“Tower of Babel” (Brugel, oil on canvas, 1563, Vienna version)

When Christian Klein walked the Sapphire stage in Orlando last Tuesday, he delivered the most accurate diagnosis of enterprise AI I have heard from a major software CEO in years.

“For the mission-critical processes of our customers, ‘almost right’ just isn’t good enough.”

Eighty percent accuracy is fine for chatting with a model. It is unacceptable for closing the books, scheduling production, or releasing a payment. The Autonomous Enterprise — where agents execute end-to-end processes accurately, compliantly, and securely, where humans stay in the loop on decisions that matter — is where this is going. Klein got everyone excited about the destination which is the consensus among builders.

And yet the market dropped SAP 4.31% the day after the keynote.

Year to date, the stock is down roughly 30%. Over twelve months, roughly 45%. That gap — between a correct vision and a punished stock — is the most interesting question of the week, and it is not answered by saying that investors are short-termist.

The market is pricing two specific premises in Klein’s pitch that are wrong. Both are wrong for the same underlying reason. Both, if you operate an industrial business, should reshape how you think about your AI strategy for the next five years.

Premise one: that the data for agents looks like the data for systems of record

The SAP Business AI Platform, the Business Data Cloud, the Knowledge Graph — the entire architectural premise of Joule and the 200+ agents that orchestrate around it — assumes that the data needed to power autonomous operations is roughly the data SAP has already been collecting, just better organized, federated across cloud, and given a graph structure on top.

This is not what we see in the field.

We deploy agentic systems inside industrial operations every week, in companies that have lived on SAP for two decades. The number we have come back to, again and again, is that roughly 40% of the data an agent needs to execute a real operational decision does not exist in any structured system today. It lives in tribal knowledge, in three-line emails between a plant manager and a supplier, in PDF certificates of conformity, in shift handover notes on whiteboards, in the gut feel of a maintenance technician about machine #7.

That 40% is not an ERP problem. It is a representation problem. The structured-data fortress that SAP spent fifty years building was designed to record transactions — postings, allocations, movements between states. Agentic execution requires a different shape of data: signals about what is changing now, events with the context that produced them, intent about what should happen next and why. The transactional substrate of an ERP is not the same substrate as an agentic substrate, and you cannot get the second by wrapping a knowledge graph around the first.

The remaining 60% — the data that does exist in SAP — has been shaped by twenty years of finance-driven schema decisions. It is correct, audited, reportable. It is also brittle to reason over, because every field was designed to answer “what did we book?” not “what is the operation telling us?”

This is the part the keynote could not address, because addressing it would mean admitting that the next ten years of value will not flow cleanly out of the systems that captured the last forty.

Premise two: that SAP captures the value of the layer it is announcing

The second premise is more consequential. Klein’s pitch is that the Autonomous Enterprise stack — Joule, Business AI Platform, agents, Industry AI — will run on top of S/4HANA, governed by SAP, monetized through cloud subscriptions and a new agent runtime. One vendor. One control layer. One throat to choke.

Angela Strange, a general partner at Andreessen Horowitz, posted a framework this week that captures the alternative more cleanly than any analyst note has so far. Her argument: enterprise buyers facing the AI transition have three real paths, and only one is the durable winner.

The first path is the SAP path. Keep your incumbent system of record — Salesforce, SAP, Workday — and build agents on top, treating the underlying platform as headless backend infrastructure. Familiar, low-disruption, the comfortable choice. Strange’s caution is direct: risk aversion will tempt you here, but you set yourself up to lose to competitors who lower costs and grow revenue with AI. “Try adding an effective voice agent that needs to read/write to 1990s software,” she writes.

The second path is to rebuild the system of record yourself with AI-assisted development — what she calls vibe coding a replacement. The data model, the permissions, the workflows, the audit trail, all coded from scratch. Strange dismisses this directly: your risk appetite should rightly terrify you away from it.

The third path is to buy an AI-native replacement — a system built from the ground up for agents, machine readability, and orchestration. Strange’s argument is that this is the smart choice and the path where the most durable enterprise AI companies of the decade will be built. She points to two live examples in financial services: Valon, AI-native mortgage loan servicing, and Vesta, AI-native mortgage loan origination. Both are already delivering two to three times the efficiency of incumbent peers, with public-company customers willing to confirm it.

This is precisely the structural argument the SAP narrative cannot survive. What Klein announced on Tuesday is, in Strange’s taxonomy, path one — keep SAP, add Joule on top. The same path that the most disciplined enterprise capital allocator in Silicon Valley is publicly identifying as the losing path against path three.

And path three, in our experience deploying inside industrial groups, is not built by horizontal vendors. It is built by vertical players who own the operational verb of a specific industry — the WMS picking the box, the MES running the line, the CMMS scheduling the maintenance — with agentic tooling built ground-up around the operational signal rather than bolted onto a transactional posting layer.

The non-commonality of operational data across industries makes this even more pronounced than in financial services, where Strange’s examples come from. The bill of materials of a turbine has nothing in common with the bill of materials of a yogurt SKU has nothing in common with the bill of materials of a t-shirt. There is no universal ontology of industry. The ontology gets built bottom-up, vertical by vertical, use case by use case, by the players who live closest to that data. SAP can wrap a knowledge graph around what exists. They cannot create the part that doesn’t.

