SAP and Salesforce: the ledger, the relationship, and why neither one leaves

BLUF: SAP and Salesforce sit on opposite sides of the enterprise — SAP is the financial system-of-record (the ledger), Salesforce is the system-of-engagement (the relationship). They are converging on the “quote-to-cash” middle, but the deeper story is stickiness: both are complained about constantly and churned out of almost never, because leaving is a multi-year, nine-figure operational risk with a public failure record. This brief maps six dimensions — competencies, overlap, deployment, integration, APIs/lock-in, and AI — and closes on the mechanism of entrenchment and what could break it.

Worked example throughout — NVIDIA. NVIDIA is the collision made concrete: SAP (SD module, mid-migration ECC→S/4HANA) is the financial transaction core, while Salesforce (SFDC) occupies the customer-facing order/case-management slot SAP SD would otherwise fill [[2026-07-06-nvidia-rma-acronym-glossary|NVIDIA RMA glossary]]. Even a cash-rich, engineering-elite company wraps rather than replaces — layering custom tooling and a data lake (“ODP”) on top instead of ripping either incumbent out. That choice is the thesis of §7 in miniature.

Anchor sources. The motivating internal thread is the NVIDIA engagement ([[2026-06-17-logistics-catch-up-wgreg-lonny-nvidia|Greg + Lonny @ NVIDIA]]; [[nvidia-primer-2026-06-18|NVIDIA primer]]), and the strongest prior synthesis is [[sap-in-semiconductor-supply-chain-2026-05-27|SAP in the semiconductor supply chain]]. Interview grounding comes from [[2026-04-29-call-holly-rawlins-re-renesas|Holly Rawlins (Renesas)]], [[2026-05-02-zoom-dustin-ross-skiffra|Andy (Skiffra)]], and [[2026-05-06-stramgt-3862-athey|Susan Athey]].

Outline changes: none. Executed as the approved 7-section plan (§7 “why so sticky” was agreed during planning, not added mid-run). Every section did independent web research; citations use inline confidence labels with a consolidated Sources list at the end.

This is a landscape map, not a thesis verdict. Per RDI, it surfaces evidence and ends in open questions. It does not draw a TBD-strategy conclusion — that’s the founders’ call.

1. Core competencies & value propositions

SAP and Salesforce sit on opposite sides of the enterprise: the ledger vs. the relationship. SAP is the financial system-of-record; Salesforce is the system-of-engagement. In NVIDIA’s stack, this split is literal — SAP (SD module, mid-migration ECC→S/4HANA) books the money, while Salesforce (SFDC) runs the customer-facing order and case-management slot that SAP’s SD module would otherwise occupy [Synthesis].

Origin-story moats

SAP’s moat is the closed general ledger. Founded 1972 in Walldorf, SAP built its franchise on the ERP core — finance, controlling, materials management, production planning — where every transaction reconciles to an auditable financial record. Once a company’s chart of accounts, month-end close, and statutory reporting run through SAP, ripping it out means re-plumbing the balance sheet. That switching cost, not features, is the moat. Holly Rawlins (Renesas) described SAP as “the unavoidable financial core” [Interview: Holly Rawlins, 2026-04-29]. Andy at Skiffra sharpened the boundary of that competency: “SAP is great at reporting, bad at predicting” [Interview: Andy/Skiffra, 2026-05-01] — strong on the backward-looking record, weaker on forward-looking engagement.

Salesforce’s moat is the multi-tenant cloud CRM it invented. Founded 1999 with a “No Software” banner, Salesforce pioneered SaaS-delivered CRM and rode that into the #1 CRM position — IDC has ranked it the world’s #1 CRM provider for 12 consecutive years, at roughly 20.7% worldwide CRM share in 2024 [Public: cirrusinsight.com / IDC, 2025]. Its moat is the sales/service rep’s daily workflow plus a developer/AppExchange ecosystem, not the financial ledger.

Scale and financials

Dimension	SAP	Salesforce
Latest FY revenue	€36.8B FY2025, +8% (+11% cc) `[Public: SAP FY2025]`	$37.9B FY2025 (ended Jan 2025), +9% `[Public: Salesforce FY25 Q4 release, 2025-02-26]`
Cloud revenue	€21.02B, +23% `[Public: SAP FY2025]`	Subscription & support $35.7B, +10% `[Public: Salesforce FY25 Q4 release, 2025-02-26]`
Cloud backlog	Total cloud backlog €63.29B, +43%; current €18.08B, +32% (Q4’24) `[Public: SAP FY2024 6-K, 2025-01]`	Data Cloud & AI ARR $900M, +120% YoY `[Public: Salesforce FY25 Q4 release, 2025-02-26]`
Customers	~425,000; ~30,000 still on legacy ECC on-prem `[Public: SAP fact sheets/Constellation, 2026]`	#1 CRM 12 yrs running (IDC) `[Public: cirrusinsight.com/IDC, 2025]`
Market cap (context)	~$189–332B across 2025–2026 (Europe’s most valuable at ~$332B, Mar 2025) `[Public: CNBC, 2025-03-25]`	~$252.6B (Jul 2026) `[Public: companiesmarketcap, 2026]`

Worth flagging: SAP’s ~€21B cloud revenue trails Salesforce’s ~$35.7B subscription revenue, yet SAP’s cloud compounds faster (+23% vs. +10%) off a large migration backlog [Synthesis]. Market-cap leadership between the two has flipped multiple times across 2024–2026 [Public: Cloud Wars, 2025; companiesmarketcap, 2026].

