractors.

Set against Thursday's data centre story -- AI facilities pulling specialist MEP trades away from residential projects -- the picture gets more complex. The government is signalling major AI infrastructure investment at the same moment when the construction labour market for that infrastructure is already under pressure. Contractors positioning for this pipeline need to be thinking about security clearances, compliance requirements, and workforce capacity now.

Analysis from Build News this week put a direct number on something that construction professionals have been feeling but not naming. The surge in AI data centre construction is pulling specialist MEP trades -- particularly those with cooling and power infrastructure skills -- away from residential projects. Kao Data's 17.6MW liquid-cooled facility in Harlow and QTS's Cambois campus in Northumberland are just two of the projects creating this demand.

The timing is brutal. The government is targeting 1.5 million new homes this Parliament, with London aiming for 88,000 homes per year. Emergency housing powers were confirmed this week to restart stalled London schemes blocked by cost and financing pressures. The MEP trades needed for those homes are competing for work against higher-margin data centre projects.

BCIS forecasts construction costs up 14% and tender prices up 15% by 2031. Cost pressures are coming regardless. The AI infrastructure boom is accelerating them.

Siemens has been unusually direct about a problem most AI vendors won't touch. Gartner projects that 60% of AI projects unsupported by AI-ready data will be abandoned by end of 2026. In manufacturing and construction, the current failure rate is already 76.4%. The culprits aren't the AI tools. They're OT/IT integration gaps and data quality.

In construction terms, this translates directly: fragmented project data, inconsistent naming conventions, information locked in siloed systems, and sensor data that was never structured for analysis. These aren't exotic problems -- they're the normal state of most construction firms' information environments.

The UK Admin Drain Report from HeyBRB, published separately this week, adds context. UK tradespeople lose 8 hours a week -- 10 working weeks a year, worth GBP 17,000-25,000 in lost billable time -- to manual admin. The processes generating all that manual work are also the processes generating the fragmented, inconsistent data that makes AI fail.

Skanska UK CEO Katy Dowding has made the clearest case for construction robotics I've heard from a major contractor: AI and robotics are safety tools, not efficiency tools. Her team has deployed the Schindler Robotic Installation System on the 105 Victoria Street office build in London, removing workers from elevator installation tasks. She's explicit about the ambition -- using AI and robotics to remove humans from the most dangerous positions on site: electrical work, heavy lifting, working at height, people-plant interface.

The UK construction sector averages 35-50 fatalities per year. Electrical accidents have risen 75% in recent years. The robotics story and the safety story are the same story.

AMC Bridge has built an MCP Connector that lets AI agents work across both Autodesk Platform Services and Procore simultaneously. Natural language queries can now span project documents, drawings and cost data across both systems, replacing point-to-point integrations. The Model Context Protocol is becoming the standard plumbing for multi-tool AI workflows in construction -- and this demonstration is the first time the two dominant construction platforms have been bridged this way.

Combined with Procore's Agent Builder (above), this is the week that multi-system AI agents went from concept to deployable technology for UK construction teams.

Anthropic put a significant capability into research preview this week: Claude can now control a Mac directly. Open apps, navigate browsers, scan emails, fill spreadsheets -- all driven by AI. It works via Claude Cowork and Claude Code on Pro and Max plans.

It's a research preview for a reason. Browser control and computer use in real-world conditions is harder than demos suggest, and the community reaction was a mix of genuine excitement and pointed concern about reliability. But the direction is clear. AI is crossing from "thing you query" to "thing that operates software."

For construction, the target zone is obvious: chasing submittals, cross-referencing drawing revisions, updating document registers, generating meeting summaries. These are repetitive, time-consuming tasks that sit precisely in the target zone for AI agent control. The capability is real. The reliability in complex project environments needs proving.