Teams that estimated cost from agent count were wrong and teams that estimated from skill mix were close, because autonomy draws five times what retrieval does
There is no standalone price line to negotiate. You negotiate the credit pool and the rate card behind it, and the mix decides how fast the pool empties.
Prepared by Redress Compliance · August 19, 2026 · Workday agent forecasts. 30 forecasts built with buyers, 2024 to 2025.
Executive summary
Teams that estimated from agent count were wrong; teams that estimated from skill mix were close. The autonomous completion skills were the swing factor.
The complimentary window was the only reliable source of action volume. Estates that ignored it forecast on guesses.
Retrieval heavy agents were cheap and predictable; autonomy heavy agents needed careful headroom. The two behave nothing alike on the meter.
A sensitivity run on the heavy case saved several buyers from signing a tier that could not absorb adoption. Adoption moves the mix, not just the volume.
How do you build the forecast?
By mapping each agent to its skills, estimating monthly actions per skill, and multiplying by the credit value on the rate card. Three steps, and the second is where the accuracy lives.
Step one, inventory agents and skills
List every active agent and the metered skills it runs, using the agent record as the source so nothing is missed. Group by function so the forecast maps to budget owners and duplicate agents are not double counted.
Step two, estimate monthly actions
Estimate actions per skill from real usage, separating retrieval from autonomous completion. Pull volume from a full month rather than a busy week so seasonality does not distort the base, and where usage is thin hold the number conservative.
Step three, multiply and keep the blended rate visible
Multiply actions by the per skill credit value and sum. Keep blended credits per action visible as a single figure, because a blended rate drifting above 2 signals an autonomy heavy estate that needs more headroom. The catalog of agents and skills is published at the agents catalog.
What does the meter actually charge for?
A completed task, not a prompt or a token. That single design choice is why skill type matters more than usage volume.
| Skill type | Credits per action | Meter trigger | Forecast behaviour |
|---|---|---|---|
| Retrieval, self service | 1 | Answer returned | Cheap and predictable |
| Guided draft | 2 to 3, illustrative | Draft produced | Mid range, confirm your own card |
| Autonomous completion | 5 | Task completed end to end | The swing factor, needs headroom |
There is no standalone line to negotiate
The agents draw on the same universal credit rate card as every other agent on the platform, so what you negotiate is the credit pool and the rate card behind it. The published rates are set out on the credit rate card.
Only the published values are firm
One credit for retrieval and five for autonomous completion are published. Mid range values are illustrative, so confirm your own rate card rather than modelling from a table.
The credit pool negotiation playbook
How the meter behaves, what the complimentary window proves, and the buyer side moves before the pool is sized.
Get the brief →What 30 Workday agent forecasts showed
Across roughly 30 Workday forecasts Fredrik Filipsson built with buyers in 2024 and 2025, the pattern was consistent. Three findings recur.
- The complimentary window was the only reliable source of action volume, and estates that ignored it forecast on guesses.
- Retrieval heavy agents were cheap and predictable, while autonomy heavy agents needed careful headroom.
- A sensitivity run on the heavy case saved several buyers from signing a tier that could not absorb adoption.
The autonomous completion skills were the swing factor, because a modest rise in their use moved the monthly burn sharply.
- Commitments sized from your actual consumption curve rather than last year's footprint
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A ranked savings queue with dollar values, not license counts
Why does the mix beat the count?
Because the arithmetic is plain. Hold actions flat and raise the autonomous share, and the burn climbs fast.
The worked example
Thirty five thousand retrieval actions at 1 credit, six thousand guided drafts at 2, and five thousand autonomous completions at 5 produce 72,000 credits from 46,000 actions. The blended rate is 1.57, and that single number is the fastest sanity check at the table.
Move five thousand actions and the total climbs
Shift 5,000 actions from retrieval to autonomous and the burn rises by 20,000 credits a month, because each of those actions now costs 5 instead of 1. Volume held flat, the total still climbs.