Product portfolio breadth

SAP: ERP core (S/4HANA), delivered increasingly via RISE with SAP — repackaged in 2025 into a modular per-SKU menu, private edition rebranded SAP Cloud ERP Private [Public: panaya.com / ciodive.com, 2025]. Five line-of-business bundles (finance, supply chain, HR, procurement, CX) [Public: CIO Dive/Sapphire, 2025]. Front office via SAP Customer Experience (CX): Sales, Service, Marketing, Commerce, Emarsys, CDP [Public: sam-solutions.com, 2025]. Data layer via SAP Business Data Cloud (Datasphere + embedded Databricks + SAP Analytics Cloud + Knowledge Graph) [Public: SAP/LeverX, 2025].
Salesforce: Sales Cloud, Service Cloud, Platform+Slack, Marketing & Commerce, and Integration & Analytics (MuleSoft/Tableau), now wrapped under Agentforce 360 branding [Public: demandsage.com, 2026]. Plus Revenue Cloud (quote-to-cash), Data Cloud / Data 360, and Salesforce Industries (ex-Vlocity vertical clouds) [Public: Noltic, 2026; salesforce.com, 2020/2025].

One-sentence pitch to a CIO

SAP: “Run your entire enterprise — finance, supply chain, manufacturing — on one auditable cloud ERP system-of-record” [Synthesis].
Salesforce: “Unify every customer touchpoint on one AI-agent-native platform your teams already live in” [Synthesis].

What each is genuinely best-in-class at

SAP: the financial general ledger and deeply integrated manufacturing/supply-chain execution (MM, PP, SD, ATP, statutory close) — the transactional record NVIDIA’s SD module holds [Synthesis; Interview: Holly Rawlins, 2026-04-29].
Salesforce: the sales/service rep’s engagement workflow and the surrounding developer/ISV ecosystem — best at making customer context actionable “today,” the OMS/case slot SFDC fills for NVIDIA [Public: zennify.com, 2025; Synthesis].

2. Overlapping capability & head-to-head

Both companies are converging on the “quote-to-cash” middle — SAP pushing up from the ledger into CX, Salesforce pushing down from CRM into ERP-adjacent billing and data. NVIDIA is the collision made concrete: SAP SD and SFDC both plausibly own order and case management, and NVIDIA resolved the overlap by keeping the financial record in SAP and the customer-facing order/case slot in Salesforce [Synthesis].

Where they collide

Order management / OMS. SAP order management (capture, ATP, fulfillment, returns) runs inside S/4HANA and SAP Commerce Cloud with deep finance/logistics integration; Salesforce OMS is API-driven, Einstein-routed, pre-wired to Commerce + Service Cloud [Public: ranosys.com; fynd.com, 2025–2026].
CPQ / quote-to-cash. SAP entered via the $2.4B CallidusCloud acquisition (announced Jan 29, 2018, closed Apr 5, 2018) — CPQ + sales performance management; CallidusCloud did $253.1M FY2017 revenue [Public: TechCrunch / SEC 8-K, 2018]. Salesforce is mid-transition: legacy Salesforce CPQ went end-of-sale to net-new customers in March 2025, investment redirected to Revenue Cloud (reintroduced as Agentforce Revenue Management at Dreamforce 2025) [Public: clonepartner.com / Noltic, 2025–2026].
Commerce. SAP Commerce Cloud (headless, B2B+B2C) vs. Salesforce Commerce Cloud (SaaS, D2C-leaning) [Public: TrustRadius, 2025].
Service. SAP Service Cloud vs. Salesforce Service Cloud — the latter one of Salesforce’s largest historical revenue engines [Public: TrustRadius, 2025].
Analytics / Data Cloud. SAP Business Data Cloud (Datasphere + embedded, SAP-managed Databricks + Knowledge Graph) vs. Salesforce Data 360 (real-time unified profiles; Databricks added Unity Catalog zero-copy access in 2025) [Public: SAP/Databricks blog; customerthink.com, 2025].
Industry clouds. SAP’s LoB/industry bundles vs. Salesforce Industries (ex-Vlocity, 2020; OmniStudio) [Public: salesforce.com, 2020; getoncrm.com, 2025].

Who tends to win which battle, and why

CRM head-to-head: Salesforce wins decisively. ~20.7% worldwide CRM share, #1 for 12 straight years (IDC); SAP sits in the top 10 at ~3.1% overall, SAP Sales Cloud ~0.61% in the CRM-platforms category [Public: cirrusinsight.com/IDC; 6sense.com, 2025]. SAP’s CX push “up” has not dented Salesforce’s core [Synthesis].
Financial system-of-record / ERP: SAP wins by default. Salesforce’s Revenue Cloud reaches billing but explicitly requires customers to “define what lives in Salesforce versus ERP” — it complements, not replaces, the ledger [Public: getmonetizely.com, 2025]. NVIDIA’s split reflects exactly this boundary [Synthesis].
The contested middle (OMS, CPQ, commerce, data): tends to be won by platform gravity, not features — the incumbent platform the customer already stands on takes the adjacent module. For NVIDIA, that gravity split the middle between both vendors rather than consolidating it [Synthesis].

