A look at the industry's favorite new cost-cutting idea, and the cheaper one it keeps walking past
There is a particular kind of corporate memo circulating right now, and if you have not received one yet, you will soon. It arrives from finance, sometimes from a newly appointed "AI efficiency" lead, and it says, in effect: we spent an alarming amount of money on artificial intelligence last quarter, and we are not entirely sure what we got for it. Attached to the memo, increasingly, is a piece of advice that has become so common across boardrooms, LinkedIn posts, and consulting decks that it has started to sound less like a strategy and more like a mantra. The advice is this: stop using your most expensive model for everything. Treat your frontier AI system the way you would treat a senior partner at a law firm or a $400-an-hour management consultant reserve it for the hard, high-stakes thinking, and hand the routine work to someone, or something, cheaper.
It is a tidy idea. It has the comforting shape of wisdom, the kind that sounds obvious in hindsight, which is usually a sign that it is either genuinely important or dangerously simple. This piece is an attempt to figure out which.
Why this idea is having a moment
To understand why "treat the model like an expensive consultant" has caught on so quickly, it helps to remember what came immediately before it. For roughly two years, the dominant posture inside companies experimenting with AI was the opposite of restraint. Employees were told to use the tools constantly, on everything, as a way of building fluency and surfacing use cases. Usage itself became a proxy for progress. Some companies went as far as making AI adoption a line item in performance reviews. The logic was defensible at the time: nobody knew yet which tasks AI would actually help with, so the fastest way to find out was to point it at everything and see what stuck.
That phase produced a lot of genuinely useful discovery. It also produced staggering bills, because pointing your most capable and most expensive model at every task, including the trivial ones, is a bit like hiring a surgeon to apply band-aids. Nothing goes wrong, exactly. It is just an absurd use of a scarce and costly resource, repeated thousands of times a day across an organization, until finance notices.
So the pendulum swung, as it always does, and it swung toward discipline. That part is healthy. Any technology that moves from experimental toy to core infrastructure eventually has to survive contact with a budget review, and AI is no exception. Every prior wave of enterprise software went through the same cycle: an initial phase of enthusiastic, undisciplined adoption, followed by a correction in which someone with a spreadsheet asks what, exactly, all of this is buying. Cloud computing went through it. Enterprise software licensing went through it decades before that. There is nothing unusual about AI spend now attracting the same scrutiny; if anything, it would be strange if it hadn't.
What is a little more unusual is the speed at which a single, tidy explanation for the overspending has crystallized across so many companies at once, arriving with almost identical language, almost identical metaphors, and almost identical confidence, as though everyone arrived at the same insight independently within the space of a few months. That kind of synchronized clarity is worth being a little suspicious of. Genuinely difficult problems tend to produce a scatter of competing explanations before anyone converges on the right one. When an explanation shows up everywhere at once, fully formed, it is worth asking whether it was discovered or distributed.
None of this is to say the underlying concern is fake. The problem is not that companies are now asking harder questions about what they are paying for. The problem is with the specific answer that has become fashionable, and with how confidently it is being sold as a complete one.
The idea itself
The mechanics are simple enough to explain in a sentence: use the most powerful, most expensive model to plan, strategize, and break a problem into pieces, then route the actual execution of those pieces to smaller, cheaper models that are perfectly capable of doing narrow, well-defined work. A flagship model might spend a few seconds deciding that a customer inquiry needs to be classified, summarized, and routed to the right department, and then hand each of those three sub-tasks to a lightweight model that does the job for a fraction of the cost per token. Multiply this across a large enough operation and the savings, in principle, are real. Nobody disputes that a data-extraction task or a routine email draft does not require the same computational horsepower as, say, drafting a novel legal argument or architecting a new product from scratch.
The analogy to professional services makes intuitive sense on its surface. A law firm does not bill its most senior partner's hourly rate for photocopying and scheduling; it has a hierarchy, with the expensive expertise reserved for the moments that actually require it. Applying that same hierarchy to AI models a "senior" model for judgment calls, a "junior" model for grunt work sounds like nothing more than importing a well-tested organizational principle into a new domain.
Here is where it is worth slowing down, because the analogy is doing more work than it can actually support.
