There is a pattern playing out inside hundreds of companies right now, and it goes like this: a senior executive attends a conference, watches a demo, and returns to headquarters convinced that generative AI must be woven into everything the company does — immediately. Engineers and data professionals are then handed a mandate rather than a problem. The results are predictable, and they are not good.

The case of Malcolm — an AI engineer at a data analysis firm whose identity we are protecting at his request — is a clean illustration. When his company's leadership decided to use a large language model to segment a customer database into behavioral personas, Malcolm pushed back. The right tool for that job was a traditional machine learning classifier: deterministic, auditable, far cheaper to run, and purpose-built for exactly that kind of pattern recognition. Executives heard him out and proceeded with generative AI anyway. The output was less consistent, the costs were substantially higher, and the business problem was not solved as well as it would have been with the tool the engineer actually recommended.

This is not an isolated anecdote. It is a systemic failure of technical governance that is costing corporations real money while producing mediocre or actively harmful outcomes. The confusion stems from a category error that has been aggressively marketed into existence: the conflation of "AI" as a monolithic technology, when in reality it is a broad family of tools with radically different cost profiles, reliability characteristics, and appropriate use cases. Generative models — the kind that produce text, images, and synthetic data — are genuinely powerful for certain tasks. They are also expensive to run, prone to hallucination, and architecturally inappropriate for workloads that require consistency and repeatability. Deploying them as a default is roughly equivalent to hiring a novelist to proofread a spreadsheet.

The financial exposure is real and accelerating. Enterprise AI compute costs have surged as companies rush to stand up infrastructure before they have defined what problems that infrastructure is meant to solve. Internal budget reviews at multiple large organizations have flagged AI spending as a category that is growing faster than its demonstrable return — a dynamic that rarely ends well. The irony is that the underlying tools which generative AI is displacing — classical machine learning, rule-based automation, structured data pipelines — are mature, well-understood, and in many operational contexts still superior.

What makes this moment particularly costly is the human dimension. Workers who interact with these deployments daily report genuine confusion about what the tools are supposed to do, what outputs can be trusted, and what happens when they are wrong. When there is no clear organizational answer to those questions — and often there is not, because the strategy was built around a vendor pitch rather than an operational need — staff either over-rely on outputs they cannot verify or quietly route around the system entirely. Neither behavior is what the investment was sold to deliver.

The oversight gap runs upward as well as downward. Many executive teams lack the technical background to evaluate whether a proposed AI deployment is appropriate for a given task, and the consulting and vendor ecosystem has a powerful financial incentive not to tell them when it is not. A well-scoped machine learning project with modest licensing costs does not generate the same advisory fees or platform contracts as a full generative AI transformation program. The incentive structure actively works against the right answer.

What a functional AI strategy actually looks like is less glamorous than the conference circuit suggests. It starts with a problem inventory — what specific operational failures or inefficiencies are we trying to fix — and works backward to the technology, rather than forward from the technology to whatever problem can be retrofitted onto it. It requires that engineers and domain experts have genuine veto power over deployment decisions, not merely advisory roles that can be overruled when enthusiasm runs high. And it demands honest accounting: not just what a deployment costs to stand up, but what it costs to run at scale, what it costs when it is wrong, and what the organization is actually giving up by not using a simpler tool.

Malcolm's firm is still running its generative AI persona model. The results remain inconsistent. The costs remain elevated. And Malcolm, like a growing number of practitioners watching this from the inside, is waiting for the moment when the bills get large enough that someone in a boardroom finally asks why nobody warned them.