What AI work should cost
Flat monthly pricing assumes that serving one more customer costs almost nothing. AI work costs something every time it runs, which makes a subscription a bet placed before either side knows the odds.
Software pricing has held one shape for twenty years because the economics underneath it held one shape. Serving one more customer of a web application cost close to nothing, so a flat monthly price was both simple and honest. The vendor’s costs barely moved, the customer’s bill didn’t move at all, and everyone could plan a year ahead.
AI work breaks that arrangement. Every question answered, every document read, every hour of delegated research carries a real cost that lands the moment the work runs. Two businesses on the same plan can differ by a factor nobody would have believed under the old model. The pricing shape the whole industry inherited assumes a marginal cost near zero, and the thing it is now being wrapped around does not have one.
A subscription is a bet placed before anyone knows the odds
To price a monthly fee for AI work correctly, you would need to know in advance how much work a given business is going to delegate. On day one, neither side does. The buyer has never had a workforce like this and genuinely cannot tell you whether they will reach for it twice a week or all day. The vendor has a few months of data, most of it from people who are still poking around.
So the number gets guessed, and somebody pays for the guess. Guess high and the light users are quietly funding the heavy ones, which they work out eventually, usually at renewal. Guess low and your best customers become the ones you can least afford to keep, and that is where the rationing starts: a limit that was never in the contract, a cheaper model swapped in without an announcement, a queue that got slower for reasons nobody will put in writing.
That failure is already common enough to recognize. When usage is unpriced but not actually unlimited, the vendor still has to control it, and the only levers left are the ones the customer cannot see. A meter is the honest version of the same constraint. It is a limit you are allowed to read.
Charge for the work that was performed
The model that survives contact with these economics is the plain one. You pay for what was done for you. Amolfi measures usage as the work it actually performs on your behalf: the research, the drafting, the document processing, the meetings, the agent runs. Not raw model tokens, which are an implementation detail of somebody else’s product and a miserable unit for a business owner to think in.
The unit is the whole argument, really. A number is only honest if the person paying it can look at it and recognize what it describes. An agency owner knows what a proposal costs to write, because they have priced that work for clients for years. Tokens tell them nothing except that a bill arrived. If a meter is going to sit between a business and its software, it has to count in the currency the business already thinks in.
Two things follow from metering that a flat fee cannot offer. A slow month costs less than a heavy one, which is how every other input to your business already behaves. And it puts the vendor on the right side of efficiency: when the work is metered, making the work cheaper to perform is something that reaches you. Under a subscription every efficiency gain is margin. Under a meter it is a smaller bill.
One meter, whichever door you came through
A business does not do its work in one place. The same person asks something from the app in the morning, runs a job from a terminal after lunch, and has their own code calling in overnight while nobody is watching. Three doors, one business, and the work behind them is identical.
The meter should be identical too. A separate, higher rate for programmatic usage is a tax on the way you chose to work, and it teaches you to route around the pricing instead of toward whatever tool actually suits the job. The same work should cost the same whether a person asked for it, a script did, or an agent of yours did while you slept. One meter, every door, and no arithmetic to do before you decide how to work.
The case for a way in that costs nothing
Metering also settles an old argument about free access that subscriptions never could. A free tier under a flat model is a subsidy: the vendor absorbs the cost of serving people who may never pay, so free has to be hobbled enough to hurt, or it eats the company. Everyone involved knows this, which is why free tiers so often feel like a room with the furniture bolted down.
When the work is funded before it runs, that arithmetic changes completely. An account that costs nothing until somebody asks it for something cannot run up a bill the vendor is left holding. The only variable cost is the AI work, and the AI work has already been paid for. Which means the way in doesn’t have to be a sandbox with a fortnight’s clock ticking behind fake data. It can be the real product, pointed at the real business, at whatever size that business currently is.
That is the position. A business tool should be enterable. The meter should count work rather than headcount or tokens. And the number at the bottom should be one both sides can point at and agree describes something that happened. Get those right and you end up making your money the way your customers make theirs: by being worth what the work was worth.