Full narration of the briefing. Click a section heading to jump the player to that moment.
If you take one number from this entire series, take this one: credits per resolved interaction. Not credits per user, not credits per month, and definitely not the vendor's benchmark for organisations like yours. Credits consumed, divided by pieces of work actually finished. That number, and not a vendor estimate, sizes everything after it: your pool, your overage exposure, and whether the AI is worth having at all.
And you cannot get it from a spreadsheet. You get it by running something small and measuring it, which is what Daniel and I are going to walk through.
The pilot shape that consistently produced usable economics is tight: fifty to a hundred and fifty fulfillers, across two or three product areas, validated inside ninety days. That is it. Broad estate rollouts that skipped the pilot rarely showed credible per fulfiller return before the first renewal, which means they arrived at the negotiation with no evidence and had to accept the vendor's model. And be honest about what the pilot is for.
It is not a technology proof; the technology works. It is a meter reading. You are buying a number you can put on the table, and ninety days is what it costs to own that number rather than borrow theirs.
Then build the forecast from four inputs, because any one of them alone will mislead you. One: prompts per active user, from your pilot data, multiplied by realistic adoption rather than licensed seats. Two: agent tasks multiplied by the action weight, which is where twenty five, fifty or a hundred and fifty per action bites. Three: the non production load, because dev and test draw on the same pool.
And four: the ramp, because year one is not year three. Miss the second and you understate agents. Miss the third and you get the surprise we described last session. Most vendor models contain only the first.
Now the sizing rule, and it is deliberately conservative. Size the pool at the seventy fifth percentile of your trailing ninety day pilot consumption. Not the peak, which buys you shelfware, and not the average, which guarantees overage. And treat the pool as a usage envelope rather than a bet on take up.
The reason conservative sizing is safe is rollover: if unused capacity carries forward, under sizing costs you very little, and over sizing costs you every year of the term. Which is why we always negotiate rollover before we argue about pool size. Get the carry forward right and the sizing decision stops being frightening.
So what is this worth commercially? Buyers who measured one workflow first cut their committed pool by twenty five to forty percent. Separately, first year credit pools were over committed against measured burn by twenty to fifty percent, and on broad rollouts the median year one pool went forty two percent unused. In seven out of ten Now Assist commitments the buyer finished year one with thirty five to fifty percent of the pool unused.
Those are not small misses. And a discount on credits you never consume is not a saving, it is a discounted purchase of nothing, which then becomes the baseline for your next quote.
Then roll out in that order: one workflow, then a two product pilot, then breadth. Broad enablement burned two to three times the credits of a tuned configuration with no extra productivity, so going wide early costs you twice, once in credits and once in the baseline it sets. Put brakes on each wave: a credit ceiling with a named owner, weekly consumption reporting, and throttling logic that redirects to non AI workflow steps when a wave passes its budget rather than silently spending. And target roughly fifty to sixty percent of the licensed population in year one.
That is the phasing that avoids paying for shelf AI.
Here is the move. Start the ninety day pilot now, before the quote arrives, because evidence introduced after a quote reads as a discount request and moves far less than evidence that shaped the quote in the first place. Instrument it for one number, credits per resolved interaction, and split it by capability so you know which workflows are cheap and which are expensive. Then take the seventy fifth percentile figure into the room.
You are not asking them to believe you. You are showing them your own meter, and a meter is very hard to argue with. Next time, Daniel and Claire take the pitfalls, and the doors that only open one way.
Redress Compliance works on contingency: our fee is 25 percent of what we save you. Nothing saved, nothing paid. Independent, buyer side only, never vendor funded. Want Redress to contact you? Reach out and we respond the same day.
Talk to a ServiceNow negotiator