AI tool pricing: compare quotes on a typical and a busy month

Three quotes for the same AI tool job sit in your inbox, and none of them can be compared with the others. One is priced per user per month. One sells credits. One is a flat monthly plan with an allowance and a rate for anything above it. Each looks reasonable on its own, and each is measured in a different unit.

The way through is to stop comparing headline prices and price your own month instead: a typical one and a busy one. This guide is for small-business owners and freelancers comparing AI assistant and agent subscriptions, with a fictional worked example in Singapore dollars.

Find the unit before you read the number

Every pricing model is a claim about what value scales with. An explainer on per-seat versus usage pricing for AI agents sets out the shapes that cover almost every quote:

• Per seat: a monthly price per named person. Easy to forecast, but it grows with adoption, so every colleague who wants access adds a line.
• Per usage: a meter counting tasks, runs, credits or tokens. A quiet month is cheap and a busy one is not, and the forecasting moves onto you.
• Flat deployment: a fixed price for reserved capacity, usually with a usage allowance and a rate beyond it. People are not counted.
• Hybrid: seats plus credits, or a base fee plus per-task charges. The explainer warns that it can inherit both downsides: adoption is taxed and usage still varies.

The same explainer argues that seats fit AI agents poorly, because much of the work happens with nobody logged in and the people messaging a customer-facing assistant are customers, not staff. So ask of each quote: does its unit describe anything real about how you will use the tool?

Describe your month in jobs, not people

Write down each job the tool would do and how often. Usage follows the schedule, and the schedule is something you set. Include the work nobody watches: scheduled summaries, standing checks and anything running overnight. Then estimate three months: quiet, typical and busy.

Take a fictional renovation-design studio with three staff and a part-time bookkeeper who only reads reports. Counting in the vendors' "task" unit, its month looks like this:

• Typical month: 450 enquiry replies drafted, 150 follow-up reminders, 60 proposal outlines and 40 scheduled summaries and checks. About 700 tasks.
• Busy month, after a promotion brings a surge of enquiries: 1,100 replies, 350 reminders, 110 outlines and the same 40 scheduled jobs. About 1,600 tasks.
• Quiet month: about 300 tasks, mostly the scheduled work plus a trickle of enquiries.

A worked comparison in SGD

The four quotes below are invented to show the method. They are not real prices from any provider, and every figure is illustrative:

• Quote A, per seat: S$45 per user per month, "fair use" included. Four logins are needed, including the bookkeeper: S$180 in every month.
• Quote B, usage: S$0.20 per task, no base fee. Quiet S$60, typical S$140, busy S$320.
• Quote C, flat deployment: S$150 a month including 1,000 tasks, then S$0.15 per extra task. Quiet S$150, typical S$150, busy S$240.
• Quote D, hybrid: S$25 per user plus S$0.10 per task. Quiet S$130, typical S$170, busy S$260.

Now spread it over a fictional year of eight typical, two busy and two quiet months: A comes to S$2,160, B to S$1,880, C to S$1,980 and D to S$2,140. The quote that is cheapest in a typical month is the most expensive in a busy one, and the annual spread is much narrower than the monthly one.

Then double the headcount, as the explainer suggests, because the studio plans to hire. With eight logins, A rises to S$4,320 a year and D to S$3,340. B and C do not change, because extra readers do no metered work.

The example does not produce a winner. It shows which risk each quote carries: A taxes growth, B spikes in a busy month, C's surprise is outgrowing its allowance, and D has a bit of both. Pick the risk you can live with, then negotiate the details that shrink it.

Questions that change the number

Put the same questions to every vendor, in writing:

1. What exactly counts as a task or a credit, what does "fair use" allow, and can you show me a sample breakdown?
2. Is scheduled and background work metered, and does it appear in the usage report? The explainer calls this the invisible line: work that never shows up in a conversation but does show up on the invoice.
3. What happens at a spending limit: a warning, a switch to a cheaper model, or a stop? An alert after the event is not a cap.
4. Does a colleague who only reads the output need a seat?
5. How are customer conversations counted?
6. Is the price in Singapore dollars or another currency, and does paying annually fix the unit price?

Keep the written answers with your other checkout records.

What you keep if you leave

The cheapest quote can be the most expensive to leave. A guide to AI agent platform lock-in argues that portability comes from formats, not promises: a data-portability clause helps, but it is a promise, and a promise needs the vendor to keep it. Add an exit line to each quote:

• Can you open the tool's instructions in an ordinary text editor, or only view them inside the product?
• Is memory a document you can read, or a store you can only query through the vendor?
• Are the procedures you taught it files you can copy?
• Do integrations use an open protocol such as the Model Context Protocol, or the vendor's own connectors, which would each need rebuilding?
• If you use your own model provider account, does it keep working if you leave?

The lock-in guide also names what is lost wherever you go: connected accounts must be authorised again with whatever you move to, and a running service that keeps messaging channels open and fires schedules does not export, because it is not data. Price the exit in hours of your own time and write that beside the monthly figure. For the fictional studio, a quote where its reply templates stay as plain text is worth more than a slightly cheaper one where they do not.

Common questions