SAP, Oracle, Microsoft, Workday, ServiceNow, AWS, and Google each invented an AI credit currency in 2026. This pillar normalizes them into one table and the buyer side levers before your next renewal.
Every major enterprise vendor invented an AI credit currency in 2026. SAP AI Units, Oracle AI Units, Microsoft Copilot Credits and ACUs, Workday Flex Credits, ServiceNow Assist consumption, AWS Bedrock AgentCore, and Google Agent Engine all meter differently. This pillar normalizes them into one comparison and the buyer side levers for each.
The shift is structural, not cosmetic. From 2015 to 2024 enterprise software converged on per seat subscription pricing. In 2025 and 2026 the same vendors pivoted to metered AI currencies because agent workloads do not map to seats. An agent is not a person, so a seat cannot price it.
That leaves procurement with seven new currencies to govern at once. The rest of this pillar puts them on one table, shows which model shifts risk to whom, and sets out the levers that work across all of them.
Every vendor invented a credit currency because agent workloads break the seat model. A seat prices a human user. An autonomous agent runs on its own schedule and can fan a single request into hundreds of model calls, so the vendor needs a unit that scales with work, not with headcount.
Interactive assistants still fit a seat. A person opens the tool, asks, and reads the answer. Agentic features do not. They run in the background, chain tool calls, and consume compute whether or not a person is watching. The credit is the vendor answer to that mismatch.
Large language model inference has a real marginal cost. A flat seat that includes unlimited AI exposes the vendor to unbounded compute. Metering with credits caps that exposure and moves the variable cost onto the buyer. That is the commercial logic behind the whole shift, and buyers should read it plainly.
The vendor documentation confirms it. Oracle sets out its AI Unit model on the Oracle Fusion AI pages, and Microsoft describes Copilot capacity in the Copilot Studio capacity documentation.
On a common basis the currencies split into three families: near one cent credits, per action metering, and raw compute per vCPU hour. The table below normalizes each vendor to unit price, included allowance, overage, rollover, agentic multiplier, and the default renewal trigger date.
Enterprise AI credit currencies compared, 2026
| Vendor currency | Unit price | Included allowance | Overage | Rollover | Agentic multiplier | Default renewal trigger |
|---|---|---|---|---|---|---|
| SAP AI Units | List undisclosed, roughly 0.08 to 0.18 per action observed | 200 actions per Advanced FUE; Business AI Base in RISE | Metered per action | Often none, contract specific | 5x to 10x an interactive prompt | Cloud renewals from July 2026 |
| Oracle AI Units | 0.01 per AI Unit | 20,000 AI Units per month free per Fusion customer | 100,000 packs at 1,000 | Yes, packs roll over | Driven by model choice, not agent type | Fusion 26C from July 2026 |
| Microsoft Copilot Credits and ACUs | 0.01 per Copilot Credit; ACUs pre purchased | Varies by Copilot plan | Pay as you go or pre purchase | Pre purchased capacity typically expires | Agent actions consume more credits | Agent Store and Foundry billing from 22 April 2026 |
| Workday Flex Credits | Consumption, unit undisclosed | Included in every subscription | Negotiated at contract | Contract specific | Illuminate and Sana agents draw down | Aligned to subscription renewal |
| ServiceNow Assist consumption | Consumption on top of tier price | Now Assist bundled in every tier | Stacks on the tier price | Contract specific | Autonomous agents on Prime tier only | Tier repackaging from April 2026 |
| AWS Bedrock AgentCore | 0.0895 per vCPU hour; 0.00945 per GB hour | None, pure usage | Metered hourly | None, pure usage | Runtime scales with agent compute | On demand, no renewal lock |
| Google Agent Engine | 0.0864 per vCPU hour | None, pure usage | Metered hourly | None, pure usage | Gateway billing adds per call | Gateway billing effective 13 July 2026 |
The near one cent credit family covers Oracle and Microsoft, where the headline unit is small but agent actions multiply it. The per action family covers SAP and ServiceNow, where the unit is coarse and the multiplier hides in agentic runs. The compute family covers AWS and Google, where you pay for vCPU and memory time regardless of business outcome.
A one cent credit is not cheaper than an eighteen cent action until you know how many of each a workload consumes. Normalize every vendor to cost per completed business action, then the picture inverts for several vendors. AWS publishes its rates on the Bedrock AgentCore pages and Google publishes compute rates on the Vertex AI pricing pages.
Pure usage models shift the most risk to the buyer, and bundled allowance models keep more risk with the vendor. The dividing line is whether an unexpected spike in agent traffic converts directly into an invoice or is absorbed by an included allowance.
AWS Bedrock AgentCore and Google Agent Engine are pure usage with no allowance and no rollover. Every vCPU hour bills. A runaway agent loop is a runaway bill. The buyer carries all volatility unless a committed floor and an alert threshold are written into the deal.
Workday Flex Credits and the ServiceNow bundled tiers include allowance inside the subscription, so moderate spikes are absorbed. The trade is a higher fixed price. The buyer pays a premium for predictability, which is rational for a workload that is hard to forecast.
The standard account team pitch is that the per credit unit price is low, often near one cent, so buyers should not worry about consumption and should simply enable the AI features. We disagree. In roughly 30 to 40 negotiations we ran or benchmarked, the unit price was the least important number and the agentic multiplier was the most important, because one autonomous run consumed five to ten interactive prompts of credits and the vendor supplied first year estimate ran well below actual burn. The buyer side move is to ignore the headline unit, model the annualized burn at the agentic multiplier, and negotiate a committed floor, rollover, and a hard alert threshold before signing. That is not what the account team frames, and it is not how a per seat comparison would read.
