The discount on a credit you never use is not a saving
Service Cloud Einstein pricing is three questions stacked on top of each other: which edition you run, which AI capabilities are bundled into it, and what the consumption based autonomous layer adds on top. Most estates pay through at least two of the three, and the overlap between them is where the same capability gets bought twice without anyone noticing at renewal.
Prepared by Redress Compliance · August 10, 2026 · Salesforce advisory. Based on 15 to 25 Salesforce AI negotiations, 2024 to 2025.
Executive summary
AI line items were the least scrutinised part of the order form, quoted at full population and landing 25 to 45 percent lower. Add ons and credit commitments arrived priced against 100 percent of the user base or the projected volume, and negotiated deals settled well below that.
The gap exists because the AI line is new enough that nobody benchmarks it the way they benchmark seats, which makes it the softest number on the paper rather than the hardest.
Consumption commitments sized from vendor projections overshot actual first year usage by 2x or more in most deals.
First year consumption ran below half the committed volume, the unused credits expired worthless, and the committed floor then became the renewal baseline, which is the part that compounds.
Credits behave like seats did a decade ago: easy to buy, hard to shrink, and priced on projections only the vendor believes.
Edition against add on has no universal answer, and only modelling tells you which.
Higher editions bundle capability that costs extra below them, so an estate where most agents need AI usually models better on the higher edition, while an estate where a small team needs it models better on targeted add ons.
Step ups justified by bundled AI were cheaper in some estates and more expensive in others. Run both paths at your real population before accepting either quote.
List comparisons ignore overlap, and renewal is the moment to strip it out.
Estates routinely pay for capability bundled into the edition and again through legacy add ons nobody removed from the order form, which is invisible in any list price comparison because both lines look legitimate in isolation.
Inventory what the current edition already includes before paying for an add on that duplicates it.
The four cost paths, and where each one fails
| Path | Where it wins | Where it loses |
|---|---|---|
| Stay on edition, buy add ons | Small AI user population | Large populations stack the per user rate |
| Step up edition for bundling | Broad AI adoption planned | Paying the uplift for capability nobody uses |
| Consumption only | Deflection focused programmes | Unused prepaid credits expire as shelfware |
| Hybrid: bundle plus credits | Most large estates in practice | Complexity hides double payment for overlap |
The three pricing mechanisms stack rather than substitute, which is why the hybrid path is where most large estates actually land and where overlap hides.
Capability bundled into the edition, per user add ons for what the edition lacks, and metered consumption for autonomous work each meter differently and appear on the order form at different times, often added by different people across different renewal cycles.
That is how an estate ends up paying for case classification inside the edition it upgraded to and again through a legacy add on that predates the upgrade. The remedy is an inventory of what the current edition already bundles, taken before the next renewal rather than during it.
The wider AI packaging sits in the Agentforce pricing pillar.
Sizing the consumption commitment
- The only defensible sizing input is a measured pilot, meaning real deflection rates on your own case mix rather than the projection in the proposal deck.
- Commit to 60 to 70 percent of measured demand, and negotiate true up rates for the upside instead of prepaying it, because unused credits expire while the floor persists.
- Keep the commitment short, since packaging and rates have shifted repeatedly and a short term preserves your ability to reprice as the market moves.
- Watch the renewal baseline, because the committed floor becomes the starting point for the next negotiation whether or not the volume was ever consumed.
- Strip the legacy add ons that overlap bundled capability at the same renewal, since that is the one moment the order form is open and both lines are visible together.
The Agentforce licensing brief
The consumption model decoded, the deflection arithmetic that decides whether the economics work, and the commitment sizing that survives first year reality.
Get the white paper →What the autonomous layer changes about the economics
The consumption layer moves the unit of pricing from the human agent seat to the autonomous conversation, which makes deflection rate the single variable that decides whether the economics work at all.
That is a genuinely different question from the one seat based pricing asks, because a seat is a decision about how many people you employ while a conversation is a decision about how much work the system takes off them, and only the second can be measured before it is bought.
It also means the vendor proposal and the buyer position are built from different inputs: the proposal prices the AI future at an assumed deflection rate, and the pilot prices your actual deflection on your actual case mix. Buy the second and keep options on the first.
The standard advice runs the other way, arguing that you should commit big to credits early because committed rates beat on demand rates and AI adoption only grows.
