The licences were purchased on a Tuesday. Enterprise seats for everyone, procurement waved through in a single meeting because nobody wanted to be the executive who argued against progress in front of the board. There was a launch. There is always a launch. An all-hands, a slide with a rocket on it, a phrase like force multiplier delivered with the particular confidence of a leader who has read three articles and one McKinsey summary. Every employee now had a model on their desktop, and the announcement went out to the market, and the annual report gained a paragraph about being an AI-first organisation. The word transformation appeared eleven times in a document that described the purchase of software and nothing else. What had actually happened was that a company had bought a very fast engine and bolted it to a machine nobody had ever bothered to inspect.
We were brought in eight months later, after the transformation had quietly stopped transforming and somebody senior had begun asking, in the careful language of people protecting a budget, why the numbers had not moved. Not the adoption numbers. Those were excellent. Ninety-something percent of employees were using the tool weekly, a figure repeated in every steering meeting like a prayer. The numbers that had not moved were the ones the tool was supposed to move. Cycle times. Error rates. The quality of the decisions the organisation actually made. On every measure that mattered, the company was exactly where it had been the year before, only now it was arriving there faster.
The Model Did Not Fail. It Obeyed.
Here is what nobody in the launch had considered, because considering it would have required admitting something unflattering about the year before the launch. A large language model does not bring judgment into an organisation. It brings fluency. It brings the ability to execute whatever process it is pointed at with tireless speed and unnerving confidence. If you point it at a good process, it makes a good process faster. If you point it at a broken one, and most processes inside most companies are quietly broken in ways everyone has learned to route around, it does not notice the break. It cannot. It was never given the thing that notices breaks, which is a human being who has been trained to look for them and given the standing to say so. So it takes the broken process and it runs it at a thousand times the volume, and it does this beautifully, and it does this all day.
The proposals that used to take a junior analyst two days now took nine minutes. This was reported as a triumph and it was, in the narrow sense that nine minutes is less than two days. What went unreported was that the two-day version had been passing through a human who, somewhere in those two days, occasionally caught the error, questioned the assumption, or noticed that the client had asked for the wrong thing and gently supplied the right one. That human was not in the process because the process was efficient. That human was the process. The judgment lived in the friction, in the slowness, in the pause between receiving a request and answering it. The organisation had spent a decade treating that pause as waste and had finally found a tool that eliminated it entirely. What they had eliminated was not the delay. It was the checking.
And so the errors did not decrease. They multiplied, and they improved their disguise. A mistake made by a tired analyst at least looks like a mistake, hedged and uncertain, flagged by the very hesitancy of the person who made it. A mistake made by a model arrives in clean prose, correctly formatted, footnoted, and utterly certain. It is a far more dangerous object, because it has been stripped of every signal a reviewer used to rely on to know where to look. The company had not automated its work. It had automated its confidence, and confidence is precisely the thing you least want to scale in an organisation that has not yet learned to be right.
What the leadership told itself about all this was, predictably, a story about the tool. The model was not powerful enough. The prompts were not sophisticated enough. They needed better training on the software, another vendor, a centre of excellence, a prompt library, a Chief AI Officer with a mandate and a war chest. Every proposed remedy pointed at the technology, because pointing at the technology implicated nobody who had built the organisation the technology was now faithfully reproducing. Nobody proposed the one intervention that would have worked, which was to go back and ask whether the process the model was accelerating had ever been any good, and whether the people now supervising the model had ever been taught to supervise anything at all.
AI Does Not Fix a Broken Organisation. It Industrialises It.
This is the reframe that the entire market is currently spending billions to avoid. A model dropped into a healthy organisation is a genuine multiplier. A model dropped into a dysfunctional one is a magnifying glass held over the dysfunction, and then a printing press attached to the magnifying glass. It does not repair the thing that was wrong. It manufactures the thing that was wrong, at industrial scale, with a speed and a polish and a plausibility that make the wrongness far harder to see and far more expensive to unwind. The bottleneck in a bad company was never throughput. It was judgment, and judgment is the one thing the licence did not come with.
Consider what actually determines whether the tool helps or harms. It is not the model. Every serious company is buying from roughly the same three or four frontier providers, which means the model is now a commodity and cannot be a source of advantage, however the vendor slides insist otherwise. The variable that decides the outcome is entirely internal. It is whether the people receiving the output know enough to reject it. Whether an analyst can look at a fluent, confident, wrong answer and feel the specific discomfort that says something here does not hold. That discomfort is not natural. It is trained, over years, through exposure and correction and the accumulated scar tissue of having been wrong and been caught. An organisation that spent the previous decade optimising that friction out of its junior ranks, in the name of efficiency, has no such discomfort left to deploy. It handed a loaded instrument to people it had carefully untrained to question anything.
The companies that are pulling ahead in this cycle are not the ones with the best model. They cannot be, because there is no best model to have. They are the ones who did an unglamorous thing first. They built the judgment architecture the tool would eventually plug into. They defined, before they scaled anything, what a good decision looked like in their context, who was accountable for catching a bad one, and how a human being earned the standing to override a confident machine. They treated AI not as a workforce but as an amplifier, and they understood the iron rule of amplifiers, which is that they make everything louder, including the mistakes. So they made sure the thing being amplified was worth amplifying before they turned the volume up. This is slow, and it is invisible, and it does not photograph well next to a slide with a rocket on it.
This is the work that SSUNDAR does before a single licence is deployed. We treat the request to roll out AI as a diagnostic prompt rather than an instruction, because the tool will faithfully reproduce whatever judgment already exists in the organisation, and our job is to find out whether any exists. We map where real decisions are actually made, where the judgment that catches errors currently lives, and where it has already been hollowed out by a decade of efficiency programmes that mistook the human check for a cost. Then we rebuild that architecture, the standards, the accountability, the trained discomfort, so that when the tool is finally scaled it is scaling something worth scaling. It is the least fashionable work in the market right now. It is also the only work that determines whether the licence becomes a multiplier or a very expensive way to be wrong more quickly.
A model trained on a broken process does not fix the process. It automates it, at scale, faster, with more confidence and far less traceability. You did not buy a solution. You bought a louder version of the problem.
You Cannot Outsource a Judgment You Never Built.
The uncomfortable arithmetic is that the tool exposes the organisation with perfect fidelity. It is the most honest diagnostic a company will ever run, precisely because it has no opinion of its own and simply amplifies what is there. A company with judgment gets a force multiplier. A company without it gets its own hollowness returned to it at volume, printed on clean letterhead, ninety-something percent adopted, and entirely unable to explain why nothing improved. The adoption metric measures how many people are using the tool. It has never once measured whether the tool should be trusted to do what those people have stopped doing themselves.
So the executives who bought AI for everyone and trained no one are not wrong that something has changed. Something has. They are simply wrong about what. They believed they were installing capability. What they installed was a mirror, and the reflection is arriving faster every quarter, more confident every quarter, and no closer to being right. The organisations that will win the next decade are not the ones that bought the tool earliest. They are the ones who did the patient, unglamorous work of building the judgment the tool plugs into, before they ever scaled the tool. Everyone else bought a faster horse and forgot to check whether it was running toward the cliff.
You did not give your people a capability. You gave them a machine that does, at scale and with total confidence, exactly what your organisation already did badly, and then you called the acceleration progress.