What does an AI workflow really cost to run?

Start with a completed job

“Ten thousand users” is not a useful consumption estimate on its own. A user might ask one short question or run a document workflow that makes several model calls, searches an index and retries a failed integration. Define a unit you can count: an approved support draft, a processed application or a reviewed document.

For each unit, record the typical input size, output size, number of calls, tool use and expected review time. Include long documents and awkward cases in the sample. Keep development experiments separate from production usage so the first month does not become your forecast for every month.

Track cost per accepted job as well as cost per request. A cheap response that staff must rewrite can cost more than a better response with a higher model bill.

The costs to put in the worksheet

On a small screen, scroll the table sideways.

Specify the support window. Business-hours support and a staffed 24-hour response are different services. Name the response commitment, expected incident volume and escalation owner before calling either one “maintenance”.

A hypothetical monthly example in CAD

This is an arithmetic example, not an AKD quote, a provider price or a benchmark. Every rate below is invented for the worksheet. Replace the assumptions with measured pilot usage, current quotes and your actual staffing costs.

The input calculation is 20,000 × 2 × 6,000 ÷ 1,000,000 × CAD 3 = CAD 720. Output is 20,000 × 2 × 1,000 ÷ 1,000,000 × CAD 12 = CAD 480. The model subtotal is CAD 1,200. A hypothetical 15% allowance for extra calls adds CAD 180.

On a small screen, scroll the table sideways.

If 18,000 of those jobs pass the agreed acceptance check, the example costs about CAD 0.36 per accepted job. Initial delivery, migration, training and any extra licences, audit obligations or round-the-clock support must be estimated separately. Do not assume this allowance covers them.

Test the forecast before buying capacity

Run low, expected and busy-month cases. Keeping all other example assumptions unchanged, 5,000 jobs would give an arithmetic total of CAD 5,525; 60,000 jobs would give CAD 9,320. That is a sensitivity calculation, not a capacity promise. More traffic can require larger databases, more support hours or a different hosting tier.

Measure concurrency as well as monthly volume. Forty jobs arriving together can expose limits that forty jobs spread across a day will not. Record queue time, failure rate, response time and the cost of recovering incomplete work.

Provider pricing models differ. Use the actual billing units for the selected service, including storage or tool charges where relevant. AWS’s Generative AI Lens recommends controlling resource consumption and giving workflows explicit exit conditions. These are useful design checks even if your system uses another cloud.

Write usage limits that a buyer can understand

• Included work: define monthly completed jobs, maximum document size, users, concurrency and retention.
• Visibility: show consumption and forecast the month-end total. Choose warning thresholds, such as 70% and 90%, with the buyer.
• At the limit: agree whether the system queues work, pauses optional jobs or asks for an approved increase. A provider dashboard budget may be an alert rather than a hard cap; verify enforcement.
• Overages: state the unit, price, authorization route and maximum spend. Explain whether unused allowance rolls over.
• Changes: review longer inputs, new integrations, extra languages, different models or stricter service hours before they silently become part of the fixed fee.

The procurement pack should separate one-time delivery, recurring operations and variable consumption. It should also name who owns the cloud accounts and data, what gets handed over, and how the service can be exited. That makes competing offers easier to compare.