Show People What a Prompt Costs and They Write Better Prompts

Researchers tracked more than 1.8 million prompts from more than 61,000 users to find out what happens when you put a price tag next to the generate button. The answer had almost nothing to do with the bill. With cost visible, people spent 5% more time writing each prompt and 8% less time generating variants of the same image. They also reported being more satisfied with the results.

That finding comes from “How Costs Influence Preferences for Control in Generative Artificial Intelligence (GenAI): Human-Guided vs. GenAI-Based Delegated Search,” published in Information Systems Research in April 2026 by researchers listed at Indiana University and Pennsylvania State University. Thomas McKinlay summarized it for Science Says on 1 September 2026. Two of the paper’s authors render on that page as Wang and Lee; the third is truncated in the source, so I am not naming it here.

The number that matters is the second one. An 8% drop in variant generation is a behavioural change in how the work gets done, and it showed up because a number appeared on a screen.

The Meter Is Being Sold to the Wrong Buyer

Every AI cost control I see in an enterprise procurement conversation is pitched at finance. Caps, budgets, variance alerts, approval gates on overages. The framing is containment: the tool exists because someone might spend too much, and the control exists to stop them. Engineering treats the meter as an imposition to negotiate down, and finance treats it as a leash to hold. Both are arguing about the same axis, which is the size of the bill.

We have written this way ourselves. Governing spend by variance rather than caps is a spend-governance argument. So is the billing unit as the control surface. So is an agent approving its own budget overage. Those arguments hold. What the study adds is a second axis nobody was measuring, and on that axis the meter is not a constraint at all.

Cost Awareness Changed the Search Strategy

The mechanism proposed in the research is worth reading slowly. Cost awareness triggers a desire for control and certainty. That desire shifts people away from delegated search, where you hand the problem to the model and let it explore, toward human-guided search, where you decide more of the direction yourself before the model runs.

Delegated search is cheap to start and expensive to finish. You write a vague prompt, look at what comes back, adjust slightly, generate again. Each cycle feels like progress because something new appears on screen. The thinking happens between generations rather than before them, and the model absorbs the cost of the operator’s unresolved intent.

Human-guided search front-loads the thinking. You spend longer specifying what you want because you have a reason to get it right the first time. The 5% and the 8% are the two halves of that trade: more time in the prompt, less time in the retry loop. The satisfaction result is the part that closes the argument. If the meter were only a leash, constrained users would report being less happy with their output. They reported the opposite.

One boundary matters here. Those two percentages describe a specific behaviour, generating variants of the same image, and I am not going to pretend they transfer cleanly to coding agents or to text workflows. What transfers is the mechanism. The effect is reported as stronger among heavy AI users, which is the population most likely to have built a fast, unreflective generate-adjust-generate habit in the first place.

Feedback Is Cheaper Than Governance

Read the result as an engineering finding rather than a finance one and the recommendation inverts.

A spend cap tells you nothing until you hit it, and what it tells you then is “stop.” It carries no information about which of your prompts were wasteful. It arrives after the behaviour, not during it. Approval gates are worse on this axis, because the person approving is usually further from the work than the person spending, and the approval decision cannot distinguish a careless prompt from a careful one.

A visible per-prompt cost is a different instrument. It arrives at the moment of the decision, it is attached to the specific action, and it is information rather than permission. That is the shape of every good feedback loop engineering already trusts. A slow test suite that reports its own runtime gets faster. A build that shows bundle size in the pull request gets smaller. Nobody frames those as cost controls, and nobody sends them to finance for approval.

The equivalent instrument for AI usage is a number next to the action, visible to the person taking it, before they take it. Whether that number also rolls up into a budget is a separate question with a separate owner.

Who Should Be Asking For It

If cost visibility improves output quality, the request for it belongs to engineering leadership, and the argument for it is not a savings projection.

This changes the internal politics of the conversation in a way I think is underrated. Right now a platform team that wants per-prompt cost telemetry has to justify it as a FinOps initiative, which means it competes for budget against other FinOps initiatives and gets scoped to reporting. Reporting is monthly, aggregated, and invisible to the person prompting. It produces a dashboard nobody who writes prompts ever opens.

The version that works has to be in the loop. It has to be per action, at the moment of the action, in the interface where the work happens. That is a product requirement on internal tooling, not a finance requirement on internal spending, and it will get built correctly only if the person asking for it wants the behavioural effect rather than the report.

There is a fair objection here, and I want to state it rather than hide it. Making cost visible could also make people timid, avoiding expensive-but-correct operations to keep a number low. The study does not report that outcome, and the satisfaction result argues against it in the image-generation setting it measured. It does not rule it out for other kinds of work, and if you deploy this you should watch for it directly.

Do This Now

Pick the AI tool your team uses most, and find out whether the person using it can see what a single operation costs at the moment they trigger it. Not the monthly total. Not the team dashboard. The cost of this action, before this action runs.

For most teams the answer will be no, and the reason will be that nobody asked, because cost visibility was filed as a finance concern and finance only needed the aggregate. Change the requester. Have the engineering lead ask for it, and have them ask for it as a quality instrument: we want people to think for one extra beat before they generate, and we want fewer near-duplicate retries in the loop.

Then measure the thing the study measured. Not spend. Time per prompt, retry counts, and whether the people doing the work say the output got better.

---

This analysis synthesizes Show the cost of each AI prompt (Science Says, Thomas McKinlay, September 2026), summarizing research published in Information Systems Research in April 2026.

Victorino Group helps engineering organizations design AI cost telemetry that improves the work instead of only reporting on it. Let’s talk.