Can AI revolutionize EPC?

Start with what an engineering, procurement and construction contractor is: one firm that designs the project, runs its supply chain and builds it, engaged in a single contract after the developer has secured land, permits, power and a tenant. Modon’s complaint is that both available contract forms reward the wrong thing. Cost-plus pays a margin on whatever the project ends up costing, so the contractor earns more when the job costs more and takes longer. Fixed-firm names a price, but against a scope written tightly enough that the inevitable discoveries during design trigger change orders, which is where Modon says the margin actually gets made, negotiated when the developer can no longer walk away. He steelmans it though, and Kann presses him to: these projects run from about $100 million to several billion, involve hundreds of people, and genuinely cannot be costed accurately up front, so the flexible contract is a risk-mitigation device rather than a swindle. The misalignment worsens the more custom the project, which makes first-of-a-kind facilities the worst case and data centers bad because their interiors keep changing with chip generations and cooling choices.

The second problem is less obvious and matters more. Manufacturing has a cost per widget, so it gets a learning rate. Projects are each treated as an n of one, even a company’s tenth refinery, so nothing accumulates. Some of that is intrinsic, since every site differs, but Modon argues most is self-inflicted: because engineering has a high marginal cost, any tweak reintroduces a near-complete redesign rather than just the delta, so design learning never compounds. Kann supplies the sharpest illustration almost by accident. Modon credits solar’s cost decline to this kind of iteration, and Kann corrects him that it was module cost that fell while installed cost did not fall nearly as fast, which is the construction problem stated as a fact rather than a theory.

That sets up the actual mechanism, and it is not the one the title suggests. Engineering is only 3% to 10% or 15% of project cost, so cutting engineering fees is not the prize; spending engineering to buy certainty is. Today a conceptual design gives roughly plus or minus 50% on cost, front-end engineering narrows that to about plus or minus 10%, which is enough for a lender, and the project reaches final investment decision there with only about 30% of the engineering done. Finishing the rest first would burn development capital, and it is cheaper to spend the lender’s money later. Modon’s claim is that if engineering becomes cheap and fast enough to reach 100% definition before commitment, two things follow. You know the complete equipment and bulk materials list on the day the project is funded, so you can buy everything that day or hedge it, which removes the commodity and tariff exposure that makes contractors demand flexibility in the first place. And because you are no longer freezing design decisions early to control engineering spend, you can search a wider space instead of settling into a local optimum. Together those are what let him write contracts with no change orders and carry the overrun risk himself.

The technology claim is deliberately modest. Modon says they have not redefined engineering, they have augmented engineers so the same time and capital buys more definition. The platform is internal, not a product they sell, and its main move is consolidating a fragmented tool stack onto a single data model so no piece of data lives in ten places. That consolidation is what makes agents usable: a trade study a junior engineer would spend a week on, such as belt versus pneumatic conveyance for a material-handling problem, can be delegated to a large language model grounded in the project’s own data, vendor spec sheets and simulation tools, then handed back to the engineer who asked. His explanation for why incumbents have not built this is the incentive argument again, that a contractor has no reason to want software which reduces its billable hours, the way a law firm does not.