Why McKinsey’s 2026 agentic construction blueprint describes a demo somebody built eight years ago, and what that means for Arizona builders
McKinsey published a piece last month called their blueprint for agentic AI in the construction industry. The opening scene is beautiful.
A field superintendent walking a jobsite discovers that prefabricated pipe spools no longer fit after a late engineering change. In the agentic future, the superintendent snaps a photo of the affected area, and within minutes, AI agents cross-reference the image against the 3D model, the drawings, the procurement records, and the schedule. The agents recommend a path forward. The problem gets caught before it goes critical path.
That scene was already technically possible in 2018.
An AEC innovation veteran wrote recently about a project he led at Bechtel that year. Same problem. Same workflow. Drone photogrammetry to map every pipe spool in the laydown yard. Automated comparison against the 3D fabrication model. Ground-truth against procurement records. The tool was called Spool Hunter, and it was presented on stage at ENR FutureTech in 2018. Not a research paper. Not a prototype. A working demonstration of exactly the workflow McKinsey is now positioning as construction’s agentic future in 2026.
Every piece of technology in that 2018 demo already existed. Drones. Photogrammetry. 3D BIM comparison. Structured procurement data. Every piece.
What did not exist was the organizational plumbing to make the pieces talk to each other in production.
That is the whole story of AI in construction. And it is going to be the whole story of agentic AI in construction too, unless the industry finally learns the lesson.
What “Agentic” Actually Means
Before I go further, let me draw the line clearly, because the term is being used to mean different things in different rooms.
There are three tiers of AI showing up in construction right now, and they are not the same thing.
- Chatbots and assistants answer questions when you ask. You type or say something, they respond. Useful, but you drive.
- Copilots help you do a task faster. Draft an email, summarize a document, suggest a schedule reorder, generate a report. You still have to be there, and you still make the final call, but the copilot cuts your time in half.
- Agents actually execute multi-step workflows on your behalf, making decisions along the way. You give an agent a goal, boundaries, and access to systems. The agent reads the situation, decides what to do next, executes, and only comes back to you when it hits a decision it is not authorized to make.
The last one is the real shift. Not “AI that helps you.” AI that does the work, with humans setting objectives and monitoring outcomes rather than clicking through each step.
McKinsey’s vision is real. The direction of travel is right. My argument is not that they are wrong about where this is going. My argument is that they are late to a conversation the industry has been having in demos and pilots for the better part of a decade.
The Real Question
If the technology has been ready this long, and construction firms still are not running Spool Hunter or its equivalents in production, the honest question is not “what technology do we need next?” The honest question is “what has been in the way this whole time?”
The answer is not that construction is behind. It is not that AEC leaders are luddites. It is not that the software is not good enough. I have been in Arizona construction for twenty years and I can tell you the leaders in this industry are hungrier for real automation than the McKinsey partners writing these reports probably realize.
The answer is that the organizational conditions to run this stuff in production are much harder to build than the technology itself.
Look at what an agent actually needs to work in a construction workflow.
- Clean, structured, current data. The agent has to be able to read your 3D model, your procurement records, your schedule, your submittal log, and your RFI history. All of them. Reliably. In a format the agent can parse. Most Arizona construction firms have that data. In three different systems. In four different formats. Named differently across projects. Some of it is current. Some of it is a year old and nobody has time to update it.
- Explicit rules for what the agent can decide, and what it must escalate. In the McKinsey scene, the agent recommends a path forward. Somebody has to have decided, in advance, whether the agent is allowed to actually order the replacement parts, or whether that decision goes to a human. Whether it is allowed to notify the sub. Whether it is allowed to update the schedule. Every one of those is a governance decision that has to be made before the agent runs, not after.
- A defined handoff between agent and human. When the agent hits a decision it cannot make, who picks it up? On what timeline? How does the agent know the human received the ball? What happens if the human is on a jobsite and does not check their phone for four hours? Any construction firm that has ever tried to run a workflow across offices knows exactly how brittle handoffs are when they are not designed carefully.
- Audit trails and reversibility. Agents are going to make mistakes. Real ones, occasionally expensive ones. When that happens, you need to be able to see what the agent decided, why, what data it used, and how to reverse the action. Without that, you are not running an agent. You are running a black box with a company checkbook.
- Field discipline that makes the data worth something. This is the one everyone misses. An agent can only be as good as the ground truth you feed it. If your as-builts are red-lined on paper in the foreman’s truck, if your submittals live in a folder on someone’s desktop, if your RFI answers are in a text thread, the agent has nothing to work with. The organizational discipline of capturing real information in real time in shared systems is not a technology problem. It is a how your firm actually operates problem.
Every one of those five conditions has been the actual bottleneck for the last eight years. Every one of them is still the bottleneck now. And no amount of new agentic AI capability is going to close them for you.
Where Agentic AI Is Actually Going to Land in Arizona in the Next 18 Months
Setting aside the flagship jobsite-photo demo, let me be concrete about where agents are genuinely going to show up for small and mid-size Arizona construction firms in the next 18 to 24 months.