This pattern has played out before, at a different layer. Twenty years ago, SAP’s HR module and SAP’s CRM module were the default for any enterprise that ran on SAP. Then Workday pulled HR and payroll away. Then Salesforce pulled CRM away. Neither displaced SAP’s core; both identified a vertical domain where the user, the data, and the workflow were specific enough that a focused vendor could out-build the horizontal incumbent. They became multi-hundred-billion-dollar companies doing it.

The same pattern is now reaching the operational verb layer. In our own portfolio, OPLIT runs production planning and scheduling inside industrial groups like Teknor Apex, Decathlon — on a data model and a workflow that SAP’s APO module was never going to produce. Cognyx does the equivalent in engineering, building the parts and configuration intelligence that the SAP variant configurator was not designed to deliver. Neither replaces SAP’s general ledger. Both make the SAP modules that used to handle their domain look the way SAP HCM looks today next to Workday.

What SAP actually owns

It is worth being precise about the part of the story that is not nihilist.

SAP owns financial system of record at global scale. IFRS- and GAAP-grade close, multi-entity consolidation, audit trail, intercompany reconciliation, statutory reporting across fifty jurisdictions. This is real, this is durable, and the Autonomous Close Assistant Klein demonstrated will probably ship as advertised and save CFOs real weeks of work. That part of the announcement is credible.

SAP also owns a deep consultant ecosystem and a multi-decade lock on enterprise IT skills. Both are commercial moats, not technical ones. Both are corroding faster than SAP can publicly admit, because the same low-skill consultant base that locked customers in is the one that cannot deliver agentic engineering at the speed and quality the new vision requires. Agentic work requires actual engineering — reasoning about graphs of tool calls, validating outputs, thinking about idempotency, replays, side effects. The vast majority of the SAP services workforce is configured for ABAP and BAPI integrations. The €100M partner fund is going to discover this.

Even the most credible independent industry analyst — Forrester’s Faram Medhora — published the day of the keynote under the title “Credible, But It Comes With Concentration Risk.” His advice to operators was to commit at the architectural pattern level, pilot at the product level, and define explicit go/no-go criteria before declaring SAP the strategic AI architecture for 2030. That is the polite institutional version of the same argument.

What this means if you are operating

If you run an industrial business and you sat through Sapphire — or absorbed the LinkedIn echo of it this week — the operating instruction is straightforward.

Use SAP for the part of your stack SAP genuinely owns: financial system of record, consolidation, audit-grade close. Let the Autonomous Close Assistant ship and adopt it when it does. That value is real.

Do not collapse your agentic strategy into the SAP stack. The data needed to make agents useful in your operations is not the data that lives in S/4, and the orchestration layer of your operations will not be governed by a horizontal vendor headquartered in Walldorf. It will be orchestrated by vertical AI-native players, and the data will be yours, on your system — though that will require a more skilled labour force than the SAP era demanded.

The two layers will coexist. They will not be the same vendor. That is the structural shift the market priced in on Wednesday.

The vision is right. The premises are not.

Klein is correct that the Autonomous Enterprise is where this goes. He is correct that eighty percent accuracy is unacceptable for mission-critical work. He is correct that grounding agents in process, data, and governance is the right architectural instinct.

He is wrong that the data architecture of the next ten years looks like the transactional architecture of the last forty.

He is wrong that a single vendor — least of all the most data-closed enterprise software company in history — captures the entire emerging layer.

The market did not punish the vision. It punished the premises. That distinction is what every operator and every allocator should be drawing this week.

Sources & disclosures

This is not financial advice. OSS Ventures is the founding investor in OPLIT and Cognyx, and the author is the founder and CEO of OSS Ventures.

SAP Sapphire 2026 keynote and Autonomous Enterprise announcement. Christian Klein’s opening keynote, Orlando, 12 May 2026.

SAP stock performance. SAP SE (XETR: SAP) closed at €136.26 on Wednesday 13 May 2026, down 4.31% from the previous day’s close of €142.40 — a 52-week low. Year-to-date decline approximately 32%. Twelve-month decline approaching 50%, from a 52-week high of €273.55. Source: Xetra closing prices via stockinvest.us and Yahoo Finance, retrieved 13–15 May 2026.

Angela Strange — three-paths framework and AI-native examples. Angela Strange, General Partner at Andreessen Horowitz, public LinkedIn post, May 2026. Live examples cited: Valon (AI-native mortgage loan servicing) and Vesta (AI-native mortgage loan origination), reported by Strange as already operating at 2–3x the efficiency of incumbent peers.

Forrester analysis. Faram Medhora, Principal Analyst at Forrester, “SAP Sapphire 2026: The Autonomous Enterprise Is Credible, But It Comes With Concentration Risk,” 12 May 2026.

Historical precedent. Workday (founded 2005, NASDAQ: WDAY) and Salesforce (NYSE: CRM) as vertical AI/cloud-native disruptors of SAP HCM and SAP CRM respectively, both now multi-hundred-billion-dollar enterprise software companies.

OSS Ventures field experience and portfolio. The 40% data-gap figure and the operational characterisation of agentic deployments draw on OSS Ventures’ direct field experience across its portfolio of AI-native industrial SaaS companies. OPLIT (production planning and scheduling) and Cognyx (engineering parts and configuration intelligence) are OSS Ventures portfolio companies. Customer deployments referenced for OPLIT — Teknor Apex, Decathlon — are public. Other deployments are not.