Martin Fischer, Stanford’s “Mr. BIM,” argues that if you have $100 to spend on AI in construction, $20 should go to the tech and $80 should go to rebuilding your processes and data infrastructure. Most ConTech AI budgets are inverted, which is why so many projects underperform. The fix is not a better model. It is the digital feedback loops underneath it.
There is a quiet pattern showing up across construction technology budgets right now. Firms are racing to plug AI into estimating, scheduling, document review, and field reporting. Pilots are being signed, copilots are being trialed, and the results often feel underwhelming. Hit rate low. ROI murky. The promised productivity gains keep slipping into next quarter.
The reason is not that the models are bad. They are extraordinary. The reason is that the data underneath them was never built for this. Until that changes, every dollar spent on the AI layer is doing the work of about twenty cents.
That, at least, is the argument Martin Fischer made at our recent Bricks & Bytes event. Fischer runs Stanford’s Center for Integrated Facility Engineering and is widely credited as a founding figure of BIM. When he talks about how data should sit underneath construction workflows, the room tends to listen.
The mismatch between AI ambition and AI outcomes in construction is mostly an upstream problem. The processes that generate construction data were never designed to feed machine learning. Until that changes, AI returns will stay shallow.
The 80/20 rule for AI investment
Spend the small share on tools, the big share on plumbing
Fischer’s framing is deliberately simple. Imagine $100 to invest in AI inside a construction business. He would put $20 toward exploring what current tools can do. That covers large language models, vendor pilots, and the kind of light-touch experimentation most teams already run. Real work, but not where the value lives.
The other $80 goes somewhere far less glamorous. It funds the rebuild of how your organisation generates data in the first place. Process redesign. Schema definition. The infrastructure that lets AI sit on top of something useful rather than something messy. Without that base, you are training algorithms on a partial picture of what actually happened on a project, and the output reflects the input.
This is why so many AI rollouts in construction stall after the demo. The shiny part works. The underlying data does not give the model enough to be reliable, and the team quietly stops using it. That is not an AI failure. That is a process failure dressed up as one.
Why construction data is built for the wrong thing
Most of what you collect was never meant to teach a machine
Most of the data construction firms hold today was collected to satisfy a contractual or commercial requirement, not to train a system. Site reports get filled in to evidence progress. RFIs get logged because someone asked. Cost data gets coded the way the accounting team needs it. None of that was built with downstream AI in mind, which means the data shape is constrained by its original purpose.
Fischer’s point is that even firms with rich data are often training models on a constrained slice of reality. You see the parts of the project that got documented because someone had to document them. You miss the parts that mattered but were never captured in a structured way. Build a model on that and you get a confident answer to a half-formed question. We made a similar argument when construction was reportedly spending billions on the wrong problem.
If you have structured data, it is much easier to deploy AI. The way you structure your databases is such a great definer of how much you can truly do with AI super, super quickly.Patric Hellermann, Foundamental, on the Bricks & Bytes podcast
Digital feedback loops are the missing piece
The concept Fischer kept returning to at the event
The phrase Fischer used most that night was digital feedback loops. The idea is straightforward, even if the execution is not. An engineer or architect designs a detail. The team builds it. Someone needs to verify whether that detail actually worked: was it constructable, did it perform, did it hit the cost and schedule it was supposed to. Most of the time, no one closes that loop in a structured way.
In other industries this would be unthinkable. Software has telemetry. Manufacturing has SPC. Aviation has black boxes. Construction has photos, punch lists, and a lot of institutional memory walking around in the heads of senior staff. Verification happens, but it is not captured in a form anything else can learn from. What Fischer is pushing for is the same loop, but engineered. Design intent on one end, verified outcome on the other, and a structured pipeline in between. Once that exists, AI has something to chew on. Without it, every model is guessing.
The signal worth paying attention to in ConTech
Get sharp weekly takes on AI, data, and the deals reshaping construction. From 300+ issues and counting.
Join 2500+ ReadersWhat the 80% actually buys you
It is not a slogan, it is a budget line
The 80% does not mean buying more software. It means deliberately rebuilding the parts of your operation that produce data. How site teams report progress. How design intent is captured against built reality. How cost codes map to actual scope. How variations get documented in a way a model can learn from. Mundane work. Expensive to do properly. Almost impossible to skip.
This pattern shows up consistently in firms that get real value from AI. DPR Construction’s CTO Atul Khanzode talked about it in similar terms when he laid out his three pillars of innovation: tech only sticks when it is integrated into existing processes that have been redesigned around it. The firms quietly outperforming are the ones doing this in the background. The firms struggling are the ones treating AI like a feature flag.
How the 80/20 rule cashes out across the AI stack
| Layer | Where the 20% usually goes | Where the 80% should go |
|---|---|---|
| Estimating | AI takeoff and bid copilots | Standardising scope codes, structuring historic cost data, capturing win/loss |
| Design | Generative tools, clash detection AI | Linking design intent to as-built outcomes through structured BIM |
| Field execution | Computer vision, daily report automation | Redesigning what gets reported, by whom, against what baseline |
| Project controls | Schedule risk and forecasting models | Cleaning cost code structures and tying them to scope and progress |
Frequently asked questions
Martin Fischer is the Kumagai Professor at Stanford’s School of Engineering and Director of the Center for Integrated Facility Engineering, the lab that helped pioneer modern BIM and 4D modelling. Nemetschek recently announced a strategic partnership with CIFE to translate that research into industry tools. (Source)
Research published through CIFE highlights that despite enthusiasm for AI in AEC, around 85% of big data projects fail. A core driver is the disconnect between teams planning AI adoption and teams responsible for implementing it on real projects. (Source)
Typical AEC IT spend sits at roughly 1-2% of revenue, against a cross-industry average closer to 3-5%. McKinsey has called this gap one of the structural reasons ConTech adoption is harder than equivalent software categories in logistics or manufacturing. (Source)
It is a structured way to verify whether what was designed actually worked once it was built. Design intent goes in on one side, the as-built outcome comes back on the other, and the gap between the two becomes data the organisation can learn from. Most construction firms do this informally. Very few do it in a way an AI model can use. (Source)
Pick one workflow with high document load and clear commercial pain, and redesign the data flow before adding any AI. Estimating, RFI handling, and design-to-as-built verification are typical starting points. Get one feedback loop working end to end, then layer AI on top of that loop rather than buying a tool and hoping the loop forms by itself. (Source)
Related Articles
Construction is Spending Billions on the Wrong Problem
Why AI in Construction Is Focused on the Wrong Problem
The Three Pillars of Construction Innovation
