Per action or per month

What we are solving

In practice there are two models: people pay per action, or people pay per month. Everything else is a variation on those.

You pick a line in a price list and you are really picking what you will be doing all next year. A subscription drags one kind of work behind it, packs another, and both kinds start before your first revenue.

So do not choose on which earns more in theory. Choose on what job your product does, and on which model you can service on your own.

Steps

First question — is the job one-off or recurring

A subscription works where the person comes back by themselves, without being reminded — not because you retained them, but because the job arrives at them again.

If the job is one-off, a subscription is churn on a delay: somebody pays once, does not open it for a month, and cancels. They are entirely right to.

Here is what that looks like for me. Of 155 people who got at least one successful result, 117 used it on exactly one day, another 17 stopped at two, and six got past five.

Days of use among the 155 people who got a result

1 day3–5 days6+ days

Всего 155. Наведите на столбик, чтобы увидеть значение.

Days of use among the 155 people who got a result. 1 day — 117, 2 days — 17, 3–5 days — 15, 6+ days — 6.

Measured 13 August 2026 on the Voice AI database. Three quarters of them fit inside one day, which leaves a subscription nothing to sell.

There is no recurring job in those numbers. A subscription would not have held together on them at all: I would have been selling access to something nobody comes back to.

What a subscription demands beyond the product

People do not buy a subscription for a feature. They buy it on trust extended several months forward, and that conversation runs on your name rather than on your button.

Hence the first demand: markedly more brand-led marketing. A side project without a name has a harder time selling a recurring charge than a one-off one, at identical product quality.

The second demand is refunds, and it is not a detail. People forget to cancel, notice the charge, come and ask for their money back, and say they did not use it.

One query against your own database settles it: charged, and not a single action in that period. They are usually right, and the decent answer is a refund — a separate process you build once and then run by hand, month after month.

Why I chose pay-as-you-go

Three one-off packs: Starter — 800 credits for $5, Standard — 2,000 for $12, Pro — 7,500 for $40. Credits are spent per action and do not expire on a calendar.

The reason is the one I just gave: there is no recurring job visible in my audience, and somebody who arrives with a recording gets text back and goes about their day. That is a normal scenario rather than a failure.

The second reason is workload. Packs require neither name recognition nor a cancellation process. The person pays for the thing they already intend to do right now.

What pay-as-you-go does not fix

It does not save you from refunds. The only genuine payment in my database was refunded 21 minutes after the purchase, and the model had nothing to do with it.

It gives you no predictable revenue. A month without purchases is simply a month without money, and there is nothing there to plan on.

And it moves the entire load onto the moment of purchase. Somebody has to decide to pay exactly when they want the result, not later, and any hesitation at that moment costs you the payment.

How to decide in one evening

Work out, from your own database, the share of people who came back on a second day. Count actual use rather than visits.

Then the decision is simple. Almost nobody returns — charge per action. A noticeable share returns — a subscription becomes honest, and that is when brand and a cancellation process start earning their keep.

Either way the price does not come out of your head: getting the first person to pay.

What did not work

• Treating a subscription as the more grown-up model. It is not more grown-up, it is more expensive to run. On a one-off job it is just churn postponed by a month.
• Choosing the model before looking at repeat use. The number was already in the database: 117 of 155 used it for one day. I could have known that before I thought about a price list.
• Assuming packs spare you refunds. My only genuine payment came back 21 minutes later. A refund is about the product, not about the billing model.
• Shipping three packs without testing one. Starter, Standard and Pro appeared at the same moment. One was ever bought, so about the other two I know nothing at all.

Verify

• Compute the share of users who returned on a second day. If that number does not exist, the event row is not carrying it.
• Write in one sentence what the person is paying for: an action performed, or the ability to perform it all month. If it will not come out, the model is not chosen yet.
• If you picked a subscription: write down the refund procedure before launch. You will need it, and inventing it during the first email is worse.
• Check that the price and the unit being spent are visible before payment rather than after. A credit that spends on something unclear stops a purchase as effectively as a high price.

What one action actually costs you is counted separately: what one user costs.