Bid response and preconstruction. An agent that ingests an incoming RFP, reads every attachment, cross-references your bid history, flags scope ambiguities, drafts a preliminary estimate structure with your standard cost assumptions, and hands it to your estimator with initial questions already answered. Not “AI that helps you estimate faster.” An agent that has already done half the work by the time your estimator opens the file.
Submittal management. An agent that receives a submittal, checks it against the spec, flags conflicts with prior approvals, routes to the right reviewer, and follows up when the reviewer does not respond. Not “faster document routing.” An agent running the workflow end to end and only escalating when the spec is genuinely unclear.
RFI drafting and response. An agent that reads an incoming RFI, matches it against similar historical RFIs from your past projects, drafts a suggested response with references to the relevant specs and drawings, and lets your PM review and send. Two-day RFI turnarounds become two-hour turnarounds on the routine ones. Your team’s time gets spent on the actual hard ones.
Change order tracking and exposure. An agent monitoring project communications and drawings for change signals, drafting change order documentation, and flagging financial exposure before it hides in a spreadsheet for three weeks.
Closeout package assembly. The one we wrote about a few weeks ago. An agent gathering as-builts, O&M manuals, warranty documents, and lien waivers throughout the project instead of scrambling at the end. Turning closeout from a fire drill into a final step.
None of these are magic. All of them are approaching production readiness in construction-specific platforms right now. And every single one requires the same five prerequisites I listed above.
The Honest Grade
If I had to grade McKinsey’s report for the Arizona builder audience, here is where I would land.
They are right about the direction. Agentic AI is going to change how construction firms operate. Not in five years. In the next 18 to 36 months for firms that are ready. The direction of travel is not in doubt.
They are early on the timing for most of the industry. Their examples are mostly Fortune 500 EPCs like Bechtel, or large enterprise deployments with resources that a mid-size Arizona GC does not have. The agentic capabilities that will land for a 35-person Tempe contractor are narrower, more specific, and more workflow-focused than the McKinsey vision suggests.
They understate the organizational lift required. Almost every case study in the McKinsey report is a firm that had already spent years cleaning up their data, their processes, and their governance. That work is invisible in the report. But it is the entire foundation. Firms that skip it will not scale their AI investment. They will just spend more on agents that cannot work with their data.
They ignore the field. Every McKinsey agentic scenario I have seen assumes the field is a source of clean digital signals. Photos, drone imagery, sensor data, real-time updates from mobile devices. On most Arizona projects, the field is still the source of paper notes, text threads, and a superintendent’s memory. Until the field-to-office pipeline is genuinely digital and disciplined, the agentic vision runs on air.
I am not saying McKinsey is wrong. I am saying their vision is aspirational, and the gap between the vision and the Monday morning reality of a mid-size Arizona construction firm is enormous. That gap is not a technology gap. It is an operations gap.
What This Means for You
If you have been reading these TechTips for the last few months, you already know what I am about to tell you. Because it is the same thing I have been telling you all year.
Your data is the lever. Not the tool. The moment you get serious about capturing your data cleanly and consistently, in structured systems that agents can read, the technology conversation gets a lot easier. Every one of the McKinsey scenarios becomes achievable. Every one of the agentic use cases I listed above becomes possible.
Your SOPs are the training material. Agents need to know how your firm actually operates. That knowledge lives in documented, current, followed standard operating procedures. If your SOPs are a binder from 2019, no agent will ever run your workflows correctly.
Governance before capability. The firms that will actually deploy agentic AI safely in construction are the firms that built the governance layer first. Who decides what the agent can do. Who owns escalation. Who monitors outcomes. All of that comes before the tool, not after.
Field discipline is the whole game. The agentic vision fails at the field-to-office handoff on almost every project. The firms that fix that are the firms that will run agents. The firms that do not are the firms that will keep watching demos.
Your Move
Do not buy the first “agentic construction platform” that lands on your calendar next quarter. There will be several. Some of them will be genuinely useful. Some of them will be Copilot with a new logo.
Instead, spend the next 90 days getting honest with yourself about your operational plumbing. Ask three questions.
THE 90-DAY READINESS CHECK
- What are the five workflows that would benefit most from agent-level automation if we could run them today? Preconstruction. Submittals. RFIs. Change orders. Closeout. Something else. Pick five.
- For each of those five, what would we need to fix about our data, our SOPs, and our field-to-office flow before an agent could actually run the workflow safely? Be specific. Write it down.
- Which of those fixes could we make in the next six months without buying any new software? The answer is usually most of them.
Do that work and the next 18 months of agentic capability will land in your firm cleanly. Skip it and you will spend a lot of money on software that never delivers on the demo.
The tech was not the bottleneck in 2018. It is not the bottleneck now. It never has been.
The firms that figure that out early are going to look, in five years, like they had a magic AI advantage. They did not. They did the boring work that made the technology possible. A common theme with most technology.
A demo was built in 2018 that McKinsey is now calling the future. The reason it is not standard practice yet has nothing to do with capability. It has everything to do with whether firms did the work of getting ready.
Get ready.
For over 20 years, Dimensions has been the trusted IT partner for Arizona’s architecture, engineering, and construction industry. We help AEC firms communicate better, collaborate smarter, and actually use the technology they’ve invested in. Because in construction, the tools only work if your team does.
IT Built For Builders.