In the negotiations we advised, first year consumption ran below half the committed volume in most deals, the unused credits expired worthless, and the committed floor became the renewal baseline.
So the discount was applied to capacity that was never used and the overcommitment then travelled forward into the next term.
Five levers move the price: edition mix modelling at your real population, removing the add on overlap, negotiating the credit rate rather than accepting the published one, sizing the commitment to measured rather than projected demand, and timing the close against the vendor fiscal year end.
The related Data Cloud consumption mechanics sit in the Data Cloud and Agentforce guide.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A negotiation playbook, talking points, and a two page executive brief on day one
What we saw across Salesforce AI pricing engagements, 2024 to 2025
Across roughly 15 to 25 Salesforce negotiations involving Einstein and autonomous agent components we advised between 2024 and 2025, the AI line items were the least scrutinised part of the order form:
Reduction achieved against the quoted AI line items, which arrive priced at full population or projected volume by default.
Typical overshoot of consumption commitments against actual first year usage, with the unused portion expiring worthless.
Three patterns recurred: AI add ons and credit commitments quoted at 100 percent of population or volume while negotiated deals landed 25 to 45 percent lower, consumption commitments sized from vendor projections overshooting actual first year usage by 2x or more.
And edition step ups justified by bundled AI proving cheaper than add on stacking in some estates and more expensive in others, with only modelling telling which.
The buyer side move is to pilot, measure deflection on your own case mix, commit to 60 to 70 percent of measured demand, and negotiate true up rates for the upside instead of prepaying it. The wider picture sits in the Salesforce practice.
Your first five moves
- Inventory which AI capabilities your current edition already bundles, because estates routinely pay for the same capability inside the edition and again through a legacy add on.
- Model the edition step up against add on stacking at your real user population, since neither path wins universally and only the modelling tells you which applies to you.
- Run a measured pilot before any credit commitment, because deflection rates on your own case mix are the only defensible sizing input available.
- Commit to 60 to 70 percent of measured consumption with negotiated true up rates for the upside, rather than prepaying capacity that expires if adoption lags.
- Strip the overlapping add ons at renewal and time the close against the vendor fiscal year end, which is the one moment both the order form and the discount range are open. The Salesforce practice runs the modelling with you.
Frequently asked questions
How is Service Cloud Einstein priced?
Through three stacked mechanisms: capabilities bundled into your edition, per user add ons for what the edition lacks, and metered consumption for autonomous agent work.
Most estates pay through at least two of the three, which is why the overlap between them is the most common source of duplicate payment on the order form.
Which edition makes Einstein cheapest?
There is no universal answer. An estate where most agents need AI capability usually models better on a higher edition with bundling, while an estate where a small team needs it models better on targeted add ons.
The mistake is stacking add ons without modelling the step up, or stepping up and paying an uplift for capability nobody uses.
Why are credits the new shelfware?
Because they behave like seats did a decade ago: easy to buy, hard to shrink, and priced on projections only the vendor believes.
In our file first year consumption ran below half the committed volume in most deals, the unused credits expired worthless, and the committed floor became the renewal baseline for the next term.
How should a consumption commitment be sized?
From a measured pilot rather than a proposal projection. Real deflection rates on your own case mix are the only defensible input.
Commit to roughly 60 to 70 percent of measured demand and negotiate true up rates for the upside, which converts the risk of overcommitment into a manageable rate question instead of an expiring prepayment.
What does the autonomous layer change?
It moves the unit of pricing from the human agent seat to the autonomous conversation, which makes deflection rate the variable that decides whether the economics work.
A seat is a decision about how many people you employ; a conversation is a decision about how much work the system removes, and only the second can be measured before it is bought.
Where do list price comparisons mislead?
They ignore overlap. Estates routinely pay for capability bundled into the edition and again through legacy add ons that nobody removed from the order form, and both lines look legitimate in isolation.
Renewal is the moment to strip the overlap out, because it is the only point at which the whole order form is open.
How much movement is available on AI line items?
In our engagements, negotiated deals landed 25 to 45 percent below the quoted price, with a median cut of about 35 percent.
The line arrives quoted at full population or projected volume by default, and it is the least benchmarked part of the order form, which makes it the softest number on the paper rather than the hardest.
