Five percent of AI programs in commercial real estate achieved their stated objectives. One in twenty.

That’s McKinsey’s finding, and it deserves to be read without softening. Organizations spent real money, ran real pilots, stood up real dashboards — and nineteen out of twenty walked away without the outcome they paid for. In most industries, a 95% failure rate would end the product category. In CRE, the budgets keep arriving, because the 5% — cases like the documented $470K in annual savings at one university — prove the ceiling is real.

The mechanics: how a program gets to “failed”

The number isn’t produced by bad algorithms. It’s produced by a sequence so consistent that operators we interviewed described it almost identically without knowing each other.

The program launches against the operation’s existing data. The model needs to know what assets exist; the asset registry says one thing, the building says another — equipment added, removed, or replaced without the record changing. The model needs maintenance history; the history is a spreadsheet column reading “Done ✓” with no findings, no readings, no parts, no costs. The model needs failure data; failures were never coded as failures, just as repair invoices in accounting. The model needs sensor signals; the critical assets have none.

So the model does what models do with thin, contradictory input: it produces output anyway — confident, plausible, and untrustworthy. The team spends months chasing false positives and explaining misses. Trust erodes per alert. Within a year the dashboard joins the category of software everyone pays for and nobody opens, and the post-mortem politely blames “change management.” The data was never usable. Everything after that was theater.

What “data AI can’t use” actually means

The phrase sounds abstract, so make it concrete. AI-usable maintenance data has five properties, and most operations fail at least four:

  • Complete — every asset in the registry, every work order captured, including the ones resolved with a phone call and no record.
  • Structured — findings as fields, not as “fixed the thing, runs fine now” in a notes box.
  • Historical — depth measured in years, because pattern detection needs patterns, and a decade of spreadsheet rows recorded dates while destroying everything else.
  • Connected — the work order linked to the asset linked to the cost linked to the sensor, so the model can trace cause to effect.
  • Verified — reflecting work that actually happened, which is its own problem in operations where sign-offs and reality diverge.

Run your operation against those five properties honestly. That gap — not the model, not the vendor — is the 95%.

What most organizations do with this number

They proceed anyway, and the reasons are familiar: the budget was approved for AI, not for data plumbing; the demo looked spectacular (demos run on the vendor’s clean data, never yours); and “fix the data first” sounds like delay while “deploy AI” sounds like progress. So the data work gets skipped, which guarantees membership in the nineteen — and worse, it salts the earth: the failed pilot becomes the organizational memory that “we tried AI and it doesn’t work,” deferring the real fix by years.

The unglamorous truth: the path to working AI runs through plumbing, and the plumbing is more valuable than the AI.

The condition that moves an operation into the 5%

The unglamorous truth: the path to working AI runs through plumbing, and the plumbing is more valuable than the AI. Data that is complete, structured, historical, connected, and verified doesn’t get produced by a cleanup project — cleanup projects decay the day they end. It gets produced by infrastructure where capture is a byproduct of operating: work orders that carry asset, findings, parts, cost, and verification as structured fields filled in the flow of the job; sensors feeding readings into the same asset records; payments releasing on verified completion so the record and reality can’t diverge.

STAGE 1 Structured Capture

Capture complete asset data, findings, parts, and costs as structured fields directly within the workflow.

STAGE 2 Connected History

Link work orders to assets, costs, and sensor signals to establish a deep, pattern-rich historical record.

STAGE 3 Verified Execution

Tie payment releases to verified completion to ensure the digital record matches operational reality.

Build that layer and something underrated happens — the operation starts getting smarter before any AI arrives, because clean connected data answers most operational questions with plain arithmetic. That’s the sequencing Sweven FM is built around: the data infrastructure is the foundation, and the intelligence is what the foundation eventually makes trustworthy.

The Data Census

The 5% statistic, read correctly, isn’t a verdict on artificial intelligence. It’s a census of data infrastructure in commercial real estate — taken by an expensive instrument. The question it leaves every operation is the same one: if the census visited you tomorrow, which side of it would your data put you on?

Sources: