Lists of AI opportunities are cheap and everywhere; lists of AI non-opportunities are rare, which is strange, because the second list saves more money. What follows is not a catalog of AI's technical limits — those move constantly — but four recurring situations in companies where the demonstration succeeds, the enthusiasm is genuine, and the value still never arrives. They are worth naming before the pilot budget is spent, not after.
Automation of any kind earns on repetition, and the same arithmetic governs AI: a task performed a few times a month, however elegantly automated, returns nothing worth the integration, the maintenance and the new failure modes it introduces. The tell is the pilot that «works» — a small-volume task pilots beautifully and scales into nothing. Volume is the first filter in use case selection, and the one most often skipped, because low-volume tasks are precisely the ones that look charming in a demonstration.
AI supplies intelligence-shaped effort, so it pays only where intelligence-shaped effort is the constraint. Many processes are constrained by something else entirely: waiting for a counterparty's signature, waiting for a decision owner's attention, waiting for market feedback, waiting for a physical step. Inserting a model into a queue does not move the queue; it makes the queue's paperwork prettier. The diagnosis runs through the whole process — where the time and the cost actually accumulate — and the honest outcome of that diagnosis is often an uncomfortable finding that the constraint is a person, a contract, or the company's own decision speed. Where AI moves the P&L is where the constraint is genuinely intelligent work; everywhere else it is decoration.
Some decisions are made rarely enough and matter enough that removing the human does not save anything worth risking: pricing a flagship product, a public communication, a commitment with legal weight. In these decisions AI can draft, summarize and prepare — a real contribution — but the last mile of judgment is where the liability lives, and companies that delegate the last mile to save minutes end up buying back the minutes at a different price. The line is not about the model's accuracy; it is about the asymmetry of consequences, which no amount of accuracy fixes.
Automating a broken process accelerates the brokenness. AI is a powerful amplifier of whatever workflow it is pointed at, and when the workflow itself is incoherent — redundant approvals, contradictory rules, steps that exist only by habit — the model executes the incoherence faster and with more confidence than the tired humans did. The fix sequence is inverted from the intuitive one: repair first, automate second, and the repair itself often shrinks the task until the cost of the process no longer justifies any automation at all — which is a saving the automation budget never gets credit for.
A use case that fails one of these four tests is not a weak opportunity. It is a well-demonstrated way to spend integration effort on nothing.
Running candidate use cases against the four filters — volume, real bottleneck, error asymmetry, process health — is the opening stage of opportunity assessment and standard work in the AI practice: the cheapest point in the cycle to say no, and therefore the most profitable one.
We use cookies to analyse traffic and improve the site. Analytics is enabled only with your consent. Cookie Policy