In a law firm, the decision about which task goes to the partner and which goes to the paralegal is made by people who already understand the case, the client, and the stakes involved. That judgment is not itself billed as a separate, premium service. It is simply how organizations with any functioning structure operate; a competent office manager, a team lead, or frankly any employee who has been at the company for more than a few weeks already knows, more or less instinctively, which tasks are routine and which ones deserve real thought. That knowledge is not scarce. It is not expensive. It is, in most functioning teams, close to free, because it is a byproduct of simply understanding your own business.
The version of this idea currently being marketed asks companies to do something stranger: pay a premium AI system, on every task, to rediscover from scratch a categorization that the people running the business could already produce for nothing. The expensive model is not being used to solve the hard problem. It is being used to figure out which problems are hard a determination that, in almost every real organization, someone on staff could make in about four seconds, without spinning up an API call. Dressing that determination up as "strategic planning" performed by a senior AI consultant does not make it more valuable. It just adds a billing line to a decision that used to be free.
This is not a small technicality. It is the entire hinge on which the "expensive consultant" framing turns. The framing works only if you accept that triage deciding what deserves deep thought and what does not is itself a task requiring frontier-level intelligence. In practice, triage is one of the most human-native skills there is. People do it constantly, cheaply, and well, because they have context that no model, however expensive, is handed for free: they know the client, they know what happened last quarter, they know which shortcuts are safe and which ones will blow up in six months. A model, however capable, is starting from zero on all of that, every single time, unless someone pays to feed it that context at which point you are paying twice: once for the context-gathering, and again for the "expensive" judgment that context was supposed to inform.
A quick thought experiment
Picture a mid-sized company that handles customer support, contract review, and internal reporting, and imagine it adopts the full "expensive consultant" model faithfully. Every incoming task a support ticket, a contract clause, a weekly summary first passes through the flagship model, which decides how complex the task is and routes it accordingly. On paper, this looks efficient. In practice, ask a simple question: who decided that these three categories of work were the ones worth analyzing in the first place, and roughly how each one should be weighted?
The answer, almost certainly, is a person. Someone in operations sat down, probably in an afternoon, and worked out that contract review usually needs careful attention, that most support tickets are repetitive and low-stakes, and that internal reports fall somewhere in between depending on the audience. That person did not need a frontier model to reach these conclusions. They needed to have worked at the company for a few months and to have paid attention.
Once that basic map exists, the expensive model's "planning" role shrinks dramatically. It is no longer discovering which tasks are hard; it is executing a categorization a human already made, task by task, at inference-time prices, forever. The company is not buying strategic judgment on an ongoing basis. It bought it once, for free, from an employee doing their job, and is now renting a much costlier version of the same judgment on a per-token basis, indefinitely, because the marketing around the tool made that ongoing rental sound like the sophisticated choice. This is not to say routing itself is useless sending simple tasks to cheaper models is a sound engineering practice once the categories are established. The issue is specifically with the idea that the categorization itself needs to be an expensive, ongoing AI function rather than a one-time act of institutional knowledge that any reasonably attentive team already possesses.
The part nobody quite says out loud
It is worth asking, plainly, who benefits from a framing that treats the categorization step as premium work. The honest answer is: whoever is positioned to sell you the categorization step. An entire small industry has emerged over the past year around exactly this idea platforms that promise to monitor your AI spend, route your workloads intelligently, and tell you, for a fee, which of your tasks deserve the expensive model and which don't. Several have raised meaningful venture funding on the strength of this pitch alone.
There is nothing inherently wrong with a company building a product around cost optimization; plenty of legitimate infrastructure gets built this way. But it is worth noticing the shape of the argument being made. The pitch is not simply "AI is expensive, here is a tool that helps you spend less." The pitch is "the thing you are already good at knowing your own workflows is actually a sophisticated strategic function that requires our platform, or a frontier model acting as your consultant, to perform correctly." That is a much more convenient claim if your business depends on people believing it. It reframes a basic managerial competency as a gap that only a paid product can fill.
