With Workday's usage-based pricing, an agent's bill is set mostly by how it was designed, not by how many people use it. Workday meters Flex Credits "when a defined task with known value is complete," and gives each skill its own credit price. On Workday's published example, the Self-Service Agent's information retrieval costs 1 credit per action and its autonomous task completion costs 5.

Custom agents became generally available in Workday Extend AI at Workday Rising on October 13, 2026. As teams start building their own, the design choices below decide whether a useful agent is also an affordable one.

1. Give each agent one job

An agent that "helps with HR" will try to do everything, and every extra thing it does is another possible metered action. An agent that "prepares a manager's time-off approvals for review" has a short list of things it may do and a clear point at which its work is finished.

Narrow agents are also easier to secure and test. Workday's custom agents inherit the user's permissions, but you still choose which tools an agent can call. Fewer tools means fewer ways to spend credits by accident.

2. Answer before you act

Retrieval is cheap and action is expensive. Design the conversation so the agent first shows the person what it found ("You have 6.5 days of vacation; the request would cover 3") and only performs the task when they ask for it.

Many requests end at the answer. If the agent submits something every time it is asked a question, you pay the task-completion price for work nobody wanted.

3. Ask once, at the right moment

Workday's custom agents can "pause for human approvals when needed." Use that pause where it saves the most: immediately before the one action that changes data. One clear confirmation with a summary of what will happen avoids a costly wrong action and a second action to undo it.

Do not scatter confirmations through every step. Each extra round trip is friction for the user, and does nothing to protect the credit balance.

4. Stop loops before they start

The most expensive agent is one that retries. If a submission fails validation, an agent that keeps resubmitting with small guesses may trigger more metered work and still fail.

Tell the agent what to do when Workday rejects something: report the reason in plain language and stop. Then test that behaviour deliberately. Feed it a request that Workday will reject and watch what it does.

5. Batch where the process allows it

Some work is naturally one-at-a-time; some is not. A manager reviewing twelve pending requests does not need twelve separate conversations. Where Workday's process and the agent's tools support it, gather the items, present them together and let the person act on them as a set.

Check the Rate Card for how batched work is metered before you assume it is cheaper. Workday defines consumption by capability, event or skill, so a batch is not automatically one action.

6. Do your experimenting where it is free

Workday says customers can "test most agents for free in non-production," and metering happens almost entirely in production. Every prompt revision, tool change and edge case you try in a sandbox is one you are not paying for in Production.

That means a real non-production test plan:

  • a list of realistic requests, including awkward ones;
  • requests that should be refused because of the person's security;
  • requests that Workday should reject;
  • a count of actions per request, so the budget uses observed numbers.

Watch the console after go-live

Workday's consumption console shows current use, usage over time and alerts you can set yourself. Compare the first month's real actions per request with your sandbox numbers. A large gap usually means the agent is doing work nobody designed it to do, which is a design fix, not a reason to buy more credits.

How kiweely helps

kiweely is a desktop co-pilot that works against a client's real Workday tenants, with consent at each change. It is built for the work around an agent that decides its cost: setting up and checking security in the tenant, building and validating the Workday Extend apps and business objects an agent relies on, and testing the result in the browser as different users before anything reaches Production. Its AI runs on the ChatGPT or Claude account you connect, so all of that iteration happens without touching the client's Flex Credits.