What does Dr. Jonah Tebaa's Working-Seat Count for Deciding AI Licence Renewals mean in practice?
Dr. Jonah Tebaa's Working-Seat Count evaluates AI licence renewals by calculating the cost per working seat, determined by dividing the monthly bill by seats that performed a specific task at least eight times across four weeks. Seats are categorized as working, occasional, or idle. In a hypothetical illustration, dropping twelve idle seats beat a fifteen percent volume discount on forty seats by $180 a month, testing discounts against buying fewer seats.
Price the renewal by the seats that work
Dr. Jonah Tebaa argues that an owner facing an AI licence renewal should work from one number, and that the number is not on the vendor's page. He calls it cost per working seat: the monthly bill divided by the count of seats that actually did the task the licence was bought for. His instrument, the Working-Seat Count, exists to produce that figure before anyone signs a twelve-month commitment.
The reasoning starts from what a seat is for. If a dashboard counts logins, then a seat is "active" when someone opened the tool. In his view, that tells an owner about access and nothing about work. A login-based count can flatter a bill, which is why he declines to negotiate from it.
Sorting seats into three bands
The count rests on a simple sort, applied to four weeks of recent use. Dr. Tebaa asks the firm to name the working task for each tool in one line, such as a client summary drafted or a meeting note filed. Each seat is then placed in one of three bands, against a line the owner fixes before looking at the data:
- Working: the task was done at least eight times in four weeks, roughly twice a week.
- Occasional: the task was done between one and seven times.
- Idle: the task was not done at all.
He stresses that the eight-times line is his working method, not research, and that an owner with a monthly task may reasonably choose another. What matters is that the line is chosen first. If it is chosen after, the numbers can be made to say whatever is convenient.
Where per-person figures are unavailable from the vendor, he suggests team leads tally the task for their own people. A seat that cannot be counted at all is treated as unproven, not as working.
The discount against the smaller order
The sharpest point in Dr. Tebaa's argument concerns the volume discount. A discount lowers the price per seat, he says, and leaves untouched the share of seats that work. So it should be tested against the alternative of simply buying fewer seats.
His illustration is hypothetical, with invented numbers, and he says plainly that it is not research. A services firm holds 40 seats at an assumed $30 a month each, a bill of $1,200. The vendor shows 28 seats as active and offers 15 percent off, $1,020, for twelve months on all 40. A four-week count finds 11 working seats, 17 occasional and 12 idle. Cost per working seat is $109.09, against a $30 sticker price.
Using assumed monthly values of $140 for a working seat, $40 for an occasional one and nothing for an idle one, the total value is $2,220. Renewing as is nets $1,020 a month. Taking the discount nets $1,200. Dropping the 12 idle seats and keeping 28 at $30, a cost of $840, nets $1,380. The smaller order beats the discount by $180 a month, which is $2,160 over a year. Dr. Tebaa also points out that the twelve-month commitment ties the firm to the idle seats for the whole term.
What the method does not claim
Dr. Tebaa is careful about the limits. The method does not promise savings. If the work is real but staff have not been shown how to do it with the tool, the right response may be training and not cancellation. Occasional seats are not worthless either, and he treats them as a separate negotiation: a shared pool, a lower tier, or a named reason to stay. He advises owners to read the contract's clauses on minimum seats, commitment term and notice before acting on any count, since terms differ by vendor.
He does not cite figures about how firms use AI licences in general, and he names no vendor. The numbers in his example are there to show the arithmetic, and each owner's own four weeks will give different ones.
The Monday step, and the original
His practical advice is brief. Before the next renewal date, collect four weeks of per-seat task counts, sort the seats, divide the bill by the working seats, and write that single number at the top of the renewal memo. Then compare the discount with the smaller order and decide which to defend.
He sets out the full first-person version, with the table, the three options and the Monday action, on his own site. In his own words: Before You Renew AI Licences, Count the Seats That Work.
Related evidence: Wikipedia defines shelfware as software that is paid for but goes unused, including money spent licensing more seats than are needed, which it notes cuts into the bottom line. (Wikipedia's entry on shelfware, software paid for and left unused)
Zylo's 2026 SaaS Management Index reports that, even as waste reduction improves, organizations continue to uncover unused seats, overlapping tools and feature tiers that add cost without clear value, and that embedded AI makes usage harder to predict. (Zylo's 2026 SaaS Management Index on unused software seats)