What Your AI Failures Are Trying to Tell You
Somewhere in your organization right now, there is an AI pilot that has stalled. Not failed, exactly — failure would require someone to declare it, and nobody will. It launched with a kickoff deck, produced a few impressive demos, generated a press-release-grade anecdote about saved hours, and then settled into a state best described as renewed but unverified. The invoice arrives on schedule. The transformation does not.
If your organization is typical, there are several of these. And if your organization is typical, the question being asked about them is the wrong one.
The question being asked is: Why isn’t the AI working?
That question assumes the AI is the variable. It sends teams off to evaluate different models, hire prompt consultants, attend vendor briefings about the next version that will finally close the gap. It treats the stalled pilot as a procurement problem — we bought the wrong thing, or configured it badly, or trained people insufficiently.
The better question — the one almost nobody asks, because the answer is uncomfortable — is: What is the AI’s failure telling us?
Because here is the thing about these systems that the case studies will never say: AI does not fail randomly. It fails precisely where your organization was already broken. You just couldn’t see the breakage before, because human beings were quietly absorbing it.
The workaround economy
Every organization runs on an invisible layer of human compensation. The analyst who knows the Tuesday report is wrong and mentally corrects it before forwarding. The operations manager who knows that “shipped” in the system means “probably shipped, check with Dave.” The salesperson who maintains a private spreadsheet because the CRM’s data can’t be trusted. The veteran who knows which policy is real and which is theater.
None of this appears in any process diagram. It is unbooked, unmeasured, and largely unconscious — people don’t experience themselves as compensating for broken information systems. They experience themselves as knowing how things work here. That knowledge is the connective tissue that lets dysfunctional information flows produce functional outcomes anyway.
This is why organizations can carry astonishing amounts of dysfunction indefinitely. The dysfunction is real, but it is priced in — absorbed, distributed, and hidden inside the salaries of people doing two jobs: the job on their title, and the unacknowledged job of routing around everything that doesn’t work.
Now introduce AI into this environment.
The machine does not know that the Tuesday report is wrong. It does not know to check with Dave. It has no access to the private spreadsheet, no tenure, no instinct for which documents are load-bearing and which are corporate fiction. It takes the organization’s information at face value — because face value is all it has — and it amplifies what it finds, fluently, confidently, and at scale.
The result is not transformation. The result is your existing dysfunction, re-emitted at machine speed, with better grammar and an invoice attached.
This is what most “AI failures” actually are. The model didn’t malfunction. It faithfully processed what you gave it. The failure is a reading — an instrument detecting something that was always there.
The contrast agent
In medicine, a contrast agent is a substance you introduce into the body not to heal anything, but to make the invisible visible. The dye doesn’t cause the blockage. It reveals where the blockage already was.
This is the most useful way to think about what AI deployment is actually doing inside organizations right now — usefully, accidentally, and at enormous expense. Every place the AI produces confident nonsense is a place where your source information was already unreliable. Every workflow where the AI’s output requires so much human review that it saves no time is a workflow whose inputs were never as clean as the process diagram claimed. Every department where the pilot quietly died is a department whose actual decision process never matched its official one — the AI was automating a fiction.
Mapped honestly, your AI failures are a diagnostic image of your organization’s information health. Where the failures cluster, the dysfunction was already clustered. The blockages predate the dye.
This reframing matters because it changes what the spending bought. Most executives are currently trying to decide whether their AI investment “worked,” using success criteria borrowed from vendor decks. By those criteria, much of it didn’t. But the same investment, read as instrumentation, has produced something organizations have never had before: an honest, involuntary audit of where their information cannot be trusted. Consultants charge seven figures for worse maps than the one your failed pilots have already drawn — if anyone bothers to read it.
We have been here before
If this pattern feels new, it isn’t. It is a rerun, and the original aired in 1997.
