Vendors claim hundreds of AI agents inside the enterprise stack. This is the neutral ledger of the ones actually documented — what each does, whether it's shipped, who pays for it, and a primary source per row.
Segment = licensing status. The green share is the activation gap in its purest form: capability included in subscriptions customers already pay for.
Filter, search, sort. Click any source to verify — a row without a source doesn't make it into the ledger.
Start here if you own a process. Each cell counts the entries this ledger holds for that business function at that vendor. Click a function to filter the table below it.
Counts include both tiers. Vendors bucket processes differently — Oracle files most procurement agents under Supply Chain, SAP splits Spend from Finance — so functions are normalised here by rule to make columns comparable. That mapping is ours, not the vendors’. A zero means the vendor publishes nothing individually named for that function, not that the product does nothing.
| Vendor | Name | Layer | Function | What it does | Licensing | Status | Announced | Source |
|---|
Marketing counts what it wants to count. The ledger only counts what can be verified. The distance between those two numbers is itself a finding.
| Vendor | Public claim | Named & sourced in this ledger |
|---|
All six suites were re-checked against their live catalogues on 25 August 2026. No two count the same thing.
SAP publishes a Joule Assistant per domain — 50 across HCM, Finance, Spend, CX and Supply Chain — each coordinating agents underneath, with an Autonomous Domain Blueprint above as the commercial package. No flat agent catalogue, no per-agent GA dates. Oracle is the most auditable: 207 named agent listings plus 22 agentic applications, in separate public tables — though with no status or date column. Microsoft lists 148 rows for Dynamics 365, but only 16 are typed as “AI agent”; the other 132 are Copilot features, a different object entirely. Salesforce publishes agent templates to build from, with the rest on AgentExchange. ServiceNow claims “thousands” and names almost none. Workday names agents per release announcement, never as a catalogue.
So one SAP assistant ≈ one Oracle agentic application ≈ a bundle of Microsoft agents ≈ a Salesforce template you still have to build. Any single cross-vendor agent total compares unlike things. Every row here carries a layer tag — agent, assistant or agentic app — and the Assistants-only and Agentic-apps-only views isolate the container layer so you can compare like with like.
Correction, 25 August 2026. Three SAP rows previously listed as agents (HR Service, People Intelligence, Career and Talent Development) are assistants in SAP’s current taxonomy, with agents beneath them. Reclassified. Their earlier placement conflated two layers — the error this section exists to prevent.
Fully documented (148 entries). The original bar: a name, what it does, a release status, a licensing tag and a primary source — vendor newsroom, documentation or official product page. Every headline figure on this page is computed from this tier only.
Vendor catalogue (252 entries). Everything the vendor lists publicly but does not document to that bar. Oracle’s catalogue gives a name and a description with no status, release or date column; SAP’s domain pages give assistant names with agents nested underneath. These rows are included because a practitioner scanning for their process needs the full landscape — but they are tagged, and they are deliberately excluded from the percentages.
The two-tier split is itself the finding: the gap between what vendors name and what vendors document is where activation planning goes wrong.
An entry is a fully documented AI agent, assistant, or agentic application shipping inside one of six enterprise suites (SAP, Microsoft, Oracle, Salesforce, ServiceNow, Workday). Fully documented means all of: a name, what it does, a release status, a licensing tag, and a primary source — vendor newsroom, documentation, or official product page. Generic platform capabilities ("AI-powered search") are excluded unless the vendor names them as a distinct agent.
A name alone doesn't clear the bar. Oracle's own catalog page names 217 agents, advisors and agentic workspaces plus 28 agentic applications across both tiers — far more than its announcements document with a status and a date. Those names are tracked in Claimed vs. named below, and rows get promoted into the ledger as vendors document them. The distance between naming and documenting is itself one of the ledger's findings.
Licensing is recorded as: included (no incremental charge on the qualifying subscription), metered (consumes credits/units — SAP AI Units, Salesforce Flex Credits, Workday Flex Credits, Oracle AI Units), paid add-on (a separate per-user SKU), or unclear (vendor hasn't said publicly — which is a data point in itself). Oracle nuance: included covers the prebuilt agents and Agent Studio on Oracle-provided models with the qualifying Fusion subscription — premium-model usage beyond the default token allocation, marketplace/third-party agents, and custom agents require a paid subscription (Oracle pricing FAQ, 2026).
Status: GA, preview, or announced. Vendors rename and re-bundle constantly; the ledger records the state at the compile date and each release gets a version stamp.
What this isn't: a completeness claim. Vendors name only a fraction of what they count in marketing (see Claimed vs. named). Corrections and additions are welcome — every row is falsifiable by design.
v2.0 — 25 August 2026. Structural release, prompted by a reader question: does a flat agent list still describe what a customer gets? It did not. Vendor catalogues were harvested live and the dataset now carries a layer (agent / assistant / agentic app) and a tier (fully documented / vendor catalogue) on every row. +252 catalogue entries — Oracle’s published catalogue (187 agents and 20 agentic applications beyond the rows already documented) and 45 SAP Joule Assistants beyond the three already documented, across five domains. Total 400 entries. Three SAP rows reclassified from agent to assistant. Headline percentages are unchanged because they are still computed from the 148 fully documented rows only — included 38%, GA 54%.
Also new: a cross-vendor function matrix at the top, so someone who owns a process (procurement, payroll, close) can see what each vendor publishes for it. Functions are normalised by rule across vendors, because Oracle files most procurement agents under Supply Chain and SAP splits Spend from Finance; that mapping is ours and is stated on the page.
v1.1 — 9 August 2026. Vendor claims re-checked against primary sources, with an independent second-pass verification before publication. Changes: four rows added (Oracle’s supply-chain agentic applications announced 29 June 2026 — Inventory Planning Command Center, Supplier Qualification Workspace, Production Readiness Workspace, Kanban Administrative Workspace), taking the dataset to 148 rows. Headline shares recomputed: included 36% → 38%, GA 56% → 54%.
Two vendor-claim corrections in the same pass. Oracle’s aggregate moved from “more than 600 embedded agents” to “1,000-plus AI agents delivered through Fusion Applications” (Oracle, 14 July 2026) — and the earlier 600 figure was never “600 across Fusion”: Oracle’s own blog broke it down as ~400 in Fusion Cloud Applications plus ~200 in Industry Applications. Separately, Oracle began metering AI consumption in Fusion AI Units from release 26C: agents still carry no separate licence fee, but usage decrements units ($0.01 each on Oracle’s published price list, with 20,000 per production environment per month included per its service descriptions). Oracle rows remain tagged included because there is no per-agent SKU — but that tag now means “no licence fee, metered usage” for Oracle, and the licensing notes say so.
On SAP’s numbers. The widely-quoted “50 assistants / 200 agents” is future tense in SAP’s own Sapphire announcement (“will deploy… a subset of over 200”) and is a target, not an inventory. The larger figures reported from the Q2 2026 earnings call (~50 assistants by end of Q3, 400+ agents by year end) appear in third-party transcripts but not in SAP’s published quarterly statement, so they are cited as such. SAP publishes no count of assistants or agents generally available today — which is precisely why this ledger counts individually documented ones instead.
v1.0 — compiled 22 July 2026. 144 documented agents across six suites, each row carrying a primary source.
23 July 2026 — independent row-level verification. Before launch, every row was re-checked against every cited source (23 distinct primary URLs). Result: all 144 names found in their sources, zero dead links — and 17 corrections, applied in full:
That four vendor GA claims didn't survive a re-read of the vendors' own announcements is, itself, a finding. Corrections are welcome — every row is falsifiable by design, and the next version will credit them.