Source: Redress Compliance advisory engagement file, 2025 to 2026.
A credit is not a price. It is a meter the vendor controls. The buyer job is to govern the meter, cap the cliff, and normalize seven currencies into one number before the renewal call.
The overage cliffs differ most in steepness and warning. Some vendors bill the moment allowance ends, some stack overage on top of a fixed tier, and some have no allowance at all so there is no cliff, only a continuous meter.
ServiceNow stacks Now Assist consumption on top of the Foundation, Advanced, or Prime tier price. The buyer pays the tier and then the consumption, so the cliff is a second bill layered on the first. Autonomous agents are gated to the Prime tier, which raises both the fixed and variable cost.
SAP bundles 200 actions per Advanced Full Use Equivalent and meters beyond that. The cliff is per action, and agentic runs cross it fast because one Joule Agent run consumes far more than one interactive Joule prompt. SAP documents the model on the SAP Business AI pages.
AWS and Google have no allowance, so there is no cliff. Instead the meter runs continuously from the first vCPU hour. That is easier to reason about but offers no free buffer, so a spend cap and alerting matter more here than anywhere else.
A multi vendor estate should standardize the forecast method, the governance gate, and the renewal calendar, not the vendor currencies themselves. You cannot make the currencies identical, but you can make your control of them identical.
Convert every vendor to cost per completed business action and forecast annualized burn at the agentic multiplier, not the interactive rate. This is the single number that lets you compare renewals and spot the vendor whose currency is quietly the most expensive.
Route every new agent through one approval gate that records the expected credit burn before it goes live. Workday frames this at the platform level with an agent registry, described on the Workday AI pages. A buyer should mirror that governance across all vendors, not just the one that ships it.
Because the 2026 change dates cluster around April and July, several renewals collide in one budget year. Put every AI credit trigger date on one calendar so you enter each negotiation with the others in view, not one at a time.
The cross vendor levers that work are the ones that apply regardless of currency: a normalized baseline, a committed floor traded for rate, rollover, and a hard alert threshold. Pull them in the same order at every vendor.
Never open a credit negotiation without a normalized consumption baseline. It is the defensible number that stops each account team from framing its own currency as cheap. ServiceNow describes its agent model on the ServiceNow AI Agents pages, and reading each vendor own definitions is part of building that baseline.
Vendors discount the per unit rate in exchange for a committed spend floor. That is a fair trade only when your baseline shows the floor is below realistic burn. Commit to what you will genuinely use, not to the vendor optimistic estimate.
Rollover, an overage cap, and an alert threshold are negotiable on committed deals and worthless if left to the renewal call. Put them in the order form. The buyer side calendar for an AI renewal starts at 180 days out, not at 30.
An enterprise AI credit is a prepaid or metered unit that a software vendor charges when its AI features run, instead of a flat per seat fee. Each vendor has invented its own currency in 2026, so a credit at one vendor is not comparable to a credit at another without normalizing to the underlying action or compute hour.
They do not compare cleanly, which is the core buyer problem. Oracle and Microsoft price a credit near 0.01 US dollars, SAP meters per action near 0.08 to 0.18 US dollars, and AWS and Google price compute near 0.086 to 0.0895 US dollars per vCPU hour. The pillar table normalizes each vendor to a common basis.
Pure consumption models with no rollover and no cap shift the most risk to the buyer, because a spike in agent traffic converts directly into an unbudgeted invoice. AWS Bedrock AgentCore and Google Agent Engine are pure usage with no included allowance, so the buyer carries all volatility unless a committed spend floor and an alert threshold are negotiated.
An overage cliff is the point where included AI allowance runs out and every additional action bills at an uncapped rate. Agentic workloads make the cliff steep because one autonomous agent run can consume five to ten times the credits of a single interactive prompt, so consumption climbs faster than a seat based forecast predicts.
It depends on the vendor. Oracle Fusion AI Unit packs roll over, Microsoft pre purchased capacity typically expires, and AWS and Google bill pure usage with nothing to roll over. Rollover terms are negotiable on committed deals, and pinning them in the order form is one of the highest value levers in the table.
The key dates cluster in 2026. SAP makes use based AI the default for cloud renewals from July 2026, Microsoft began billing Agent Store and Foundry hosted agents from 22 April 2026, Oracle Fusion 26C lands around July 2026, ServiceNow retired its legacy tiers in April 2026, and Google Agent Engine Gateway billing takes effect 13 July 2026.
Yes. A buyer with three or more AI credit currencies should run one governance model, not one per vendor. Standardize the forecast method, the alert thresholds, the approval gate for new agents, and the renewal calendar, so a spike in one vendor does not hide inside a different budget line.
Build one normalized consumption baseline across every vendor before any renewal conversation. The single most common reason buyers overpay on AI credits is negotiating each vendor in isolation without a common unit of measure, which lets every account team frame its own currency as cheap.
The cross vendor comparison, the normalized burn model, the overage cliff math, and the buyer side levers for every AI credit currency in 2026.
Used across more than five hundred enterprise engagements. Independent. Buyer side. Built for procurement and IT leaders running the next AI renewal cycle.
Benchmarks, renewal calendars, and negotiation levers across the enterprise software stack. No vendor spin. Unsubscribe anytime.
