Budgeting and capping AI usage costs: tokens, seats, and overage

Reference article · 5 min read

Agency budgets are set once a year and are difficult to amend. Most AI software is priced to move with use. That is the budgeting problem in one sentence: a tool that costs a modest amount during a controlled pilot can cost something very different once a busy department discovers it, and the discovery usually arrives mid year, when the appropriation is fixed.

The goal is not to forecast usage perfectly. It is to structure a purchase whose worst case you can state before signing, whose spend you can watch weekly, and whose ceiling is enforced by the system rather than by a vendor's assurance.

The pricing shapes you will encounter

Per seat. A fixed price per named or concurrent user. Predictable, but definitions vary. Confirm whether occasional reviewers, approvers, contractors, and service accounts consume seats, and whether a seat transfers when staff leave.
Consumption. Metered by tokens, pages, documents, minutes of audio, images, or API calls. Predictability depends on how well you understand the unit being counted.
Hybrid with overage. A platform fee plus an included allowance plus a rate for use beyond it. This is where surprises live. Ask whether unused allowance rolls over, whether it pools agency wide or is locked per department, and how overage compares to prepaid units.
Per outcome or bundled. Some tools price per processed case or call, and some include AI in a module you already license.

Learn the meter before you trust any estimate

A unit label is not self explanatory. Tokens are fragments of text rather than words, and both what you send and what the system returns are counted. A page of dense scanned tables may bill as more than one unit, and retries, background indexing, and failed jobs may or may not be billable. Ask the vendor to define the billable unit in writing, price a sample workload of your own documents, and attach that definition to the contract instead of leaving it in documentation the vendor can revise.

Estimating annual cost from a pilot

Pilots understate steady state cost, because participants are selected, trained, and supervised. Build the estimate from unit rates, not monthly totals.

Export the raw usage meter from the pilot, not a summary slide.
Compute consumption per active user per working day, and read the distribution rather than the average.
Identify the heavy tail. One unit usually finds a workflow that fits the tool unusually well and will drive a disproportionate share of spend at scale.
Multiply by the population you intend to onboard, in the order you intend to onboard it, and add growth for deepening use.
Model three cases: expected adoption, heavy adoption, and a runaway case in which one workflow is automated and runs unattended.

Budget the expected case and cap near the heavy case. Layer in seasonality specific to government work, such as budget preparation, tax cycles, permit season, storm response, and elections. Account separately for backfill jobs, which can consume more over a weekend than routine use does in a quarter.

Caps, alerting, and enforcement

Hard caps and soft caps. A hard cap stops service; a soft cap notifies and keeps running. Do not hard cap a workflow supporting emergency response or a statutory deadline.
Where the cap lives. A not to exceed amount in a contract is a legal ceiling, not a technical one. Ask whether the platform enforces a usage limit at the tenant level, and who can raise it.
Alerts that reach the right people. Route threshold notices to the budget owner and the program manager, not only to a system administrator.
Per department and per user limits. These contain a misconfigured integration before it becomes an invoice. Unattended jobs warrant tighter limits than interactive users.
A visible burn down. Require a view of consumption to date against the allowance, with a forecast to period end.

Chargeback and internal allocation

Decide early whether central IT absorbs the cost or departments are charged for it. Central absorption encourages adoption and removes spending discipline. Full chargeback creates discipline and can price small departments out of a tool the agency wants standardized. A common middle path is a central allocation covering seats and a baseline allowance, with consumption above that charged back.

Attribution makes any of this possible. Tag accounts to cost centers at provisioning, confirm the vendor's usage export carries that tag, and issue a monthly statement per department. Departments funded by grants or restricted revenue usually need separate cost centers, and allowability against a given award should be confirmed with the awarding agency and your grants office.

What to require in the contract

A not to exceed amount, with language stating that charges beyond it do not accrue without a written amendment.
Fixed unit and overage rates for the full term, and a stated cap on renewal uplift.
No automatic tier promotion without written authorization from a named agency official.
Advance notice at defined usage thresholds, a right to detailed usage records, and a billing dispute process.
The ability to reduce seats and allowances at renewal, not only to add them.
Termination and non appropriation protection. Public contracts commonly include non appropriation language, but the required form varies by state and local charter, so confirm it with counsel and purchasing.
Notice of material change to the underlying model. Consumption per task can shift when a vendor swaps models, even though your work does not. Ask for a remedy if units for the same workload materially increase.

Turn on metering and cost center tagging on the first day of the pilot, and write down the expected, heavy, and runaway numbers before negotiating. These terms, rather than the headline unit price, keep a mid year invoice from becoming a budget amendment.

HQT works with state and local agencies on technology planning and procurement support, including usage modeling for consumption priced software.

Frequently asked questions

How do we estimate annual AI costs from a short pilot?

Export the raw usage meter from the pilot and compute consumption per active user per working day rather than relying on a monthly total. Scale that rate by the population you actually plan to onboard, in the order you plan to onboard it, and add allowance for users consuming more as they grow familiar with the tool. Then model an expected case, a heavy adoption case, and a runaway case in which one workflow is automated and runs unattended, and budget and cap against those three numbers.

What is the difference between per seat and consumption pricing for AI software?

Per seat pricing charges a fixed amount for each named or concurrent user, which makes budgeting predictable but can leave you paying for licenses nobody uses. Consumption pricing charges for units such as tokens, pages, documents, or minutes of audio, which tracks actual value but varies month to month. Many vendors combine the two with an included allowance and an overage rate, so confirm how the allowance pools, whether it rolls over, and what the overage rate is before comparing offers.

Can an agency set a hard spending cap on an AI vendor?

A not to exceed amount in the contract is a legal ceiling, but it does not stop the system from running. Ask during evaluation whether the platform supports an enforced usage or spending limit at the tenant level, who is able to raise that limit, and what happens to users when it is reached. Avoid hard caps on workflows that support emergency response or statutory deadlines, and use graduated alerts and per department limits there instead.

Should we charge departments back for AI usage or absorb it centrally?

Central absorption encourages adoption but removes the spending discipline that comes from seeing a bill. Full chargeback creates discipline but can price smaller departments out of a tool the agency wants standardized. Many agencies use a hybrid in which central IT funds seats and a baseline allowance and charges back consumption above it, which requires tagging accounts to cost centers when they are provisioned and confirming the vendor's usage export carries those tags.

What contract terms help prevent a surprise AI bill mid fiscal year?

Useful terms include a not to exceed amount stating that charges do not accrue beyond it without a written amendment, unit and overage rates fixed for the term, no automatic promotion to a higher tier, and advance notice at defined usage thresholds. A right to detailed usage records and a defined billing dispute process help when an invoice looks wrong. Requirements for termination and non appropriation language vary by state and local charter, so confirm the required form with your counsel and purchasing office.