Comparison table

Dimension	SAP	Salesforce	Who tends to win / note
Core identity	ERP / financial system-of-record	CRM / system-of-engagement	Different jobs; NVIDIA runs both `[Synthesis]`
Financial system-of-record	Native (S/4HANA GL)	Reaches billing but defers to ERP	SAP
CRM (sales/service)	Sales/Service Cloud, ~0.61% share	#1, ~20.7%, 12 yrs	Salesforce (decisive)
CPQ / quote-to-cash	CallidusCloud ($2.4B, 2018)	Revenue Cloud; legacy CPQ EOL to net-new Mar 2025	Contested; platform gravity `[Synthesis]`
Order management	S/4HANA + Commerce; ATP, returns	API-driven, Einstein routing	Incumbent-platform wins; split at NVIDIA `[Synthesis]`
Data / analytics	Business Data Cloud (Datasphere + embedded Databricks)	Data 360; +Databricks zero-copy 2025; ARR $900M +120%	Different centers of gravity; shared Databricks partner `[Synthesis]`
Direction of travel	Moving up into CX	Moving down into billing/revenue/data	Head-on collision in the middle `[Synthesis]`

Open question (not a conclusion): where NVIDIA drew its SAP-vs-SFDC boundary — the SD-module order/case slot going to Salesforce despite SAP owning the financial record — is the practical answer to “who wins the middle” for one large customer mid-migration. Worth testing whether that split is idiosyncratic to NVIDIA’s migration state or a repeatable pattern [Speculation].

3. Deployment model & IT burden: SAP forces re-implementation choices; Salesforce forces continuous administration

BLUF: SAP and Salesforce push maintenance burden onto a customer’s IT department in fundamentally different shapes. SAP’s burden is episodic and heavy — a migration and a custom-code reckoning tied to a 2027/2030 deadline. Salesforce’s is continuous and lighter-per-event — three forced upgrades a year absorbed mostly by a declarative “admin” role. NVIDIA sits inside both burdens at once: mid-migration ECC→S/4HANA on the SAP side, and a large SFDC estate on the customer-facing side.

SAP: on-prem ECC → RISE/GROW, and the “clean core” doctrine

On-prem ECC 6 — the classic self-hosted ERP; NVIDIA’s current financial-core state [Interview: NVIDIA RMA glossary, 2026-07-06]. Mainstream maintenance for ECC 6 ends Dec 31 2027; extended maintenance runs to Dec 31 2030 at roughly +2 percentage points (~22% → ~24%, a ~9% uplift on the existing fee) [Public: Rimini Street / SAP Licensing Experts, 2025].
RISE with SAP — S/4HANA in a private cloud, single-tenant, SAP-managed on a customer-chosen hyperscaler; rebranded SAP Cloud ERP Private in 2025. Private-cloud transformation timelines run 6–18 months [Public: SAP / Panaya, 2025].
GROW with SAP — S/4HANA public cloud, true multi-tenant SaaS, always greenfield; go-lives cited at 8–12 weeks using pre-configured accelerators [Public: SAP-Press / ERP Software Blog, 2025].
Clean core doctrine — SAP’s 2025 stance: custom logic must live outside the S/4HANA core (BTP side-by-side, Key User Extensibility, ABAP Cloud). SAP now publishes a four-level extension maturity model (A–D) and a Custom Code Migration Advisor [Public: SAP News Center, Aug 2025].

The economic weight is in the custom-code reckoning: ~60% of custom code needs changes during an S/4HANA conversion, especially code touching the Universal Journal [Public: Tachyon Technologies, 2026]. Clean-core approaches remove 30–50% of custom code from the migration critical path; skipping the audit routinely produces 30–50% cost overruns, and >60% of SAP migrations exceed budget, timeline, or both [Public: SoftwareSeni / Tachyon, 2025–2026]. Consistent with our internal synthesis figure that custom-code remediation is 45–60% of total migration spend [Public synthesis, 2026-05].

What forces a re-implementation vs. an upgrade (SAP): a greenfield GROW deployment or a large brownfield-to-clean-core conversion is a re-implementation; staying on RISE/S/4HANA and applying release upgrades is an upgrade. The trigger toward re-implementation is accumulated custom ABAP that violates clean core [Synthesis]. NVIDIA’s SAP estate is visibly Z-heavy (ZRMA, ZMFG, ZMRR, etc.) [Interview: NVIDIA RMA glossary, 2026-07-06], exactly the profile that raises remediation load [Synthesis].

Internal evidence on switching cost: Holly Rawlins framed displacing SAP at a semiconductor IDM as “hundreds of millions of dollars and years” [Interview: Holly Rawlins, 2026-04-29]. Renesas’s custom order-entry tool “Rainbow” persisted through years of SAP migration planning because SAP-native order entry never reached parity [Interview: Holly Rawlins, 2026-04-29].

Salesforce: multi-tenant SaaS, three forced releases a year, an admin-centric change model

Deployment — single multi-tenant SaaS; customers get orgs and sandboxes, not infrastructure [Public: Salesforce Help, 2025].
Release cadence — three seasonal releases/year (Spring/Summer/Winter); sandbox preview instances upgrade 4–5 weeks ahead. Upgrades are automatic and non-optional — the burden is regression-testing customizations, not scheduling [Public: Salesforce Help, 2025].
Declarative vs. code — Salesforce is “Flow-first” (Workflow Rules / Process Builder retired). Admins build automation in Flow with no developer; drop to Apex only when Flow can’t do it [Public: Nandann / Salesforce Admins, 2025–2026].

