The construction industry has spent three years being told that AI will transform it and almost no time being told by how much. On 16 September, Suffolk — the Boston contractor that ranks among ENR’s largest US builders — published a white paper with the MIT Center for Real Estate and the MIT Media Lab’s City Science group that tries to answer the question with arithmetic instead of adjectives.
The paper is called Construction in the Age of AI: An Industry White Paper and Research Roadmap, and its headline finding is that the coordinated application of six AI-enabled levers “could generate approximately 17–20 percent total cost savings and 22–25 percent total schedule savings on the sample multifamily residential project analyzed.”
That is a big number, and it will be the number that travels. But the most valuable sentence in the document is the footnote explaining where it came from, and that sentence deserves to travel with it.
The six levers, and what each one is worth
The study identifies six processes where it sees the greatest near-term potential, and attaches a maximum efficiency gain to each. Per the white paper’s own table:
| Lever | Described as | Potential efficiency gain (maximum) |
|---|---|---|
| Design automation | Generative design and BIM AI | Up to 39% reduction in design cycle time; up to 21% design and engineering cost savings |
| Offsite manufacturing | AI-optimized factory prefabrication | Up to 32% reduction in project timelines; up to 14% total cost savings |
| Permitting | Automated code-compliance review and application preparation | Up to 24% reduction in project time; up to 8% cost savings |
| Scheduling | ML-driven sequencing and optimization | Up to 18% reduction in project timeline; up to 11% total cost savings |
| Skilled labor and subcontracting | AI trade development, coordination and performance management | Up to 17% project time savings; up to 13% total cost savings |
| Supply chain and procurement | Predictive lead time, delivery optimization and process automation | Up to 15% reduction in total project time; up to 10% cost savings |
Design automation carries the biggest headline percentages in that table, though the columns are not measured against the same baseline — its figures are stated against design cycle time and design-and-engineering cost, while offsite manufacturing, scheduling, labour and procurement are stated against total project time and total project cost. The levers are not directly rankable by those numbers alone, and the paper does not rank them that way. It makes a different argument: “Design Automation is the clearest upstream enabler.” Its argument is dependency, not magnitude — an AI-enabled design process is what produces the kit of parts that makes offsite manufacturing work, the structured model that makes automated code checking possible, and the bill of materials that makes procurement automation more than autocomplete.
The paper backs this with a ranking exercise among its expert participants, and then does something unusual: it discloses the bias in its own instrument. Design automation ranks first on raw Borda scores with 206 total points, “though 89 of those are self-votes from the design cohort.” Permitting ranks third overall — but when each expert’s votes for their own primary lever are excluded, permitting drops to the least valuable lever of the six. Publishing the figure that undercuts one of your own findings is more candour than this genre of document usually volunteers.
The project the model was run on
The 17–20% figure is not an industry average. It is the output of a financial model applied to one reference case: a multifamily development of 180,000 square feet, completed in 2024, with a total duration of 4.5 years from initial design through commissioning and closeout. The caveat sentence describes it as “a $180M multifamily San Francisco project.”
The paper’s own table for that project puts a $179 million cost baseline against $32 million of modelled savings — 18% — and a 51-month schedule baseline against 11 months saved, or 22%. The table carries its own asterisk, noting the total “doesn’t equal to sum of previous lines due to overlap of project phase.” The savings are concentrated where the paper’s logic says they should be: 20% off the design phase cost, 30% off design duration, and 14% off the construction phase schedule on a base of 30 months.
Then it converts that into the language developers actually fund projects in. Coordinated application of the six levers, the paper says, could increase illustrative unlevered IRR by approximately 5–6 percentage points on the sample project — “for example from 15–20 percent to 20–25 percent.” For a marginal deal, that is the difference between a project that clears an investment committee and one that does not. It is also, the paper is careful to say, illustrative.
