What does the myth say?
That once you have an AI employee, the work is essentially free: no salary, no benefits, no office, and a subscription that rounds to nothing next to a person. Under this belief, any function should move to a Digital FTE because the cost comparison is not close, and any pilot that does not show enormous savings has been run badly. It is the mirror image of the replacement myth, and it fails the same way: by not doing the arithmetic.
The examination is about what the cost actually includes, and how to compute it rather than assume it.
Where does the belief come from?
From consumer AI pricing, where a monthly subscription buys unlimited conversation, and from vendor decks that show a salary next to a subscription and let the reader draw the line. It also comes from the true observation that a model call costs fractions of a cent, and from the mistaken inference that a business function therefore costs fractions of a cent. The belief is what a buyer arrives with, and a vendor that does not correct it is setting up a disappointing review.
This one corrects it, because the platform is built to be measured and an unmeasured expectation is the enemy of that.
What is true in it?
That for repetitive, evidence-based work, cost per outcome from a Digital FTE is often well below what the outcome cost before, once the approval habit exists and the routine classes have earned autonomy. That there is no salary, no benefits, and no notice period, and the block can be changed or cancelled from the workspace. That a function which was never done at all, overnight order-status replies, weekly citation sampling, now gets done for the cost of the tokens. Cheaper is frequently the right word; free is never.
The myth's error is in assuming the answer instead of computing it.
What does the cost actually include?
The plan block: a purchase of tokens for the organization, drawn down by every role, with a hard stop at zero and nothing on credit, so the maximum exposure in a period is the block already bought. The manager's time: approvals in minutes a day, send-back reasons weekly, a monthly review, and setup in the first week, all of which the baseline sheet should log. External spend a role controls: ad budgets for a Performance Ads FTE, billed by the ad platform under its own cap, never by ftes.ai. And the model tier choice per role, which trades cost against quality and is labelled honestly in Setup. Usage is reported per role, so each role's cost can be read on its own.
The hidden-costs piece in the library lists what to look for in any vendor's number, including this one's.
- Block: tokens, ceiling, hard stop
- Time: approvals, reviews, setup
- External: platform spend under its own cap
- Tier: cost against quality, per role
When is the AI employee not the cheap option?
When the work is mostly judgement, because every item escalates and the person does the work anyway plus the approval. When the volume is too low for the setup and management time to amortise. When the function needs a relationship, a negotiation, or a decision about a person, none of which the FTE does. And when the baseline was already efficient, a well-run team with good tooling, so the outcome was not expensive to begin with. The four-option framework in the library puts hire, agency, AI employee, and doing nothing on the same five dimensions, and the AI employee does not always win.
A vendor that says it always wins is doing the arithmetic for you, which is the thing to refuse.
How can you check for yourself?
Record the baseline before the pilot: hours per outcome at your team's rates, plus any external cost. Run two full cycles. Read the role's usage in its Performance tab and multiply by the plan rate. Count the outcomes in the Targets tab with their evidence. Add the approval, review, and setup hours you logged. Divide. Compare with the baseline, labelling the first cycle as first-cycle. The number is yours, from your own record, and it is defensible to a board because nothing in it came from a vendor.
The whitepaper on the economics of digital labor works the arithmetic through with the inputs left blank for your figures.