Who bears the burden (Salesforce): a dedicated Salesforce admin role owns config, Flows, release regression, user management. Much changes without code (objects, fields, layouts, validation, most automation); code (Apex/LWC) is reserved for complex logic and external callouts [Public: Salesforce Ben, 2025]. A re-implementation is typically triggered by data-model debt or org sprawl, not the release cadence [Synthesis]. NVIDIA’s stated SFDC pain — keeping installed-base/asset records accurate — is org-data-quality debt, not a version problem [Interview: NVIDIA RMA glossary, 2026-07-06].

NVIDIA as worked example

NVIDIA carries both burden shapes at once: an ECC→S/4HANA migration in progress, with $2M+ committed to SAP to automate the planning layer, plus EWM/TM rollouts being pushed into 3PL partners [Interview: NVIDIA RMA glossary, 2026-07-06; NVIDIA primer, 2026-06-18]. SFDC sits in the customer-facing slot “SAP SD would otherwise fill.” Neither system alone resolves Greg’s complaint that everyone isn’t looking at the same data — SAP holds the financial transaction, SFDC holds the case, and the reconciled operational view lives nowhere native to either [Interview: Greg + Lonny, 2026-06-17].

4. Integrations between SAP and Salesforce: point-to-point is the norm, “zero-copy” is the emerging pattern, and the common layer still lives outside both

BLUF: There is no single blessed SAP↔Salesforce integration; the canonical pattern is middleware-mediated API sync (MuleSoft or SAP Integration Suite, or a third-party iPaaS), poll-based and API-limited. A newer pattern — zero-copy data federation — lets a lake query each system without moving records, but as of mid-2026 there is no announced direct Salesforce Data Cloud ↔ SAP Datasphere link; both federate to third-party lakes (Snowflake/Databricks/BigQuery) as the meeting point.

First-party integration

MuleSoft (Salesforce-owned) — SAP connectors/templates; structural first-party advantage inside the Salesforce ecosystem. MuleSoft now markets Agentforce↔S/4HANA context hand-off [Public: MuleSoft, 2025].
SAP Integration Suite (on SAP BTP) — SAP’s own iPaaS; connects Salesforce and SAP “without custom code,” aligned to the clean-core doctrine [Public: SAP Community / Skyvia, 2025].
Partnership history — a caution, not a co-product: SAP and Salesforce have no durable joint product. The clearest fragility data point: SAP CPQ’s native Salesforce integration was retired (SAP CPQ release 2308, ~2023) after Salesforce retired REST API versions 21–30, forcing a partner-based integration [Public: SAP KB #3253678, 2022–2023]. The “first-party” path is two vendors’ middleware meeting in the middle, not a co-engineered connector [Synthesis].

Third-party iPaaS

General-purpose iPaaS used for SAP↔SFDC sync: Boomi, Informatica Cloud, Workato ($25K–$300K/yr; 6K+ connectors), SnapLogic, Jitterbit, Celigo, DBSync [Public: OneIO / Celigo / Houseblend, 2025–2026]. Common use cases: order creation, master-data sync, inventory lookups, invoice automation [Public: Skyvia, 2025].

What breaks (point-to-point failure modes): these platforms run on external infrastructure, call Salesforce REST/SOAP APIs from outside, count against the subscriber’s daily API limit, and run on 5–15 minute poll cycles [Public: AppNigma / Houseblend, 2026]. Practical failure modes: (a) API-limit exhaustion, (b) latency/staleness from polling, (c) versioned-API breakage of the kind that killed the SAP CPQ connector. Master-data conflicts (which system owns the customer/asset record) are a semantic failure middleware moves around but doesn’t resolve [Synthesis].

Salesforce side: the Zero Copy Partner Network (announced Apr 25 2024) federates Data Cloud/Data 360 with Snowflake, Databricks, AWS, Google Cloud, Microsoft via metadata + query pushdown — no record movement, including Apache Iceberg support. Salesforce cites ~4 trillion records queried without moving data [Public: Salesforce press release, 2024-04-25].
SAP side: SAP Business Data Cloud (BDC) Connect (announced Oct 2025) plus SAP Datasphere provide bidirectional zero-copy with Databricks, Google BigQuery, Snowflake (Snowflake productive availability planned Q1 2026) and Microsoft Fabric [Public: SAP Community / CIO, 2025].
The gap: No announcement of a direct Salesforce Data Cloud ↔ SAP Datasphere/BDC zero-copy link surfaced in current sources [Public: negative result, 2024–2025 scan]. Both ecosystems federate to the same third-party lakes, so today the shared layer between them is a neutral lakehouse, not a native SAP↔SFDC bridge [Synthesis].

The canonical pattern, and where the “common data layer” actually lives

Canonical SAP↔SFDC pattern: middleware-mediated, API-based sync (MuleSoft / SAP Integration Suite / third-party iPaaS), typically poll-based and API-limited [Public + Synthesis].
Where the common layer lives: not inside SAP and not inside Salesforce. In the sync pattern it’s implicit in the middleware’s mappings; in the modern pattern it’s an external lakehouse both sides federate into [Synthesis].

NVIDIA’s ODP ambition against these patterns

NVIDIA’s ODP data lake is described internally as “the intended single source of truth the planning and control-tower layers would read from, versus today’s emails and spreadsheets” [Interview: NVIDIA RMA glossary, 2026-07-06]. Architecturally ODP is an external-lakehouse / zero-copy-target play, not a point-to-point SAP↔SFDC sync — the neutral layer both systems feed, precisely the shape the industry is converging on [Synthesis]. Two observations, not conclusions:

NVIDIA’s own framing — data scattered across “SFDC, SAP, the NVIDIA data lake, customer reply-all chains, and 3PL spreadsheets, with no single source of truth” [Interview: engagement one-pager, 2026-06-18] — includes sources (customer email chains, 3PL spreadsheets) that live entirely outside any SAP/SFDC zero-copy federation. Even perfect ODP↔SAP↔SFDC federation would not capture the customer-side and 3PL legs [Synthesis].
The zero-copy story addresses reconciling internal systems; it does not address the external visibility gap [Synthesis].

