Tight scope, named benefit, defensible measurement: every instinct that gets an AI investment approved also forecloses its largest return. The by-products your programme will depend on in three years are being priced at zero today.
The 2.4-metre mirror inside NASA’s Nancy Grace Roman Space Telescope was built to look down at Earth. The National Reconnaissance Office handed NASA the surplus spy-satellite optics in 2012, and the telescope built around them will map dark energy. Somewhere in that survey data, with no line item in the science case, it will also find asteroids, including ones that cross our orbit. Nobody funded that. It is by-catch, and the by-catch may be what justifies the telescope to people who never once worried about cosmic expansion.
Almost every serious capability investment has this shape. Enterprise AI business cases are structurally unable to describe it, and the gap costs more than most failed pilots do.
The second mission is free only if the first was built wide
Roman will detect moving objects for an unglamorous reason: it revisits the same patch of sky every fifteen minutes, for months. Repeated imaging of a fixed field is the exact recipe for spotting anything that shifts between frames. The asteroids fall out of a cadence chosen for something else entirely.
The enterprise version plays out constantly and gets noticed rarely. A finance function buys document extraction to cut accounts-payable costs: clean case, defensible payback. Two years on, the pipeline has quietly produced a line-item time series of what every supplier charged for every component, month by month. To procurement, that is a negotiating position worth multiples of the AP saving that paid for it.
It exists only if someone kept the structured output instead of treating it as scaffolding. Whether that asset exists was decided by an engineer choosing a retention policy, not by anyone in the room where the investment was approved.
Three legitimate gates, one foreclosed upside
Ask a CFO for money and you will be asked for tight scope, a named benefit, and a measurement plan. Each instinct is correct. Together they routinely destroy the larger return, in three places.
Legal. Under purpose limitation, data collected for one specified purpose cannot simply be repurposed later. An impact assessment drafted narrowly, precisely because narrow drafts get signed off quickly, becomes the ceiling on everything the data can subsequently do.
Procurement. The standard clause reads: customer data may be processed solely to provide the services. Fine, until the outputs you generated on the vendor’s platform become the input to a use case the contract never contemplated, and the export path turns out to be a CSV of the last ninety days.
Organisational. The sponsoring function owns the system, so it owns the output. The supplier-price series sits in a finance data store that procurement cannot query and does not know exists.
None of these is governance failing. Each is governance working as designed, on an assumption the evidence does not support: that a system’s value is bounded by its stated purpose.
Cadence beats resolution, and overwriting is the expensive habit
The asteroid detection comes from revisit rate, not optical quality. A sharper single image finds nothing that moves. A blurrier image taken repeatedly finds everything.
Enterprise data estates are built for resolution and almost never for cadence. The dominant pattern is current state: a lead moves through eight status changes and the field holds one. A price is revised eleven times and the catalogue holds the eleventh. Each overwrite saves trivial storage and permanently destroys the only evidence of how the thing behaves over time.
A dashboard needs the current number. A model that predicts which deals stall, which suppliers will raise prices, or which customers are drifting needs the sequence, and the sequence cannot be reconstructed after the fact at any price. The cheapest capability decision available to most organisations is to append rather than replace. Almost nobody makes it deliberately, because it never appears in a business case as a benefit.
Capability only its sponsor can see produces only what its sponsor imagined
Roman’s asteroid papers will largely be written by people outside the dark-energy team, because NASA’s survey data lands in a public archive where strangers are allowed to look. Most companies have no equivalent archive and no equivalent culture.
The second cheapest capability decision, after appending: widen who may query what a system produces. An interface reachable only through one team’s UI is an asset priced at one team’s imagination. This is also why portfolio reviews organised around stated intent systematically miss where value accrues. The revealing question is not “which of our AI projects delivered ROI” but “which of our systems now produce data or interfaces something else could be built on, and did we plan any of them”.
The substrate question, asked from Dubai
Dubai’s 295,000-company agentic AI program will put agents into ordinary firms at speed. What separates a two-week agent deployment from a two-quarter one is almost always substrate: whether the process history was retained, whether the records are structured, whether internal systems are reachable programmatically rather than only through a UI.
Those are infrastructure decisions with an AI consequence that arrives years after the decision is taken. Firms that default to retention, append-only records, and exposed interfaces will capture their by-catch. Firms that scoped everything tightly will fund each new capability from zero, and the 95% pilot failure pattern will keep looking like a technology problem.
Before the next AI line item is approved, map what your existing systems already throw off. The five-step consulting method starts exactly there: which processes hold retained history, which outputs are assets nobody owns, and where an agent could run on substrate you already paid for. The build that follows starts wide enough to keep its by-catch.
Frequently asked questions
Why do narrowly scoped AI business cases destroy long-term value?+
Three legitimate gates do it: data-protection assessments drafted narrowly to get signed off become the ceiling on what the data may later do; vendor contracts restrict processing to the contracted service; and the sponsoring function owns the output, so nobody else knows the asset exists. Each gate works as designed, on the false assumption that a system's value is bounded by its stated purpose.
What is the cheapest high-value data decision most organisations get wrong?+
Append instead of overwrite. Most systems store current state: a lead's latest status, a price's latest revision. Each overwrite saves trivial storage and permanently destroys the sequence, and models that predict stalling deals, price rises, or churn need the sequence. It cannot be reconstructed later at any price.
How do you find the hidden value in AI systems you already run?+
Ask which systems now produce data or interfaces that something else could be built on, rather than which AI projects delivered their stated ROI. A document-extraction pipeline bought to cut invoice-processing costs may have quietly produced a line-item price history of every supplier, worth multiples of the saving that justified it.