That year, Paul Strassmann — former CIO of Xerox and of the U.S. Department of Defense, and a man who had spent decades actually measuring what computers do to businesses — published The Squandered Computer. His core finding, built on years of analyzing corporate financials against technology budgets, was simple and devastating: there was no correlation between how much companies spent on information technology and how well those companies performed. None. Big spenders did not outperform small spenders. The decades-long assumption that computerization translated into competitive advantage was, on the evidence, unsupported.
The industry’s response is instructive. It was not to engage with the measurement. It was to wait Strassmann out — to keep publishing case studies, keep running transformation programs, and keep treating IT spending as self-evidently strategic. It took roughly a decade, and an enormous amount of squandered capital, for his position to migrate from heresy to common knowledge: technology spending only pays when the organization’s information practices and management discipline allow it to pay. The computer never was the variable. The organization was.
Strassmann’s deeper point — the one that matters for this moment — was about where the returns actually came from. When he found companies that did profit from technology, the technology wasn’t the differentiator. Management was. The same dollar of IT spending produced wildly different outcomes depending on the organizational context it landed in. Spending was easy to measure and easy to approve; the organizational conditions that determined whether spending paid were hard to measure and therefore ignored.
Replace “IT” with “AI” and you have a precise description of 2026. We are measuring adoption because adoption is easy to measure. We are not measuring the conditions that determine whether adoption pays, because those conditions — information quality, decision rights, validation burden — are hard to count and embarrassing to examine.
The last cycle took a decade of denial before the measurement caught up with the spending. The only question this time is whether you personally need to repeat that decade, or whether you’re willing to skip to the part where someone counts.
What the instrument is for
Treating AI as a contrast agent is not a metaphor to admire. It is a practice, and it has three components.
First: read the failure map. Inventory your AI initiatives — not by status (everyone lies about status) but by failure mode. Where did output require heavy correction? Where did the pilot produce nothing anyone used? Where did people quietly revert to the old way? Then ask, for each: what would have to be true about our information for this to have worked? The answers are a list of specific, located, previously invisible breakages — wrong data treated as right, undocumented knowledge living in individual heads, official processes that nobody actually follows. This is not a list of AI problems. It is a list of organizational problems that AI was the first instrument sensitive enough to detect.
Second: count the validation labor. The largest unexamined number in corporate AI is the human cost of checking the machine. Every hour spent reviewing, correcting, fact-checking, and re-doing AI output is a real cost, paid in your most expensive people’s time, and it is booked nowhere. In many deployments, validation labor silently consumes most or all of the productivity the tool was purchased to create — the work didn’t disappear, it changed shape and moved off the books. Until that number is on a ledger next to the subscription cost, every ROI claim in every internal deck is fiction. (This is exactly what the Validation Ledger that accompanies this essay is for: one page, the full cost picture, including the hours nobody books.)
Third: fix what the dye revealed — not the dye. When the contrast scan shows a blockage, no competent physician prescribes better dye. Yet that is precisely what organizations do when a pilot fails: switch vendors, upgrade models, hire prompt engineers. The high-leverage move is the unglamorous one: repair the information condition the failure exposed. Clean the source the model was hallucinating from. Write down the knowledge that lived only in Dave. Reconcile the official process with the real one. These repairs are not “AI work,” which is why no AI budget funds them — and they are the only work that determines whether the next deployment pays.
The honest position
None of this is an argument against AI. The technology’s capabilities are real, and organizations that learn to deploy it on sound information foundations will get returns that the case studies promise and almost nobody achieves. The argument is against unaccounted AI — deployment without instrumentation, spending without counting, transformation claims without verification.
The executives who navigate this well will not be the ones who adopted fastest or spent most. The last cycle settled that question definitively, in public, with a decade of squandered capital as the tuition. They will be the ones who treated every failure as a reading, every unverifiable win as a flag, and every renewal decision as a measurement problem rather than a faith problem.
Your AI isn’t failing. It’s reporting.
The only question is whether anyone in your organization is reading the instruments.
The Validation Ledger — a one-page worksheet for counting what your AI deployment actually costs, including the hours nobody books — is available to dispatch subscribers.