The system went live months ago. It works. The savings you put in the business case have not shown up in your numbers, and nobody can tell you exactly why.
That gap has been measured. A 2024 survey of C-suite executives on AI value (BCG, 2024), 1,000 executives across 59 countries, found 74% of companies have yet to show tangible value from their use of AI. The obvious reading is that AI does not work, and it is the wrong reading. That figure measures the distance between a system that runs and a business that has changed, and that distance is crossed, or not crossed, on the far side of launch day.
Every dollar of an AI investment is committed before launch and every dollar of return is earned after it, in the stretch when the build is finished, the project budget has closed, and the system belongs to nobody in particular. Payoff is not a function of elapsed time; it is a function of whether anyone is still holding the system when the return is supposed to arrive.
The Unowned Payoff Window
ML LABS engineered the backend of a medical-device platform that has reached clinical production in two countries and keeps expanding: the backend behind HeartSciences' MyoVista Insights. Their Director of Software Engineering put it on the record: "Omar did outstanding work designing and building the backend for our cloud-native AI-ECG platform." The platform was designed so that onboarding a new organization is configuration and never code: HL7 field mappings, enabled AI models, invoice pricing, storage provisioning. None of that paid anything on launch day; it pays as the platform grows, for as long as someone keeps the property true. An AI return is not an event but a position that has to be held.
Nor does the model hold still: research on temporal degradation in deployed models (Scientific Reports, 2022) tested 128 model-dataset pairs across healthcare, weather, finance, and transportation and found 91% of them degrading over time, many with no concept drift at all. Those systems were not attacked from outside; they aged.
So the payoff window and the ownership vacuum are the same window. Costs are legible from day one: invoices, cloud bills, a line in the plan. The return is illegible unless someone builds the instrument that makes it visible and keeps it honest, the discipline the measurement article is about. An unmeasured return is indistinguishable, to whoever signed the original check, from no return at all, and spend drifts upward inside the same blind spot.
Returns We Can Actually Show
Three systems produced returns ML LABS is cleared to state, each for a different reason.
The roaming optimization platform built for a top 10 global telecom company delivered over 12x return against engagement cost in the first year. It paid fast because its output was the decision rather than a report about it: routing moved from a manual review cycle the analysts could not finish before the traffic pattern shifted, to decisions the platform made itself. It also surfaced optimization corridors across 128% more of the network than the original scope targeted. Nobody had to change their behavior for that return to land.
The property valuation engine built for a PropTech platform hit its accuracy target, within 10% of closing price in dense metro areas for 90% of cases, in seconds instead of days, and the accuracy is not what made it pay. The confidence band did: the calibration deciding which properties the system was allowed to price automatically and which ones were routed to a human appraiser. Calibrating that band correctly was harder than producing the point estimate, and it is what made the 90% commercially meaningful. Their innovation director's account is that the system "lifted our real estate insights and client engagement."
A model that knows where it is wrong can be automated around. A model that does not know is a demo with a good headline number.
The third return was never designed. A hedge fund stored large volumes of unnecessary and polluted data without noticing: the aggregation step discarded information the models needed, and the storage structure duplicated what it kept. Correcting it cut storage costs by more than 60%, and their models performed 2% better. Their Head of Data's published account ends "it held up in production, reliable in a way this field rarely is", and the full telling belongs with the ownership argument, because that is what it is evidence of.
In none of the three did the payoff arrive because the model was good. It arrived because the output was wired to a decision, because the system knew when to disqualify itself, or because someone was still examining a platform everyone else had stopped examining.
When The Return Cannot Exist
All of this assumes there is a return there to collect. Sometimes there is not, and the most valuable thing an engineering partner can do is say so before the money moves. A major US TV network arrived at a scoping call holding a quote to build a full software system for a workflow that did not require one, and the deliverable of that call was "don't build this"; their AI Program Manager's published account is that it "saved us from a $200K mistake."
The gate is whether the cost of the problem is quantifiable today. When nobody can state what the current process costs in labor, errors, delay, or lost revenue, there is no way to prove a return later, which makes the quantification itself the honest first investment.
Ownership As The Return Mechanism
Match the move to where you stand. If nothing is live, the work is quantifying the problem. If a system is live and the number it was meant to move is still theoretical, the missing piece is not a better model; it is a standing owner who watches drift, spend, and failure classes, keeps the measurement honest, and ships what the measurement asks for.
Returns are held, not banked. Holding them is what Advisory exists to do: engineering ownership that stays attached to the system through the window where the money arrives, catching decay before it becomes a rewrite, and going looking for the returns nobody budgeted for, the way a hedge fund's storage bill turned out to be one. The systems that pay off are the ones that somebody is still holding when the return shows up.
References
- Boston Consulting Group. Where's the Value in AI?. BCG, 2024.
- Vela, D., Sharp, A., Zhang, R., Nguyen, T., Hoang, A., and Pianykh, O. S. Temporal Quality Degradation in AI Models. Scientific Reports, 2022.