None of this is a conspiracy. It is simply the ordinary way that markets work: when a genuine cost pressure appears, someone will show up with a product that requires you to believe the pressure is more complicated to solve than it actually is. The tell, in this case, is how thin the actual evidence for the framing tends to be. Search for hard, independently verified numbers behind the "treat frontier models as expensive consultants" idea controlled before-and-after comparisons, published cost breakdowns, anything resembling a rigorous case study and what you mostly find instead is a chain of confident quotes from people whose companies are built to sell you the solution. That is not proof the idea is wrong. It is a reason to hold it a little more loosely than the confidence of its advocates would suggest.
The mundane truth hiding underneath
Strip away the consulting metaphor and what remains is a genuinely useful, genuinely unglamorous engineering practice: match the tool to the task. Nobody should dispute that a document-classification task does not need the same model that can architect a distributed system, any more than a filing job needs a surgeon. That part of the advice is correct, and companies that are indiscriminately routing every request through their most expensive model are, in fact, leaving money on the table. But the interesting, useful version of that advice is boring in a way that doesn't sell well. It sounds like this: understand what your teams actually do, task by task; measure where quality genuinely depends on a more capable model and where it doesn't; build simple, unglamorous rules for routing work accordingly; and revisit those rules periodically, because model capability and pricing are both moving targets right now. None of that requires believing that AI models occupy a professional hierarchy analogous to a law firm. It requires the same thing good operations has always required: paying attention to your own workflows. The "expensive consultant" framing survives not because it is the most accurate description of what's happening, but because it is the most flattering one. It lets a fairly ordinary act of resource allocation borrow the prestige of professional services the language of partners and associates, of strategic counsel and billable hours and in doing so, makes a spreadsheet exercise sound like sophisticated judgment. Companies like being told that the fix for their overspending is to hire a smarter kind of expensive thing, rather than to look, carefully and without much drama, at what their own people already know.
Where this actually goes
There is a reasonably confident prediction to make here, and it does not require a model of any size to make it: the price of frontier-level AI intelligence is falling, and falling fast. Every major lab has spent the last year introducing cheaper tiers, faster small models, and aggressive promotional pricing, and there is no sign of that trend reversing. Within a relatively short window, the price gap between a "premium consultant" model and a "junior" model is likely to compress significantly, and the entire architecture of the expensive-consultant metaphor the idea that you need an elaborate hierarchy to avoid paying senior rates for junior work starts to look like a solution to a problem that is quietly shrinking on its own.
What will not shrink, and what will remain the actual differentiator between companies that get real value from AI and companies that don't, is the unglamorous work of understanding your own operations well enough to know what needs real thought and what doesn't. That capability was never for sale, which may be exactly why it gets talked about so much less than the products built to replace it.
It is worth sitting with how strange this is, once it is stated plainly. The industry has spent the last year discovering, with great fanfare, that not every task needs the most powerful available tool an observation that predates computing itself, and that shows up in some form in nearly every operations textbook ever written. The genuine achievement, if there is one, is not the insight but the packaging: taking a piece of common sense old enough to have been true of typewriters and photocopiers, and successfully repositioning it as a strategic breakthrough specific to this moment in AI. That is a marketing accomplishment, not an intellectual one, and it is worth being able to tell the difference, especially when the bill arrives.
Companies that actually save money over the next few years will likely be the ones that quietly did the boring version of this work auditing what their teams actually need, task by task, without paying anyone extra to tell them what they could have found out by asking rather than the ones that bought the more expensive story about why that work required a consultant, human or artificial, to begin with.
The next fashionable idea in AI cost management will probably arrive wearing a different metaphor orchestras instead of law firms, maybe, or supply chains but it will very likely be selling the same basic insight, repackaged: don't use your most expensive tool for your simplest job. That insight was true before anyone thought to charge a subscription for it, and it will still be true after this particular metaphor has quietly retired to wherever expired consulting frameworks go.
___________________________________________________________________________________
Articles published in the Reviews section provide analytical, interpretive, and occasionally forward-looking perspectives on scientific, technological, and policy developments. While grounded in available evidence and referenced sources where appropriate, they may include reasoned critique, synthesis, or informed judgment. They should not be interpreted as representing scientific consensus or definitive conclusions.
___________________________________________________________________________________
