Can six levers really map construction in the age of AI?

Six levers are not the map.

Last week Suffolk and MIT put the full whitepaper behind their Construction in the Age of AI research on the public record: a top-five US contractor working with MIT's Center for Real Estate and the Media Lab's City Science Group, built on a March roundtable of more than fifty industry leaders [1]. This is not a vendor brochure. MIT co-authored it, the evidence review leans on academic literature, and the appendix shows its working, including the votes that flatter it.

The coverage settled on six value levers: design automation, offsite manufacturing, permitting, scheduling, skilled labor and subcontracting, and supply chain and procurement.

The levers are real. Suffolk's first-pass model stacks all six on a sample $179 million project and finds roughly 17–20% cost savings, 22–25% schedule compression and a 5–6 point lift in unlevered developer IRR [1]. Permitting is the best-documented of the six: Boston alone divides 48 square miles into 429 zoning districts governed by nearly 3,800 pages of code essentially unrevised since 1964, and 42% of its parcels are non-conforming before anyone files for a variance [1]. Anyone who has waited on a permit recognizes that world.

Then Chapter 4 says the quiet part:

"Capturing that value requires a shared data infrastructure layer, something the levers themselves do not provide, that standardizes outputs at every handoff and gives every actor in the chain — owner, general contractor, subcontractor, supplier — a centralized system and industry data reporting standards to build on. Building it is the prerequisite to capturing the multiplier effect at scale, and the most consequential near-term opportunity the industry has on which to act." [1]

Read the middle clause again: something the levers themselves do not provide. The paper's own arithmetic puts the multiplier in the connective tissue between the levers, not in any one of them. Standardized outputs at every handoff is a description of a shared language for what construction produces and hands over. Chapter 1 supplies the diagnosis: "Institutional knowledge is rarely codified. Data is siloed across owners, contractors, subcontractors, and regulators." [1]

This is, nearly word for word, the thesis a small band of builders has been working on for years: fragmented data, not any single workflow, is construction's real bottleneck, and the way through is a connected knowledge layer where contracts, safety, programme, quality, carbon and procurement sit in one graph and answer questions no single-domain tool can. We are building one such graph (its schema runs to 115 node types and 89 relationship types; the counts matter less than the claim). A top-five contractor and MIT just arrived at the same claim independently and called building it the industry's most consequential near-term opportunity. When you have bet years of nights and weekends on a thesis, that is what third-party validation reads like.

The test is concrete. What did the subcontractor's contract say about delay notices, what did the programme actually record, and do the incident reports line up with the method statements? A contract repository, a scheduling tool and a safety system each hold one fragment. Only a connected layer answers the question, because the question crosses the walls between them.

Now the honest footnote: their layer and ours are not the same object. Theirs leans toward project-execution plumbing: BIM, schedule and procurement interoperability plus reporting standards for the delivery phase. A graph that also carries contracts, claims, competencies and carbon is broader, adjacent rather than identical. The economic claim is identical either way: point solutions capture fractions of the value, and the multiplier lives in the connective tissue.

Look at who the paper names against each lever: Procore, ALICE Technologies, OpenSpace, PermitFlow, Kaya [1]. Credible single-lever players, every one. Not one is building the connective layer, and the paper says outright that the levers will not provide it themselves. That white space is the most important paragraph of the whitepaper, and it is not in the whitepaper. You have to assemble it yourself from Chapter 4 and the vendor list.

And still: six levers are not the map. Consider what the frame cannot see.

Asset operations and facilities management never make the list, though over a building's life operations typically outearns the construction that produced it. Infrastructure and civils are absent: the roundtable roster (Wynn, MP Boston, SHoP, Hensel Phelps) is entirely vertical-building people, not one roads, rail or energy voice [1]. Contract administration and claims, the delay analysis and extension-of-time machinery of FIDIC and NEC that consumes months of every major project, is missing entirely. So is training and competency, the RAMS and workforce-skills lane. Materials production and decarbonization, insurance and underwriting, quality and compliance, handover to operations, carbon reporting, disputes: all zero, or footnotes.

None of this is a coincidence, and none of it is a conspiracy either. Three selection effects produced the six.

First, the seat at the table. Ask a GC-centric roundtable where the near-term money is and you will get a project-delivery answer. The six levers describe how Suffolk delivers faster and cheaper, because the people in the room were the people whose delivery the levers would speed up.

Second, the value function is two-dimensional. The model measures cost and schedule, and both point the same way. Quality outcomes, safety records, carbon, durability: invisible to an IRR calculation, therefore invisible to the map.

Third, evidence availability. The six levers are where pilot data already exists, not where the ceiling is. McKinsey's earlier map of the same territory counts more than 150 workflows across 25 AEC domains, worth around $228 billion a year to US construction by 2030 [2]. Suffolk picked six cells off that map and modeled them carefully. Picking cells is not a sin. Mistaking the cells for the territory would be.

The tell is in the paper's own logic. Chapter 4's argument escapes the six levers: once you concede that value concentrates in the connective layer rather than the levers, the list becomes an implementation detail, because a layer is domain-agnostic by construction. It spans exactly the domains the roundtable could not see from its seat. A layer built to answer design-schedule-procurement questions stops at the paper's perimeter. One built to answer contract-delay-claim-programme questions, or RAMS-competency-incident questions, keeps going, and covers ground the paper does not know it is missing. Same physics, wider geography.

Carry the caveats too, because the paper's own appendix supplies them. Design automation ranks first on 206 raw Borda points, but 89 of those are self-votes from the design cohort, more than 40% of the score voting for itself [1]. Suffolk sells AI-enabled delivery, so it has every incentive to flatter this thesis. And the headline savings are a first-pass model, not observed projects; the paper is careful about this, and its readers should be too. Treat the document as a research agenda with unusual institutional weight, not a measurement.

So cite the paper for the physics. Don't let its perimeter define the map. The shared layer will get built, and the whitepaper is the strongest mainstream evidence yet that the industry's establishment expects it. The open question is whose questions it answers first: the roundtable's, or everyone who wasn't in the room?

Sources

[1] Suffolk and MIT. "Construction in the Age of AI" (whitepaper). September 2026. https://suffolk.com/wp-content/uploads/2026/09/Suffolk_MIT_AI_Whitepaper_FINAL.pdf

[2] McKinsey Global Institute figure, via The Economic Times. "AI could unlock $228 billion in annual value for US AEC sector by 2030." https://economictimes.indiatimes.com/topic/workflow-transformation