Design ROI is the wrong question. Ask what your interface is costing you.
Flip the ledger: bad interfaces are a live expense, and in the AI era they can quietly burn the entire model budget
Design leaders keep building investment cases for a function the company treats as optional. Flip the ledger: bad interfaces are a live expense, and in the AI era they can quietly burn the entire model budget.
AI Key Takeaways
“Design ROI” arguments fail in boardrooms because they accept the wrong frame: design as an optional investment with speculative returns, argued with numbers too good to be believed.
Interface costs are already on the books, filed under support, training, error remediation, abandoned funnels, and compliance exposure. They are simply never attributed to design.
Safety-critical industries already think this way: aviation prices the cost of the failure a checklist prevents, and nobody computes the checklist’s return.
AI raises the stakes. The model is rented plumbing available to everyone, and the interface is where adoption, trust, and the entire AI investment succeed or die.
The practical move is a cost audit: pull the support logs, count the workarounds, measure the steering effort, and attach the existing spend to the screens that generate it.
Somewhere right now, a design director is assembling a deck to prove her team deserves to exist. She has the McKinsey study showing design-led companies outgrow their peers. She has the famous claim that every dollar invested in UX returns a hundred. She has case studies, before-and-afters, a slide titled “The Business Value of Design.”
She will lose, and the numbers won’t be the reason. The exercise concedes the argument before it begins.
The frame is the defeat
The ROI argument accepts three premises the moment it opens its mouth. That design is optional: a discretionary investment competing with everything else on the CFO’s list. That its value lives in the future: a return to be realized, someday, if the projections hold. And that the burden of proof sits with design, in a way nobody asks of legal, security, or accounting.
Accept those premises and the strongest evidence in the world won’t save you, because the evidence has a credibility problem of its own. A hundred-to-one return is a vendor’s number, and a CFO hears it as one. If design reliably returned a hundred dollars on one, capital would have flooded the function decades ago without a single deck being written. The ROI genre’s numbers are so good they refute themselves. Executives reject the design case because it arrives shaped like marketing, and for no deeper reason than that.
The frame that survives contact with a boardroom runs the other way. Your interface is not a pending investment, but a live expense. It is spending your money right now, today, whether or not anyone is watching the meter. So ask yourself what your current design already costs.
The ledger nobody reads
Interface costs are real, present, and sitting in your books under other names.
They’re filed under customer support: every ticket that exists because a flow was ambiguous is a design cost wearing a support badge. Under training: every hour spent teaching employees to operate internal software is the price of decisions someone made in a tool, years ago, and nobody has audited since. Under error remediation: rework, refunds, reconciliation. Under revenue operations: the abandoned carts, the stalled onboarding, the enterprise feature nobody adopts. Under shadow process: the spreadsheet an entire department secretly uses because the official system is unbearable. You paid for the system, and now you pay again in duplicated labor and unmanaged risk.
Occasionally the ledger surfaces where everyone can see it. Jared Spool’s famous “$300 Million Button” documented a retailer whose registration form (one form, one forced login) suppressed roughly three hundred million dollars a year in sales before anyone thought to question it. In 2020, Citibank wired $900 million to Revlon’s lenders by mistake, an error a federal court traced in part to a loan software interface whose checkbox logic three separate employees misread the same way. Citi eventually recovered the money, after two years of litigation it had no guarantee of winning. Nobody had ever asked the ROI of that screen. The screen billed them anyway.
These cases make headlines because of their size, but their structure is ordinary. Every organization runs interfaces that misprice themselves this way. The losses are distributed finely enough (a ticket here, an abandoned session there, a workaround nobody mentions) to stay below the threshold of attribution.
The cost sits in plain sight, unassigned.
Industries where failure is priced already know this
I spent five years inside aviation manufacturing. Nobody in that world computes the return on investment of a checklist. Nobody builds a business case for making a warning label legible.
The entire discipline runs on the opposite calculation: what does the failure cost, and what does it cost to prevent it?
Safety-critical industries figured out long ago that prevention argued as “investment” loses to prevention argued as “exposure,” because exposure is already on the risk register, and investment is competing with next quarter’s priorities.
Healthcare knows it. Finance knows it, at least after each fine. These are also the industries where design is least discussed as a differentiator and most embedded as a control, and that is no coincidence. When the cost of interface failure is legible (a grounded aircraft, a mis-dosed patient, a nine-figure wire), nobody needs persuading that the interface is a serious object. The design conversation only degenerates into ROI theater in industries where the failure costs are just as real but nobody has done the accounting.
AI turns the leak into a flood
This argument mattered before AI. It’s existential now, for one structural reason: the model is plumbing.
When every competitor can rent comparable intelligence from the same three vendors, the model is a shared cost line, and the layer where a human decides whether to trust the thing is the layer that isn’t shared. That layer is the interface. I’ve made this argument at length before. What matters for the cost math is this:
An enterprise AI deployment is a stack of sunk costs (licensing, integration, security review, data work) capped by a thin surface where an employee decides, in the first handful of sessions, whether this tool is worth steering. If the interface fails to signal what the system can and can’t do, fails to make uncertainty legible, fails to help the user recover the first time the model stumbles, adoption dies quietly. And when adoption dies, the failed interface bills you for the entire stack underneath it.
The widely cited MIT finding that some 95% of enterprise generative-AI pilots deliver no measurable P&L impact is usually read as a model problem or a strategy problem. Having watched these deployments up close, I’d put it differently: the pilot dies in the gap between the demo and the Tuesday afternoon workflow, and that gap is a designed surface. The most expensive interface in your company is now the one wrapped around your AI spend.
This is a cost argument, and I want to keep it one; promising that better design will make the AI investment soar would be the ROI trap again. The claim is that the current interface is where your AI investment is presently evaporating, at a measurable rate, and that the meter runs whether or not you look at it.
Run the audit
So retire the ROI deck. Replace it with cost accounting, presented the way a CFO already thinks.
Pull twelve months of support tickets and tag the share caused by comprehension failures, as distinct from product defects. In most organizations that share is the majority, and it converts directly into dollars.
Count the workarounds: every unofficial spreadsheet and side channel is a receipt for an interface that failed, so price the duplicated labor.
Put a number on training that exists only to compensate for unintuitive internal tools.
Measure abandonment where money leaks (onboarding, checkout, renewal) and attach the existing revenue loss.
For AI products, measure steering effort: how many turns users need to get an acceptable output, and where they give up and go back to the old way.
Then present one number: this is what the current interface costs per year. Fixing it stops an expense the company has already verified.
Do you know what your interfaces cost you last year, and if you don’t, who in your organization is supposed to?
NA: AI-assisted tools were used for transcription, reference formatting, and language editing. All intellectual content and conclusions remain solely the author’s.