5. APIs, extensibility & lock-in risk

BLUF: Salesforce is the easier surface to extend (broader marketplace, more accessible languages, cleaner API story); SAP is arguably harder to leave (deeper data gravity, ABAP’s steeper egress, a decades-deep on-prem installed base). Both push proprietary languages and metadata models that create real switching cost [Synthesis].

SAP developer surface

BTP is the extension platform — custom logic routes off the digital core onto SAP Business Technology Platform as “side-by-side” extensions (ABAP Cloud, CAP, low-code SAP Build) [Public: SAP News Center, Aug 2025].
“Clean core” is a formal A–D maturity model — Level A (fully clean) → Level D (upgrade-unsafe) [Public: SAP News Center, Aug 2025; avotechs.com, 2025].
RAP is the core programming model — the ABAP RESTful Application Programming Model auto-exposes Fiori apps as OData/REST [Public: SAP Community / help.sap.com, 2025].
API discovery lives in the SAP Business Accelerator Hub (api.sap.com) [Public: help.sap.com, 2025].
Marketplace: SAP Store lists 2,300+ solutions [Public: sfapps.info / SAP Store, 2025].

Salesforce developer surface

Multiple API families draw from one shared daily pool. REST, SOAP, Bulk, Connect all count against a single per-org 24-hour allocation — Enterprise Edition ≈ 1,000,000 + users × 1,000; Developer Edition capped at 15,000/day; sandboxes 5M/day [Public: Salesforce Developers limits cheatsheet, 2025].
Apex governor limits constrain per-transaction work: 100 SOQL queries, 150 DML, 10,000 rows, 100 HTTP callouts; Bulk API 2.0 allows up to 150M records/rolling 24h [Public: Salesforce Developers, 2025].
Build surface: Apex (proprietary), Lightning Web Components, Flow, plus REST/SOAP/Bulk/Streaming APIs.
Marketplace: AppExchange ~6,000 apps (May 2025) → 6,233 (Dec 2025) — roughly 2.5–3× the SAP Store count [Public: sfapps.info, 2025].

Head-to-head

Dimension	SAP	Salesforce
Proprietary language	ABAP / ABAP Cloud	Apex
Modern build model	RAP, CAP, SAP Build (on BTP)	LWC, Flow, Apex
API catalog	Business Accelerator Hub (REST/OData/SOAP)	REST/SOAP/Bulk/Streaming
Governor model	Clean-core compliance levels (A–D)	Hard per-org API + per-txn Apex limits
Marketplace apps	2,300+ (SAP Store)	~6,200 (AppExchange)

Lock-in comparison (candid)

Easier to extend? Salesforce leads on ecosystem breadth and language accessibility — a larger AppExchange, Apex/LWC more widely staffed than ABAP. SAP’s clean-core push deliberately narrows what counts as sanctioned extension, improving upgrade safety at the cost of governance overhead [Synthesis].
Harder to leave, and why? SAP: S/4HANA is the system of record for finance, supply chain, and manufacturing; ABAP is niche; decades of ECC customization make migration costly even within SAP. Salesforce: Apex, custom objects, and Flow logic are expressed in a proprietary metadata model; hard per-org API limits make bulk egress operationally slow. Net read (to test, not conclude): SAP’s lock-in is rooted in data centrality + migration cost of the record itself; Salesforce’s in proprietary application logic + API egress friction [Synthesis].

NVIDIA worked example: NVIDIA runs SAP (ECC→S/4HANA) and Salesforce (SFDC) side by side, portals (ZMRR portal, SFDC) fronted by SSO — the common enterprise pattern the lock-in analysis describes, each system carrying its own switching cost [Synthesis].

6. AI integrations

BLUF: Both vendors now ship a copilot plus an agent layer priced on consumption; SAP meters “AI Units,” Salesforce meters conversations/actions (Flex Credits). The clearest shipping facts are the pricing mechanics and protocol adoption (both back MCP and A2A). The most announced-but-maturing piece is deep autonomous, cross-vendor agent execution — including NVIDIA’s GTC 2026 Agent Toolkit tie-ins [Synthesis].

SAP: Joule, Joule Agents, Business AI

Copilot: Joule is SAP’s generative-AI assistant across the Business Suite. Collaborative Joule Agents were unveiled at SAP TechEd (Oct/Q4 2024), first agents GA Q4 2024 [Public: SAP News Center, Oct 2024]. (Shipping.)
2025 expansion: At SAP Connect (Oct 2025), ~14–15 new Joule Agents embedded across the suite [Public: SAP News Center, Oct 2025]. (Mix of shipping + rolling GA.)
2026: Joule Studio (custom agent builder) GA in Q1 2026 [Public: aimultiple.com, 2026]. (Verify GA scope.)
Pricing — AI Units: SAP reset AI/data licensing in July 2025, introducing AI Units as a consumption currency billed outside the S/4HANA user metric. List ~$0.40/unit; overage ~$7/unit in 100-unit minimums; document grounding ≈0.005 units/record; customer-built agents ≈0.005–0.025 units/step. Annual purchase, 12-month expiry, no rollover [Public: saplicensingexperts.com, 2025–2026]. (Unit-rate figures from licensing analysts, not an official SAP rate card — directional.)

