The cost of Amazon Quick is not driven by headcount, and that is the single most common budgeting error. A per-user licence is only part of the bill; the second part is a fixed account-level fee that appears on the higher tiers, and the third is the capability tier itself, which changes what the platform can do rather than how many people can use it. A hundred-person estimate built on seat count alone will be wrong in both directions — too high on users, too low on the total.
The three components of the bill
Treating this as a seat-based product produces a number that survives until the first invoice. The structure has three moving parts, and each behaves differently as the deployment grows.
Per-user licences scale linearly and are the part everyone models correctly. The fixed account fee does not scale at all, which means it dominates small deployments and disappears into the noise of large ones. The capability tier is a step function: nothing changes until a required capability sits above the current tier, at which point the whole account moves up.
The failure mode is predictable. A pilot is budgeted at the lowest tier for a small group, the economics look excellent, and the rollout then hits a capability boundary that moves the entire account to a higher tier — at which point the per-user maths that justified the project no longer holds.
What a hundred-user estimate should include
Building a defensible figure means separating what is known from what is assumed, and being explicit about the second category.
- Split the user base by need, not by headcount. Most organisations have a small group who build and maintain agents and a much larger group who consume the output. These are not the same licence, and conflating them inflates the estimate substantially.
- Identify the capability boundary first. List the actions the agents must perform, map each to a tier, and take the highest. Doing this before the user count prevents the step-function surprise.
- Add the fixed component once. It does not multiply, and forgetting it is the most common reason a pilot budget looks better than the production one.
- Model the consumption that sits outside the licence. Connected data sources, storage and the compute behind the actions are billed separately in most architectures. A licence estimate that stops at the licence is incomplete.
Where the real money goes
In the deployments that go badly, the overspend is rarely in the licence line. It appears in three places that no one modelled.
Agents that nobody retired. Configuration accumulates. An agent built for a process that changed six months ago still runs, still consumes, and still counts. Without an owner and a review cycle, this grows quietly.
Tier migration mid-project. Covered above, and worth repeating because it is the largest single item when it happens.
The integration work that was assumed to be included. Connecting a system that has no ready connector is a project, not a configuration step. This is the line most often missing from a first estimate.
How to present the number
A finance director does not want a licence table. They want a three-year total against the current cost of doing the same work, with the assumptions visible enough to argue with.
That means one page: the current cost in hours and money, the projected cost including all three licence components plus consumption, the difference, and a short list of what would have to be true for the projection to hold. The last item is what makes the document credible — an estimate with no stated assumptions reads as a sales figure, not an analysis.
Key points
- The bill has three parts: per-user licences, a fixed account fee, and the capability tier.
- Tier changes are a step function, not a gradient — identify the required capability before counting users.
- Builders and consumers of agent output are different licence needs; conflating them inflates estimates.
- Consumption behind the actions is billed separately and is frequently omitted from first estimates.
- Present a three-year total against current cost, with assumptions stated explicitly.
Does the cost per user fall as the deployment grows?
Effectively yes, because the fixed account component is spread across more users. This is why very small deployments can look disproportionately expensive per head, and why shrinking a pilot below a working team rarely saves what people expect.
Which tier should a first deployment choose?
The lowest tier that includes every capability the planned agents require — determined by listing the actions first. Choosing on price and discovering the gap later costs more than the tier difference, because migration carries project cost as well as licence cost.
What is usually missing from a first estimate?
Three things: the fixed account fee, the consumption behind connected systems, and integration work for systems without a ready connector. Together these often exceed the licence line in the first year.
How do you justify the spend when the savings are in staff hours?
By converting hours into a figure the organisation already accepts — either fully loaded cost per hour for the affected roles, or the value of the capacity released. The conversion rate matters less than agreeing it before the pilot rather than after.