This sits under our 2026 guide to agentic AI ITSM pricing and negotiation. The guide covers strategy; this article is the apples-to-apples worksheet, because the most common AI pricing mistake we see is comparing two quotes that were never measuring the same thing.
Why can you not just compare rate cards?
A rate card is denominated in the vendor's chosen unit, and the unit is chosen to flatter the vendor. An assist, a session, a per-user add-on and a credit are not interchangeable, and the conversion between them depends on your traffic mix, not on a published ratio. A vendor with a low per-unit rate but a high mandatory minimum and list-priced overage can land well above a vendor whose rate looks higher but whose pool is right-sized and whose overage is capped. The headline number is the bait; the commitment and overage terms are the hook.
What dimensions do you benchmark on?
Build one row per vendor and one column per dimension below. Most rate cards give you only the first column, so the remaining columns are what you have to extract in the RFP and the demo.
| Dimension | What to capture | Why it moves the number |
|---|---|---|
| Metering unit | Assist, session, per-user add-on, credit, token | Sets the denominator you must convert from |
| Mandatory minimum | Smallest commitment the vendor will sell | A high floor inflates cost at low adoption |
| Overage rate | Price above the pool, and whether it is fixed or at list | List-priced overage is an uncapped liability |
| Annual cap | Whether a not-to-exceed ceiling is available | Caps convert a variable cost into a budgetable one |
| True-down rights | Whether you can reduce the commitment at renewal | Without it, an over-sized year one locks in |
| Cost per resolved action | Fully loaded annual cost divided by projected resolutions | The only figure that compares vendors like for like |
The contrast between metering models is the heart of it, and the trade-offs between them are set out in consumption vs per-seat ITSM AI pricing. Note as you fill the grid that consolidation is reshaping the field: ServiceNow has acquired Moveworks and Automation Anywhere has acquired Aisera, so some standalone AI vendors you benchmark today will be sold as part of a wider stack tomorrow, which changes both the leverage and the lock-in.
Book an AI cost control review.
We normalise every vendor's AI quote to cost per resolved action at your volume, then negotiate the pool, the overage and the cap. Fixed fee or gainshare.
Book an AI cost control review →What are the steps to run a clean AI benchmark?
- Fix the workload first. Define a single projected volume of actions per year, drawn from your own pilot data, and hold it constant across every vendor. The workload is the control variable; if it moves, the benchmark is worthless.
- Extract the full model, not the rate. In the RFP, require each vendor to state the minimum, the overage rate, the cap availability and the true-down terms in writing. A rate without these is incomplete by design.
- Convert to a common denominator. Apply each vendor's model to the fixed workload and compute fully loaded annual cost, then divide by projected resolved actions. Now the rows are comparable.
- Separate the AI line from the subscription. Benchmark the AI commitment on its own, so a vendor cannot bury an aggressive AI uplift inside an attractive platform price.
- Pressure-test at two volumes. Re-run the benchmark at half and double your projected volume. The vendor whose cost stays linear is safer than one who looks cheap at plan and expensive at scale.
Running the benchmark at multiple volumes is what exposes the vendors who price for the demo rather than for production. It is also where a credible alternative comes from: a normalised grid lets you tell your incumbent exactly what an equivalent workload costs elsewhere, which is the foundation of any AI cost negotiation.
Our gated 2026 ITSM Negotiation Benchmark Report includes the AI cost-per-action grid and the normalisation worksheet behind this method.
What will vendors argue against a like-for-like benchmark?
Each vendor will argue that their unit captures value the others miss, and that a straight cost comparison undervalues their model. Treat that as a reason to normalise harder, not to abandon the comparison. The value claim belongs in the qualitative scoring of the RFP; the commercial benchmark stays in cost per resolved action at fixed volume. Keep the two separate, because the moment you let a value narrative move the cost grid, the vendor is pricing your benchmark for you. Across $420M+ in ITSM contract value negotiated, the buyers who held the workload constant and made vendors compete on cost per outcome consistently paid less than those who compared rate cards.
Frequently asked questions
- Why is agentic AI pricing hard to benchmark?
- Each vendor meters in a different unit, so a rate card tells you little until you convert every quote to fully loaded annual cost per resolved action at your projected volume.
- What is the right unit to benchmark on?
- Cost per outcome at projected volume, not headline rate per unit. A low rate with a high minimum and list-priced overage can beat a higher rate with a fixed pool and a cap.
- Should I benchmark AI separately from the subscription?
- Yes. Break the AI commitment, overage and cap into their own line and benchmark them alone, so the uplift is negotiated rather than hidden in the platform total.
What to do next
Fix one workload, extract each vendor's full model, and convert everything to cost per resolved action before you score anything. When you want that grid built and the AI line negotiated rather than absorbed, it runs through our AI cost control service on fixed fee or gainshare.
Compare on cost, not claims.
We build the normalised benchmark and take the AI uplift apart, vendor by vendor.
Book an AI cost control review →