Ask AI 'what will it do, and what will it cost' at the idea stage── Have value and cost set side by side for the same plan early, and you can decide to continue or drop it before building

01Before building, the operator kept asking the AI about both the gains and the fees

Take an exchange from 5 April 2026 in the record. The operator was weighing a plan to extend a system that ran on one AI model so that it would run on four: the original model plus three models from other companies.

The requests that day were not instructions to build anything. First, the operator asked what the system could now do at its current stage. Next came a question about the monthly cost of adding the other companies' models.

Then the operator asked how large the difference would be between running one model and running four. After that, the operator asked again what the whole thing would make possible. Finally, the operator asked the AI to look up each provider's per-token prices, estimate how many tokens the work would consume, and give a daily and a monthly cost.

Every question that day was about one of two things: value or cost. Value was asked about twice, cost three times, each time in a different form. The cost questions grew more precise as they went, from a rough monthly figure, to the gap between one model and four, to daily and monthly totals built from unit prices and estimated volume.

The order is the point. Before any assembly began, the operator moved back and forth between a question about value and a question about cost, several times, on the same day.

02Put value and cost side by side and an idea becomes a keep-or-drop decision

Ask only what something will do, and the answer tends to be a list of benefits. Ask only what it will cost, and you get a figure with no way to tell whether it is high or low.

Ask both about the same plan and you can compare. What does a four-model setup allow that a single model does not? How much extra do you pay for that difference? Once those two answers sit in the same table, the idea stops being an impression and becomes something you can decide on.

03The earlier you estimate, the cheaper it is to walk away

The cost of abandoning an idea grows with every step taken. At the idea stage, dropping it costs a handful of messages. After assembly, it costs the configuration you wrote, the connections you set up and the fees you already paid.

With setups that combine several AI models, the gap is wide. Writing about its own research feature, Anthropic reports that agents (AI that chooses its own steps to carry out a task) typically use about four times as many tokens as a chat, and that multi-agent systems use about fifteen times as many. Tokens are the units in which a model processes text, and they are what you are billed for.

The same report states that such systems pay off only for tasks whose value is high enough to cover the added cost. Weighing value against cost is therefore a condition to check before building, not a lesson to learn afterwards.

04An estimate here means the order of magnitude and the difference, not a budget

The estimate you ask for at the idea stage is not a forecast of the invoice. Its purpose is a number accurate enough to decide whether the idea is worth pursuing.

It has two parts. The first is value: being able to name, as concrete tasks, what the plan would newly make possible. The second is cost: price per token, times tokens per run, times runs, giving the order of magnitude for a day and a month.

There is also a boundary. An idea-stage estimate should not go into a budget request or a contract. Those need figures measured on a running system. What the early estimate decides is only whether to proceed to that measurement.

05In promotional material review, it pays off before you decide the number of models per item

The same question comes up when a pharmaceutical company brings AI into the review of promotional material. Do you have one model read each item? Or do you split the work, with one model checking wording, another checking claims against their evidence and a third checking regulatory points?

Splitting looks as if it should catch more. But every extra model reads the same material again, and the fee rises roughly in step with the number of readings. Multiply by the number of items reviewed each month and the difference is not small.

So at the idea stage, ask two things. How many additional findings, specifically, would a split setup catch that a single model misses? And how much per month would that cost? With both in hand, you can decide whether every item goes through several models or only the high-risk ones.

The same holds in medical affairs and marketing: having several models summarise a paper and comparing the results, or having a second model check a draft reply to an enquiry. In each case, setting value against cost early lets you fix the scale before you try anything.

06Costs can be estimated quickly because they are public prices multiplied by volume

Most generative AI is billed at published prices. Anthropic and Google both publish separate input and output prices per million tokens. Once the prices are public, a working assumption about volume is all you need to get a figure by arithmetic.

This arithmetic suits an AI well. Reading a price list, estimating how many tokens one run reads and writes, and multiplying by the number of runs is a fixed procedure. The person using it only has to supply the volume and frequency assumptions.

Value is different. Asked what a plan will do, an AI tends to return a list of things that might be possible. Choosing which of those the work actually needs is the user's job.

Anthropic's guidance on building agents recommends starting with the simplest solution and adding complexity only when it is needed. It describes agentic systems as trading latency and cost for better task performance. The value question exists to check whether that trade holds.

07From tomorrow, put four questions to the AI for each plan, on the same day

When a new plan comes to mind, ask the AI these four questions before asking it to build anything.

1. What would this make possible? Ask for up to three tasks, by name.
2. How do a simple version and a complex version differ? Specify the pair: one model against several, manual against automated.
3. What does each cost per day and per month? Have it calculate from public price lists and stated volume assumptions, and write the assumptions down.
4. Is the difference worth the money? If not, ask how far the plan would have to shrink before it is.

With the answers in hand, choose one of three: continue, scale down, or drop. Keep the questions and estimates for dropped ideas too; if prices fall, you can revisit them.

The figure the AI gives changes with its volume assumptions. So ask once more what happens if volume doubles. If the order of magnitude holds, the decision holds. If it shifts, measure that volume for real first.