Spending more was never the AI strategy

For a while the answer to almost every AI question was to spend more. Buy licenses for everyone. Point the largest available model at every task. Approve the budget and work out the value later.

That stretch appears to be ending. It was reported this summer that the spending spree has passed its peak, and that companies are starting to mix cheaper models in alongside the expensive ones instead of reaching for the best available model on every job. An executive at one of the AI coding companies compared running a top model for routine work to driving a Lamborghini to pick up milk. The comparison landed because everyone recognized their own invoice in it.

Here is the part worth betting on. When spending slows, the advantage does not disappear. It moves. It moves to whoever knows which model to point at which job, and to whoever can tell when the cheap one quietly got something wrong.

Five practices follow from that. None of them require a bigger budget, and most of them get cheaper as a team gets better at them.

Match the model to the task

The common failure is running one model for everything, chosen because it scored best on a benchmark somebody posted.

Most of what a development team asks an AI to do is not hard. Summarizing a long document. Renaming things consistently across a module. Drafting boilerplate. Writing a commit message. A cheap, fast model does these as well as an expensive one, and in the finished result you will not be able to tell which one did the work.

Then there is the other kind of task. Reasoning across a codebase nobody on the team wrote. Planning a migration where the ordering constraints matter and getting them wrong means a weekend. Debugging something where the obvious answer is the wrong answer. That is where an expensive model earns its price, and it earns it in one pass instead of five.

The practice is to classify your recurring work once, out loud, as a team. Two buckets is enough to start with. The teams that skip this step are not making a technical choice, they are paying a premium to avoid a twenty minute conversation.

Do not marry a vendor

The second failure is standardizing on a single provider because procurement prefers one invoice, and then discovering that the invoice has no ceiling.

The teams saving real money run more than one provider side by side and move work between them as prices and capabilities change. In this market that is not disloyalty, it is basic hygiene. Capability leapfrogs every few months, and whoever led when you signed is usually not the one leading by renewal.

What this asks of your engineering is small but it has to be deliberate: keep the seam. The model is a configuration value, not an assumption baked into forty call sites. Teams that treat the provider as an interface can act on a price change in an afternoon. Teams that hardcoded it will read the same news and do nothing, because switching turns into a project nobody has time to sponsor.

Report cost per finished task, not tokens burned

Most AI spend reporting measures activity. Tokens consumed, requests made, seats active. Those numbers go up when the tool gets used more and they go up when the tool gets used badly, which makes them useless for a decision.

The number worth putting on a slide is what one finished piece of work cost. Not generated: finished, reviewed, and merged.

That number behaves in ways the price sheet will not predict. Sometimes the expensive model is the cheaper choice, because it gets to a correct answer in one attempt while the cheap one takes four tries and a rewrite. Sometimes the cheap model runs three times and still comes in under. You cannot reason your way to this from the price per million tokens. You have to measure the finished thing, on your own work, in your own codebase.

Spend more and assume better is precisely the habit that just peaked. Cost per finished task is the habit that replaces it.

Keep a person at the merge gate

Let the AI work overnight if you want. Let it open the pull request. Someone still reads what landed before it merges.

The gate is not there because these tools are unreliable in general. They are quite good. It is there because they are confidently wrong in a narrow and unpredictable band, and because the cost of one bad merge into a system that already carries real customers is far higher than the cost of the review that would have caught it.

Moving to cheaper models makes this more important rather than less. A weaker model’s mistakes are not louder, they are quieter. It will produce something that reads correctly, passes a shallow look, and is subtly wrong about a case that shows up in production three weeks later. If you are going to capture the savings, you have to fund the verification. That is the actual trade, and teams that take the savings without funding the gate have not saved anything yet. They have moved a cost from a line item to an incident.

Judging the output is the scarce skill

Anyone can generate. Generation stopped being the constraint some time ago.

The rare skill is reading a result and knowing where it went wrong. Not whether it looks plausible, whether it is right. That skill is what decides whether cheap models are usable in your organization at all. A team that can reliably tell good output from merely plausible output can run cheap models nearly everywhere and check the seams. A team that cannot has to keep buying the expensive model as insurance, and will never know whether the insurance was necessary, because it has no way to evaluate what it is getting.

This is also why the leverage from these tools tracks engineering fundamentals so closely. Sharp thinking gets multiplied. Sloppy thinking gets multiplied too, and it arrives faster and in greater volume than it used to.

The part that does not change

More tools were never the advantage. The people who use them well are, and that was true when budgets were expanding and it stays true now that they are not.

If your AI spend is under scrutiny this quarter, do not start with the contract. Pick one recurring task, run it through the cheapest model you have access to, and compare the finished result honestly. If you cannot tell the difference, you have just learned exactly what the last year of default spending was buying you.