Salesforce: Einstein → Agentforce, Data Cloud, Atlas

Copilot→agents: Einstein evolved into Agentforce (1.0/2.0/3.0), with Data Cloud (Data 360) as the grounding layer and the Atlas reasoning engine driving planning [Public: Salesforce, 2025]. (Product-line naming shipping; per-version claims should be version-checked.)
Pricing — three coexisting models: (1) original $2 per conversation; (2) Flex Credits (May 2025) — $0.10/action = 20 credits, voice = 30 credits, 100,000-credit packs at $500, EE gets 100,000 free via Foundations; (3) per-user license from $125/user/month [Public: Salesforce press release, May 15 2025; getgenerative.ai, 2025].

AI across the two (interop)

Both back MCP and A2A. Salesforce shipped a native MCP client (pilot July 2025), an MCP server registry, MuleSoft API-to-MCP conversion, and AgentExchange; it supports Google’s A2A (launched Apr 2025) [Public: Salesforce Developers blog, Jun 2025]. SAP integrated MCP (MCP Gateway in Integration Suite) and A2A (Agent Gateway) into Joule for governed tool access [Public: SAP Architecture Center, 2025–2026]. (Some SAP A2A/MCP support is roadmap through 2026.)
NVIDIA Agent Toolkit (GTC 2026): NVIDIA launched an open enterprise agent platform with 17 named adopters including SAP, Salesforce, Adobe, ServiceNow, Siemens, Cadence, Synopsys, Palantir, Red Hat, Cisco. Components: Nemotron (open agentic-reasoning models), AI-Q (enterprise-knowledge blueprint), OpenShell (policy/security runtime), cuOpt (optimization skills) [Public: VentureBeat; NVIDIA newsroom, 2026].
Vendor-specific NVIDIA ties: SAP × NVIDIA (expanded Mar 18, 2024) — NIM microservices + AI foundry into SAP’s generative AI hub, BTP, Datasphere, RISE; SAP adopting Llama Nemotron [Public: NVIDIA newsroom / SAP News Center, 2024–2026]. Salesforce × NVIDIA — Nemotron models in Agentforce, reference architectures targeting regulated/on-prem deployment [Public: Salesforce news / Salesforce Ben, 2025–2026].
Direct SAP↔Salesforce joint AI work: none surfaced; overlap is indirect — co-membership in MCP/A2A standards and shared NVIDIA Agent Toolkit adoption, not a joint feature [Synthesis].

Sub-questions

More real vs. announced? Shipping (high confidence): both copilots, both consumption meters, MCP/A2A adoption. Announced/roadmap: fully autonomous cross-vendor orchestration, much of the GTC 2026 toolkit integration [Synthesis].
Does AI deepen or loosen lock-in? Consumption pricing tied to each vendor’s grounding layer (SAP business data/knowledge graph; Salesforce Data 360) tends to deepen data-gravity lock-in — the agent is only as good as vendor-resident context. MCP/A2A pushes the other way toward interop. Net direction unresolved and worth tracking [Synthesis] [Speculation].

7. Why so sticky despite the complaints

BLUF: Enterprise ERP and CRM incumbents retain customers who openly dislike them because leaving is a multi-year, nine-figure operational risk — and the failure record is public. The complaints are real, but they are not churn triggers. Five mechanisms compound: switching cost, data gravity, the systems-integrator economy, process embedding, and “load-bearing but clunky” inertia. The falsification question is whether AI agents, composable architectures, zero-copy data, and greenfield buyers are now loosening any of them.

7.1 Switching cost is a documented graveyard

Lidl / SAP — ~€500M written off, abandoned 2018. After ~7 years building “eLWIS,” Lidl reverted to legacy “Wawi.” Root cause: Lidl valued inventory at purchase price while standard SAP for Retail assumes retail price; it customized until the project collapsed [Public: Consultancy.uk / Computer Weekly, 2018].
Revlon / SAP S/4HANA — ~$64M unfillable orders, shareholder lawsuits, 2018. A Feb-2018 go-live disrupted the Oxford, NC plant; the stock fell ~6.9% in 24 hours and ≥4 securities class actions followed — a rare ERP failure producing investor litigation [Public: Computer Weekly / TechTarget, 2019].
National Grid / SAP (Wipro) — $585M cleanup, go-live Nov 2012. SAP went live days before Hurricane Sandy; payroll broke, book-close went from <1 week to 43 days, cleanup ~$30M/month with ~850 contractors — >150% of original cost. National Grid later sued Wipro [Public: Computerworld / TechTarget, 2017].
MillerCoors / HCL — $100M suit, March 2017. Sued its SI over a delayed SAP deployment; later settled [Public: TechTarget / UpperEdge, 2017].
Birmingham City Council / Oracle Fusion — £19M budget → £216.5M projected, Section 114 (bankruptcy) 2023. Migrating off legacy SAP R/3 to Oracle Fusion (go-live Apr 2022); the system posted transactions incorrectly and left the council unable to confirm whether fraud had occurred [Public: Computer Weekly / The Register / DCD, 2024–2026].

The pattern is the moat. Each case tells the next buyer that ripping out the system of record risks the business, not just the IT budget [Synthesis].

