How to evaluate the ROI of engineering AI

Why there is no ROI multiple here

Every vendor offers a number: five times, ten times, pays for itself in a quarter. We do not, on purpose. A multiple computed on a vendor's assumptions tells you nothing about your projects, your rates or your revision loops. But do not read the silence as modesty: the gains are large. On the mechanical work, we cut anything from a third of the time to nearly all of it, depending on the task. How large it is for your firm is the firm-specific part, and the only credible version of that number is the one you calculate yourself. What follows is not a calculator with our numbers in it. It is the list of terms to fill in with yours.

One thing before the terms: the gain has two sides, and most business cases see only one. The first is plain economics: when a deliverable takes a third of the time, the same engineers get through more of them, sooner, and with more margin left on each. In an industry that cannot hire its way out of its shortage, more throughput per engineer is a heavy term in its own right, not a bonus. The second side is that the same fee also buys more checking and more tested alternatives before a design hardens. The terms below capture both.

Cost one deliverable, from its own time sheet

Start with a deliverable you can hold: a noise assessment, an environmental report, a technical description. Take one you finished recently and open its time sheet. Where did the hours go? On one stormwater study we took apart, the hours fell four ways, with writing and calculation the largest blocks and review the smallest. The proportions will differ on your work, but the four verbs will not: research, calculate, write, check.

Multiply those hours by loaded rates and you have the production cost of the deliverable, the first term. It is the one most business cases stop at, and it is real: every hour freed can go to the next deliverable. It is also small money that governs large money: design and engineering is typically 4 to 12 percent of a project's construction cost while it shapes 100 percent of that cost and quality, in Norconsult's phrasing. The three terms that follow are where the rest of the return hides.

The revision loops are where the hours multiply

First the revision loop. A brief changes: a unit count shifts, an entrance moves, a parking ratio is revised. Each change is small. Each one ripples through the deliverable: new inputs, recalculated models, rewritten sections, results to check all over again. Studies like this routinely run over their hour budget, and under a fixed fee those extra hours are not billed onward. They are absorbed, by the firm and by the engineers' evenings.

So price it directly. Take one project and count: how many times did the brief change? Each time, what share of the calculation-and-document chain did someone re-run by hand? Multiply the hours by your rate, by the number of loops. That is the revision cost, and it is invisible on the invoice because it was eaten, not charged. It is also the term AI moves most: when re-running the chain costs an afternoon instead of a week, the loop stops being a threat.

Then the review cycles. Someone checks the numbers and the text before delivery, and on a busy project someone checks again. Review is usually a small slice of the hours: a colleague reading another's calculations on a Thursday afternoon between two deadlines of their own. Count your own, reviewer-hours per deliverable, times the number of passes, times the rate, and add it to the total. But note what the number hides. A human reviewer samples: nobody re-derives every figure in a hundred-page deliverable, so they check what experience says is risky and trust the rest. Most of the time that holds.

The two terms you forget: the error that shipped and the option you skipped

The first is catch-rate value. Ask a blunt question: what did the last error that shipped cost you? Not the one caught in review, the one that got past it. The Get It Right Initiative puts the direct cost of construction error at around 5 percent of project value, and the total, once latent defects and indirect costs are counted, at 10 to 25 percent: by GIRI's estimate, around seven times the UK construction industry's annual profit. Most of that error is born upstream, in documents. A machine that re-derives every figure and flags the deviations does not replace your reviewers; it raises the catch-rate before they sign. In live projects it has already caught errors that experienced reviewers read past. Put your own number on it: the rework, the claim, or the site cost of the last mistake that reached the field.

The second is iteration value. What was the alternative you set aside for lack of time worth? A second alignment carried through to quantities and cost, not just sketched. When Toyota's engineers kept more alternatives alive and committed late, they were roughly four times as productive and built better cars. Infrastructure design usually affords one iteration, sometimes two, because each costs weeks. The defensible version of the gain is exactly one more iteration per project, at the same fee. Price it as the value of a better-chosen option: less rework downstream, a tighter estimate, and fewer of the surprises that become claims.

The five terms, and how to fill them in

Put the five terms in one place. The left column is the same for every firm. The right column is yours to fill from a project you can name.

Fill the right column and you have your business case, in your currency, on your projects. You will notice what is not in the table: a payback period, and a multiple with our name on it. Both are missing on purpose. The return on engineering AI has two parts. One is capacity: the same engineers get through more deliverables. The other is quality: the same fee buys more checking and more alternatives before commitment, with the signature still the engineer's. That is a number only you can compute, and once you have, it tends to make its own argument.