The sentence that should travel with the number
Here is the footnote under the lever table, in full:
Potential Efficiency Gains are estimated from a comprehensive review of existing literature of means and methods applied, in Architecture, Engineering, and Construction (AEC) and adjacent industries predominantly from early-stage pilots (so they should be treated as directional rather than construction-validated benchmarks), as well as means from survey respondents anchored to a $180M multifamily San Francisco project.
Directional rather than construction-validated benchmarks. The authors are saying, in their own document, that these percentages are assembled from pilots and adjacent industries rather than measured across completed buildings. The study characterises its own contribution as, to its authors’ knowledge, among the first attempts to piece together disparate evidence into a directional view — which is a claim about the state of the evidence base, and an accurate description of what the paper does: it is a synthesis and a research agenda, not a measurement.
That framing matters because the gap between claimed and demonstrated AI value is the defining problem of this market. This newsroom has written before about the distance between what AI is forecast to automate in construction and what the industry actually uses daily. A study that publishes a 20% number and simultaneously tells you the number is directional is being more useful than one that publishes the number alone.
Where the paper is most concrete
The permitting section is the part least dependent on modelling, because what it mostly does is count things. The paper puts the number of independent permitting jurisdictions in the United States at more than 20,000, each with its own zoning codes, review processes and approval timelines. It then narrows to the authors’ own city. Boston, it says, splits 48 square miles across 429 zoning districts, under a code running to nearly 3,800 pages that has not been fundamentally revised since 1964; 42% of the city’s parcels are non-conforming under current lot-size rules.
None of that requires an AI forecast to be alarming. It is a description of a bottleneck that exists whether or not a single model is deployed against it, and it explains why the paper treats permitting as a navigation problem rather than an automation one — technology routing through fragmentation rather than removing it.
The labour section is similarly grounded. The paper cites an Associated Builders and Contractors analysis putting the US skilled-worker shortage at 456,000 additional workers needed in 2027, and frames that as a delivery-capacity constraint rather than a hiring one. On offsite manufacturing, it reports that Volumetric Building Companies, which it describes as one of the leading US producers of full volumetric buildings, reports up to 47% schedule savings across geographies, while cost savings remain market-dependent — driven by union structure, field labour cost and logistics distance. That last clause is the honest part: prefabrication’s schedule case is portable and its cost case is not.
What is not known
The paper does not claim any project has achieved 17–20%. It models what would happen if six levers were applied together on one reference case, and states that the underlying per-lever figures are drawn predominantly from early-stage pilots. Whether the levers compound as modelled is the open question, and the paper says so directly: the IRR figure “assumes the levers can talk to each other,” and capturing that value requires a shared data infrastructure layer — “something the levers themselves do not provide,” in the paper’s words, that standardises outputs at every handoff.
That is the same conclusion Suffolk reached from the other direction in June, when it launched Jobsite of the Future and the interesting part turned out to be the decade of data assembly underneath it rather than the software on top. The white paper is arguing, with more arithmetic, that the constraint on construction AI is not model quality. It is whether the six systems that would need to hand work to each other are connected at all.
Methodologically, the study drew on academic literature, case studies, expert interviews, survey input and a roundtable of over 50 experts from across the industry. Construction Dive reported the findings on 16 September, noting that Chin pointed to a roadmap of next steps including more studies and better data standards for construction.
“Construction is at an inflection point,” said John Fish, Suffolk’s chairman and CEO. “AI presents a real opportunity to transform the way we build. The most meaningful progress will come when the entire ecosystem aligns around better data, smarter workflows and shared accountability.”
The sharper line comes from James Scott, co-lead of the study at the MIT Center for Real Estate, who described it as a starting point rather than a conclusion: “The next step is to keep building the data foundation needed to understand where AI has the strongest impact, where the limits still are and how those findings can be translated into better decision-making across real projects.”
Which is another way of saying that the most valuable output of this paper may not be the 20%. It may be the paper’s own insistence that the number is not yet construction-validated — and its call for the industry to go build the evidence base that would validate it.
