A sizeable ride-hailing platform was operating smoothly for two and a half years in the US. Every ride ended in a receipt. Every receipt was computed to the cent, on time, without complaint. Tens of thousands of drivers were paid this way, week after week, by software that did exactly what it was told.
The flaw was one definition. The company's own terms said its commission came out of the fare after taxes and fees. The code took its cut before them. No engineer chose this. No test covered it. The definition lived in a legal document. The calculation lived in code. Nothing in the world connected the two.
The platform was Uber. In May 2017, it admitted it had underpaid its New York City drivers since late 2014. The refunds came to tens of millions of dollars, about nine hundred dollars per driver. A newspaper found the gap. The system never did. In thirty months, no error was thrown.
This post is about that silence. What it has cost companies for decades. Why it survives every honest attempt to fix it. And why machines asking questions of our data have finally made it impossible to ignore.
The same failure, five ways
The same failure repeats across industries and decades, and it follows rules.
In 1999, NASA lost the Mars Climate Orbiter. Ground software on one team produced thruster data in pound-seconds. The navigation software on another team expected newton-seconds. The $327.6 million spacecraft flew too close to the planet and broke apart. Both programs ran without fault. Only the meaning of one number differed between them.
In 2020, a regulator fined Citibank 400 million dollars for long-standing failures in data governance and reporting. Behind the legal language sits a plain fact. The bank could not reliably say what its own numbers meant across its systems.
Read these beside Uber and a pattern comes into focus. Five rules hold in every case.
First, the machinery always worked. The code ran, the thrusters fired, the invoices billed. The failure lived in the meaning of the numbers. Monitoring watches the machinery, so monitoring saw nothing.
Second, the truth existed the whole time. The interface specification required newtons. The terms of service defined the fare. The knowledge sat apart from the place where the data flowed, and nothing enforced the connection.
Third, discovery always came from outside. A spacecraft went missing. A newspaper called. A regulator audited. In no case did the system notice its own error. Nothing in the stack held that job.
Fourth, silence multiplied the cost. Uber ran wrong for thirty months, Citibank for years. A meaning error is cheap on the day it is born. The bill grows in the quiet between its birth and its discovery.
Fifth, people were the safety net. For decades, someone eventually frowned at a number, asked around, and traced it back. The disasters cluster wherever people were absent, outnumbered, or too slow.
What the quiet years cost
The incidents are the visible peaks. The everyday cost is larger, and it has never had a line on any budget.
The McKinsey Global Institute found that knowledge workers spend nearly a fifth of the workweek searching for internal information and chasing the colleagues who know. One working day in five goes to finding out where the truth lives. Companies that tried to price the damage land in the same range. Gartner's surveyed customers put poor data quality at an average of 12.9 million dollars a year, by their own estimate. Thomas Redman, writing in MIT Sloan Management Review, put the cost of bad data at 15 to 25 percent of revenue for most companies. And one survey of a thousand American workers found that 42 percent of what an employee knows is unique to that one person. When they leave, that part of the company's memory leaves with them.
We hold these numbers loosely. Some are self-reported, and one comes from a vendor with something to sell. But they all point the same way.
The deeper cost is trust, and it never reaches a spreadsheet. A team quietly stops believing its dashboard. An analyst takes the blame for a pipeline's mistake. Every metric review opens with the same twenty minutes: wait, which revenue is this? An organization loses faith in its own data the way a person loses eyesight. So gradually that nobody can name the day it happened.
Then the machines started asking
For decades, the human safety net kept this survivable. Slow, expensive, invisible, but survivable. Then we handed the questions to machines, and the net was gone.
An agent querying company data works at machine speed and machine volume, with no access to the hallway. It cannot ask the colleague who remembers the meeting where net revenue got its definition. It reads the schema, finds a believable path, and answers with confidence. The old failure compounded.
In 2022, Unity fed bad data from a large customer into the machine-learning model that priced its ads. The pipelines ran clean while the meaning inside them was poisoned. The fault cost the company an estimated 110 million dollars in revenue, and investors erased billions in market value within a day. The silence lasted months instead of years. The bill arrived in a quarter.
In 2024, a tribunal held Air Canada liable after its chatbot invented a refund policy for a grieving passenger. The real policy sat on the same website. The award came to about eight hundred dollars. The precedent reaches every company that puts a machine in front of a customer. When a machine states a wrong meaning, the company owns the words.
My co-founder Charu watched this arrive from the inside. Her first company built evaluation systems for AI products, and every customer taught her the same lesson. The AI layer was easy to build. Its performance was set almost entirely by the context and the checks around it.
The industry's own numbers now say the same thing. Gartner found that 63 percent of organizations lack the data practices AI needs, or cannot say whether they have them. Gartner also predicts that through 2026, companies will abandon 60 percent of AI projects that lack AI-ready data. S&P Global measured the share of companies scrapping most of their AI initiatives jumping from 17 percent to 42 percent in a single year.
A smarter model does not escape this. Your definition of net revenue is a decision your company made, recorded in a document or carried in one person's head. No amount of intelligence can discover a decision it was never told. A perfect reasoner holding the wrong definition produces a perfect wrong answer, faster.
Why the forms never worked
Companies have tried to write the meaning down. The attempts fail the same way every time, and I have lived the failure.
I spent years as a data engineer inside a large data organization. I watched weeks disappear into tables nobody could explain, tables whose meaning lived in people who had changed teams or left. I worked on data that only looked right, and paid for it in months. When the company's first AI project arrived, it needed exactly that missing meaning, and the effort collapsed into a form. Please fill up this form and tell us what these tables and these APIs mean. Sent to every team. The company was trying to serialize its own memory, by hand, after the fact.
The form fails for structural reasons. A filled form is a photograph, accurate for a moment and decaying from the day it is published. A stale page throws no error. No process halts when the wiki is wrong. Writing things down taxes today's deadline to help an unknown person later, so the deadline wins every time. And when the description drifts away from the data, nothing notices, because noticing is nobody's job. The form obeys every rule of the pattern above. It just fails more politely.
Software solved a problem of exactly this shape once before. Nobody keeps types correct by discipline. The compiler does. Nobody remembers to run the tests. The build does. A rule survives only when a system enforces it. The meaning of data has never had that system.
The gap
Everything built for this problem so far treats meaning as advice. Catalogs, wikis, and governance programs hold real knowledge, written by people who care, and they help real teams. But the description sits beside the data. Consulting it is optional. Enforcing it is nobody's job. And advice that nothing enforces drifts, silently, which is the pattern all over again.
We believe meaning must be an engineered property of the path between a question and its answer. Checked in the path where the data flows. Enforced by the system itself. Held against the live data, so that staleness surfaces as a failure while it is still cheap. That is the standard Rune exists to meet. Products will come and go against this problem. The standard will not move.
The number that is wrong right now
Right now, somewhere in your company, a number is wrong. It sits in a dashboard, a forecast, a board deck. It was computed correctly, from real data, by a system that raised no error. Everyone who has seen it believes it. And nothing in your stack holds the job of noticing.
That was always the problem. The machines have simply made it impossible to look away.
Agents ask. Rune proves.