7.2 Data gravity: the system of record is the audit trail

ERP holds the general ledger, statutory reporting, and the close process, embedded in regulatory obligation (SOX, statutory/tax filings, auditor reliance). Birmingham makes the inverse vivid: when the ERP’s audit features failed, the council could not attest its books were clean [Public: The Register, 2024]. Moving financial data means re-certifying an entire control environment [Synthesis].

7.3 The systems-integrator economy is a self-reinforcing labor market

The people who could migrate you off the incumbent make their living implementing it. The ERP SI & consulting market was ~$11.6B in 2024 inside a broader ~$553B (2025) system-integration market — Accenture, Deloitte, TCS, Capgemini, IBM, Cognizant, EY leading [Public: MarketsandMarkets, 2024–2025]. The Salesforce “economy” is projected (IDC, Salesforce-commissioned) to add ~$2.02T in revenues and ~11.6M jobs (2022–2028), partner multiplier $6.19 → $6.84 per $1 of Salesforce revenue [Public: IDC via Salesforce, 2024] (vendor-sourced; treat magnitude as directional, existence is not in dispute). A certified-admin labor market means the switching decision competes with a trained internal workforce. This is what Athey named a “Trojan horse”: an ecosystem that both extends the platform and locks the customer to it — add-ons “can be quietly strangled by the platform they depend on” [Interview: Susan Athey, 2026-05-06].

7.4 Process embedding: the software encodes the company

Lidl’s purchase-price inventory logic was a competitive process, not a config choice — it customized the software until the project collapsed [Public: Consultancy.uk, 2018]. The more common outcome is the reverse: the org reshapes around the software’s assumptions until no one can separate “our process” from “how SAP does it.” At that point the ERP is the company’s operating definition of its own processes [Synthesis].

7.5 “Load-bearing but clunky”: dissatisfaction that doesn’t cause churn

Salesforce reported ~8% attrition (≈92% retention) for Q2 FY2025, 73% of new bookings from existing customers [Public: Salesforce/SEC FY2025, 2025].
On the SAP side, 48% of ASUG respondents reported planning a migration in 2024 — but “planning is not doing” [Public: ASUG Pulse of the SAP Customer, 2024].
A Nucleus Research study found 6 of 10 SAP customers said they would not buy SAP again — yet stay [Public: Nucleus Research via InformationWeek] (older study; illustrative of durable sentiment, not a current metric).

Our interviews land the same point. Andy at Skiffra: “SAP is great at reporting, bad at predicting” — a sharp, real complaint that is not a reason to leave [Interview: Andy, Skiffra, 2026-05-01]. Holly at Renesas: SAP is the “unambiguous industry standard,” switching costs “hundreds of millions of dollars and years” [Interview: Holly Rawlins, 2026-04-29]. You don’t complain this much about software you can walk away from [Synthesis].

7.6 What would make this wrong / what could break the moat

Athey’s hedge applies: “If you don’t have a better option, or if this thing you have is so much better, people will still choose it” [Interview: Susan Athey, 2026-05-06].

AI agents abstracting the UI. If the interface most users touch becomes an agent, the incumbent’s UI moat (and the admin/certification labor market) erodes. Salesforce is trying to own this transition (Agentforce + Data Cloud), itself an admission that the UI-level moat is contestable [Public: SiliconANGLE, 2025] [Speculation].
Zero-copy loosening data gravity. Salesforce Data Cloud ingested 32T records in Q3 FY2026, 15T via zero-copy (+341% YoY). If the record can be queried in place, data gravity weakens — though Salesforce deploys it to strengthen its position [Public: Salesforce/Xillentech, 2025–2026] [Speculation].
Composable / headless ERP + best-of-breed unbundling. Gartner’s “postmodern → composable ERP” thesis (≥50% of megavendor clients evaluating multiple vendors); Coca-Cola European Partners cited replacing a monolithic SAP system with modular microservices [Public: Gartner via Infosys/TechTarget, 2024] [Speculation].
Greenfield companies picking neither — the sharpest signal for TBD’s own thesis. Most current YC startups reportedly use neither Salesforce nor HubSpot; AI-native tools (Attio — $52M Series B; Clay — $100M Series C at $3.1B) assume agents, not humans, enter data [Public: SaaStrAI / BuildMVPFast, 2025–2026]. Maps directly to Holly: “new U.S. manufacturers — perhaps they’re open to a new system” precisely because they lack SAP entrenchment [Interview: Holly Rawlins, 2026-04-29]. The moat is weakest where there’s nothing to switch from [Synthesis].

7.7 Worked example — NVIDIA proves the ceiling of “just replace it”

A company with every reason and resource to build its own stack doesn’t. NVIDIA — cash-rich, engineering-elite — runs SFDC + SAP as systems of record, is mid-migration on SAP, and rather than replace the incumbents, layers custom tooling and a data lake (“ODP”) on top [Interview: NVIDIA / Greg + Lonny, 2026-06-17; NVIDIA RMA glossary, 2026-07-06]. It committed $2M+ to SAP for planning automation while evaluating Palantir / Accenture / build-internal for the surrounding layer — not for ripping out SAP [Interview: NVIDIA, 2026-06]. The smart-money move for even the strongest engineering org is to wrap, not replace — rhyming with Renesas’s “Rainbow” outliving the migration meant to subsume it [Interview: Holly Rawlins, 2026-04-29].

