Can AI Actually Create 20% in Construction Cost Savings?

The trade email hits before six. MIT and one of the largest builders in the country ran the numbers on AI in construction, and the headline is 17 to 20 percent off total cost, 22 to 25 percent off the schedule. You read it standing up. Somewhere behind your eyes you are already doing the math on the job that closed last month at a margin you did not bid.

So here is the straight answer on AI construction cost savings, and then the part of that report almost nobody quoted, which is the part written about firms your size.

Does AI save money in construction? Yes, and not the way that headline reads.

The study is real research and the number is honest. It is also not a promise, and it was not measured on a job like yours.

Here is what it is. Suffolk Construction, working with MIT's Center for Real Estate and the MIT Media Lab, took one building: a 180,000 square foot apartment project in San Francisco, $180 million, finished in 2024. The building was already up. They went back through it and modeled where AI would have saved time and money if it had been used across six areas at once.

Their own description of the work, on page 24 of the paper: "a methodology statement, not a definitive causal analysis." They reviewed 178 sources, ran a roundtable in March with more than 50 people, and surveyed executives. They say plainly that the evidence is not yet enough to support causal conclusions.

That is a research team being careful. Read it as a direction somebody smart is pointing, not a result you can book.

The number is one building, studied backwards, and it still tells you something

Strip the headline away and two findings survive, and both are useful to a builder running $5 million to $50 million.

The first is where the savings come from. The six areas they studied are design, prefab, permitting, scheduling, subcontractor coordination, and supply chain. The report is emphatic that the gains do not add up, they compound. Suffolk's technology chief put it as one plus one plus one equaling six rather than three. Fix the design handoff and permitting, procurement, scheduling and sub coordination all get easier, because they were all waiting on the same information.

The second finding is the one they were brave to print. The 20 percent only exists if the six can reach each other. Their words: "that figure assumes the levers can talk to each other." And then, a sentence later, that capturing it "requires a shared data infrastructure layer, something the levers themselves do not provide."

A national builder and two MIT groups modeled the whole opportunity, then said the piece that makes it work is the piece nobody sells.

Then I got to page 23.

The back of the paper is a list of instructions, one section per kind of company. Owners. Architects. Regulators. Manufacturers. And one addressed to technology companies, which says this:

"Invest in tools for smaller contractors and subcontractors, not just tier-one GCs. The productivity gains available in the long tail of the industry are large but only accessible with tools designed for firms without dedicated technology teams."

That is MIT and Suffolk telling the software industry it skipped you.

Not that you are behind. Not that you should adopt faster. That the tools sold to firms your size were built for somebody else, and the people who build them should go fix that.

If you have spent four years buying construction software that half your office refuses to open, you already knew. It is worth something to see it written down by the people whose report is about to get quoted in every sales deck in the industry.

Why nobody built AI construction tools for a $15 million builder

The honest reason is not neglect. It is math that used to be true.

Software shaped around how one company runs used to cost seven figures and take two years. So software companies did the sane thing: build for the average of ten thousand firms and sell it ten thousand times. The trouble is that nobody runs the average company. A custom builder in Monterey and a remodeler in Denver have different draw cycles, different sub relationships, different words for the same job. The average serves neither.

The firms in that study have technology departments. Suffolk runs something they call Mission Control across hundreds of projects. You have an office manager who knows where everything is, and a spreadsheet that exists because two systems will not speak.

What changed is the cost of the middle. Building a system around one firm's real operation is no longer a seven-figure project. That is the genuinely new thing in construction right now, and it is not what the headline number is about.

The test you can run this week, for nothing

Do not start with a tool. Start with a count.

Pick one job you closed in the last year. Walk it from contract to closeout and write down every point where a number left one system and entered another because a person carried it. The selection that moved from the design meeting to the estimate. The vendor invoice that became a budget line. The change order that had to reach the schedule, the billing, and the homeowner. The draw package somebody assembled by hand.

Count the handoffs. Then put a name next to each one.

That list is your version of the study's finding, measured on your own job instead of a tower in San Francisco. Every line on it is a place where a number can arrive wrong, arrive late, or not arrive. You will have it in an afternoon, and it costs nothing.

Most firms that run this are surprised twice. First by the number of handoffs. Then by how many of them have the same person's name next to them.

What we build, and why it starts with a read of your firm

That list is also where we start.

A Ridgebeam is built for one firm, in three parts, in this order. First a business analyst reads how your operation runs and says where the work is stuck and what to fix first. That read is the point. It is the difference between buying software off a menu and fixing the thing that is costing you. Then automations get built for what the read found, not for what a product roadmap decided. Then the agents run in the middle of it day to day, across the systems you already own, talked to in plain words.

The paper called the missing piece a shared data layer. On a custom build it looks less abstract than that. It looks like the vendor invoice reading itself into the right budget line on the right job, with a person confirming before anything bills.

On one build we rebuilt five real jobs' billing to the penny. It found the office had under-charged by $362 that nobody had caught. Same firm, draw packages that took about 25 hours each, twice a month, now take about two.

A Tuesday in October

The draws for the month are assembled when you open your laptop. The vendor invoices came in by email overnight, got read, got matched to budget lines, and are sitting in a queue waiting for you to confirm, with the two that ran over their allowance flagged at the top. The change order from Thursday's design meeting reached the estimate, the schedule, and the homeowner, and you can see that all three happened.

Nobody stayed late to make that true. That is the whole product.

If you want to know whether your firm is a fit, the Assessment is $2,500 and you keep what it produces. And if we cannot solve your problems, we will tell you that too.

The bottom line

The report says the industry can take 20 percent out of the cost of a building, and then says the part that makes it possible does not exist yet for firms without a technology department.

It exists. It is just built one firm at a time now, instead of ten thousand at once.

Frequently asked questions