Adoption changes the mix over time
As users trust the agents they hand over more complete tasks, which is exactly the shift from retrieval to autonomy. A forecast built on today's mix underestimates a successful rollout. The credit mechanics sit in the credit reference.
Watch the briefing · 4:265 Ways to Win Your Workday NegotiationWhy the free window is not generosity, what the credit math really costs, and which four terms belong in writing.
How does the complimentary window fit?
Every subscription includes an annual allotment sized to the company and renewed each contract year. You only pay once production usage exceeds it.
It is the only honest source of volume data
Experimentation outside production draws no credits, which is why the complimentary window is safe to load test. Estates that ignored it forecast on guesses, and guesses sized the pool.
Load test the heavy case before you commit
A sensitivity run on the autonomy heavy scenario is what saved several buyers from signing a tier that could not absorb their own adoption curve. The governance layer sits in the agent governance reference and the estate view in the credits pillar.
Where the common advice on agent cost is wrong
The common advice is to forecast from the number of agents you plan to deploy. We disagree.
Agent count is the wrong denominator
Teams that estimated from agent count were wrong and teams that estimated from skill mix were close, because one autonomous agent can outspend five retrieval agents without any of them changing their volume.
The buyer side move is to inventory skills rather than agents, pull volume from the complimentary window, model the autonomy heavy case explicitly, and negotiate the pool and the rate card rather than a product line that does not exist.
The wider pricing picture sits in the agent pricing guide, and the product expansion is described at the product announcement.
What the forecasts measured, 2024 and 2025
Two cuts, and the first explains why the second is the whole exercise.
Against 1 for a retrieval action, which is why the share of autonomous work sets the burn more than the volume does.
Moving that many actions from retrieval to autonomy, with total volume held completely flat.
Neither number depends on adoption growing. Both apply the moment the mix moves.
Your first five moves
- Inventory skills rather than agents, because the metered skill sets the cost and the agent count says nothing about the mix.
- Pull action volume from the complimentary window over a full month, since it is the only reliable source and a busy week distorts the base.
- Separate retrieval from autonomous completion in every estimate, as one draws five times the other on an identical action count.
- Keep the blended credits per action visible as one number, because a blended rate drifting above 2 signals an estate that needs more headroom.
- Run the autonomy heavy sensitivity before signing a tier. The Workday practice models the heavy case first, which is what stopped several buyers signing a pool their own adoption would break.
Frequently asked questions
What should the forecast be built from?
Skill mix rather than agent count. Teams that estimated from the number of agents were wrong; teams that estimated from the metered skills were close.
When does the meter run?
When a task is complete, not per prompt or token. A retrieval skill draws about 1 credit per action and an autonomous completion about 5.
Is there a separate product line?
No. The agents draw on the same universal credit rate card as every other agent, so what you negotiate is the credit pool and the rate card behind it.
Why does the mix matter more than volume?
Because moving 5,000 actions from retrieval to autonomy raises the burn by 20,000 credits a month with total volume completely flat.
What is a healthy blended rate?
Keep it visible as one number. A blended rate drifting above 2 credits per action signals an autonomy heavy estate that needs more headroom in the pool.
Where does volume data come from?
The complimentary window. Experimentation outside production draws no credits, which makes it safe to load test and the only reliable source of action volume.
What do complimentary credits cover?
An annual allotment sized to your company and renewed each contract year. You pay only once production usage exceeds that allotment.
Which agents are predictable?
Retrieval heavy ones. They are cheap and their burn moves with volume. Autonomy heavy agents need headroom because a small mix shift moves the total sharply.
Does adoption change the forecast?
Yes, and in the expensive direction. As users trust the agents they hand over more complete tasks, which is exactly the shift from retrieval to autonomy.
What protects against that?
A sensitivity run on the autonomy heavy case before signing. It is what saved several buyers from a tier their own adoption curve would have broken.