7.8 Open questions

Where is the crossover at which a “better option” overcomes switching cost? Athey cited 2% vs. 40% share as obviously broken extremes; the crossover is unknown. Who could tighten it: Athey.
Does the wrap-not-replace pattern (NVIDIA ODP, Renesas Rainbow) represent a durable complementary layer a startup could occupy — or does the incumbent eventually absorb it? Who: Alex Zhu / NVIDIA SAP architects.
Is the greenfield “picks neither” signal real at the manufacturing/ERP core, or confined to lightweight CRM? Who: a greenfield US fab operator (Holly’s lead).
Do zero-copy and agent layers dissolve the moat, or migrate it one level up? The vendors deploying them are the incumbents [Speculation].

Convergences & divergences (internal ↔ public)

Convergences.

SAP as the unavoidable financial core. Holly Rawlins [Interview, 2026-04-29], the NVIDIA stack [Interview, 2026-06-17], and public CRM/ERP share data all agree SAP owns the ledger and Salesforce owns engagement — and that the two coexist rather than one displacing the other.
“Load-bearing but clunky.” Andy’s “great at reporting, bad at predicting” [Interview, 2026-05-01] matches Nucleus’s “6 of 10 wouldn’t buy again but stay” and Salesforce’s ~92% retention — complaint and retention coexist across both internal and public evidence.
Wrap-not-replace. NVIDIA’s ODP and Renesas’s Rainbow (internal) mirror the public composable-ERP and zero-copy patterns — the shared layer forms outside the incumbents.

Divergences / tensions to flag.

Migration intent vs. action. ASUG’s “48% planning migration” is stated intent; Holly’s “hundreds of millions and years” and the failure graveyard are why intent lags action. Treat any “X% migrating” figure as planning, not doing.
Vendor-sourced magnitudes. IDC’s Salesforce-economy figures and Salesforce’s retention/Data Cloud numbers are directionally solid but self-reported.
The ODP framing rests on a negative result. No public source confirms a direct Salesforce Data Cloud ↔ SAP Datasphere zero-copy link; both federate to third-party lakes. If that changes, the “common layer lives outside both” framing weakens.

Confidence summary

Claim area	Confidence	Load-bearing caveat
SAP = ledger / Salesforce = engagement; they coexist	High	Robust across internal + public
Financials, marketplace counts, API limits, pricing mechanics	High	Point figures may drift with each release
Migration failures (Lidl/Revlon/National Grid/Birmingham)	High	Well-documented, dated, dollar-quantified
Salesforce-economy size, retention, Data Cloud record counts	Medium	Vendor-sourced; directional
SAP AI Unit per-unit rates ($0.40 list / ~$7 overage)	Medium	From licensing analysts, not official rate card
”No direct Data Cloud ↔ Datasphere zero-copy link”	Medium	Negative result; load-bearing for ODP framing — recheck
Athey “Bloomberg Terminal” as canonical example	Low	Frame confirmed in 2026-05-06 file; Bloomberg-specific line is from prior sessions — verify before external use
Salesforce per-cloud revenue split	Low	Not broken out in official release; aggregator numbers appear AI-rebranded — treat as `[Speculation]`
Whether AI/zero-copy loosens or re-entrenches lock-in	Low	Genuinely unresolved; a question to track

Sources

Internal (memory vault)

[[sap-in-semiconductor-supply-chain-2026-05-27|SAP in the semiconductor supply chain, 2026-05-27]] — SAP portfolio, economics, migration deadline, blind-spot framing
[[nvidia-primer-2026-06-18|NVIDIA primer, 2026-06-18]]; [[2026-07-06-nvidia-rma-acronym-glossary|NVIDIA RMA glossary, 2026-07-06]]; [[2026-06-20-nvidia-pre-meeting-knowledge-map|NVIDIA pre-meeting knowledge map]] — SFDC+SAP stack, ECC→S/4HANA, ODP data lake, $2M+ planning commitment
[[2026-06-17-logistics-catch-up-wgreg-lonny-nvidia|Greg + Lonny @ NVIDIA, 2026-06-17]] — “death by a thousand emails,” data-scatter framing
[[2026-04-29-call-holly-rawlins-re-renesas|Holly Rawlins (Renesas), 2026-04-29]] — SAP as gravity well, Rainbow persistence, greenfield opening, switching cost
[[2026-05-02-zoom-dustin-ross-skiffra|Andy (Skiffra), 2026-05-01/02]] — “great at reporting, bad at predicting”
[[2026-05-06-stramgt-3862-athey|Susan Athey, 2026-05-06]] — Trojan-horse / threat-of-complements framing

External — §1–2 (competencies, overlap)

External — §3–4 (deployment, integration)

External — §5–6 (APIs, AI)

External — §7 (stickiness, failures, moat-breakers)

Lidl: Consultancy.uk · Computer Weekly
Revlon: Computer Weekly · TechTarget
National Grid: Computerworld · TechTarget
MillerCoors/HCL: TechTarget · UpperEdge
Birmingham/Oracle: Computer Weekly · The Register · DCD
IDC Salesforce Economy · SI market size (MarketsandMarkets) · ASUG Pulse 2024 (IgniteSAP) · SAP satisfaction / Nucleus (InformationWeek)
Salesforce FY2025 8-K (SEC) · Composable ERP (TechTarget) · Agentforce/Data Cloud hyperscaler path (SiliconANGLE) · AI-native CRM startups (SaaStrAI)

Prepared 2026-07-08 via the /research workflow. Landscape map, not a verdict — synthesis is a human activity per RDI.

Project TBD Memory Vault

Explorer

sap-salesforce-enterprise-stickiness-2026-07-08