How to Run a 90 Day AI Usage Simulation

A 90 day AI usage simulation is a controlled pilot that produces the one thing every AI negotiation lacks: your own consumption data. Run it before you commit and you replace the vendor's forecast with measured fact, which is the difference between negotiating a cap you can defend and accepting an allotment you guessed at. The method is four phases across the quarter: scope a representative slice of work, instrument it so every token and action is measured, let it run long enough to capture real variation, then project the results to an annual figure and price both pricing models against it. Ninety days matters because it spans more than one work cycle, so weekly and monthly peaks show up rather than hiding. This how-to is part of our complete guide to ITSM AI pricing.

Phase 1 · Scope a representative slice

The simulation only tells you something if the slice you measure resembles the deployment you will buy. Pick a fulfiller group whose work mix mirrors the broader base: a blend of incident, request, and knowledge work, not just your power users and not just your quietest queue. Turn on the specific AI features you intend to license, no more, so the consumption you measure maps to the deal you will sign. Document the starting configuration, because the result is only valid for the settings you ran; a chattier configuration later will consume differently. Scope is where most simulations go wrong, by measuring a best case and then buying for the whole organization.

Phase 2 · Instrument everything

You cannot project what you did not measure. Before the pilot starts, confirm you can see consumption at the level the contract will bill: tokens, actions, or queries, broken down by feature and ideally by user. If the vendor's reporting is thin, that is itself a negotiating point, because a meter you cannot read is a meter you cannot cap. Capture the token-to-action relationship so you can translate the abstract meter into real work, the conversion explained in token based ITSM AI pricing explained. Instrumentation set up on day one is worth more than any amount of analysis after the fact.

Phase 3 · Let it run and watch the shape

Resist the urge to read results in week two. Adoption ramps, so early-week consumption understates the steady state, and month-end or project surges only appear if you wait for them. Over the 90 days, track not just the total but the shape: how concentrated consumption is among a few users, how much a single virtual agent contributes, how usage moves with the calendar. That distribution decides which pricing model fits, the analysis behind consumption vs per-seat ITSM AI pricing. A flat, broad curve and a spiky, concentrated curve lead to opposite contract choices even at the same total.

Phase 4 · Project to a year and price both models

Now turn 90 days of data into an annual figure. Scale the steady-state consumption, not the ramp-distorted average, then add a margin for the part of the base not in the pilot and for adoption growth over the term. With a defensible annual number in hand, price it under both a per-seat uplift and a consumption model, holding the usage constant, so you can see which is cheaper at your actual shape, the discipline in how to model Now Assist consumption before you commit. The projection also gives you the number to set your ceiling against, so the cap reflects measured reality rather than a round figure the vendor proposed.

Avoid the simulation pitfalls that distort the number

A simulation is only as honest as its design, and a few predictable pitfalls produce a number that flatters the vendor or yourself. The first is the enthusiasm spike: a pilot group told they are testing a new tool uses it more than a steady-state team will, inflating consumption early and tempting you to over-commit. The second is the opposite, a quiet pilot run during a slow season that understates the peaks you will hit at month-end or during a major project. The third is feature drift, where someone enables an extra capability midway through and the data no longer reflects the configuration you will buy. Control all three by documenting the configuration up front and freezing it, running across a normal-load period rather than a holiday lull, and separating the ramp weeks from the steady-state weeks when you annualize. Note any anomaly in the record so the projection carries its caveats; a number with honest error bars is more persuasive in a negotiation than a single confident figure the vendor can poke holes in.

Turn the result into contract terms

A simulation that does not change the contract was a waste of a quarter. Use the projection to set a consumption ceiling you can defend, to choose the pricing model your curve favors, and to justify the caps that protect you, the clauses in how to cap agentic AI consumption in ITSM contracts. On ServiceNow, feed the annualized figure into the Now Assist pricing conversation and align it with the protection terms in our ServiceNow pricing 2026 guide. The vendor argues from a forecast; you argue from your own measured data, and that asymmetry is the whole value of the exercise.

The bottom line

A 90 day AI usage simulation converts a guess into evidence: scope a real slice, instrument it fully, run it long enough to catch the peaks, and project to a defensible annual cost you price under both models. The buyer who walks into the negotiation with their own data sets the cap; the buyer without it accepts the vendor's. That single shift, from arguing over a forecast to arguing from evidence, is usually worth more than any discount you could win on the rate alone. Designing and running that simulation, then turning it into contract terms, is core to our buyer-side AI cost control work, fixed fee or gainshare, so we only win when you do.

Frequently asked questions