10x Signal: Data & AI Trends · Newsletter

SE80 was the lock-in

Episode 5 - how SAP's cloud strategy can strengthen SAP, and make parts of SAP easier for AI to displace.

10x Signal: Data & AI Trends - Episode 5

Part of what made SAP hard to replace was never a feature. It was SE80, SMOD, CMOD and twenty years of transports: custom code, bespoke integrations, embedded process knowledge and the cost of reconstructing decades of undocumented decisions.

SAP Public Cloud and Clean Core remove that friction — deliberately, and for good reasons. SAP is rebuilding its position one layer higher, in business semantics, data, governance and agents. The customer decision is not whether to keep the Z-code. It is who owns the context and the decision logic once the Z-code is gone.

A recruiter said something to me recently that I have not been able to put down:

“We’ve moved from SAP to AI. That’s where we see the future.”

Not a customer. Not an analyst. A recruiter — someone watching where capable people choose to place the next phase of their careers.

One sentence is not a trend. But it is a useful way into a question the market still has not answered.

THE MARKET CAN’T MAKE UP ITS MIND

The simple story says investors feared AI would make classical software obsolete, then discovered that good software becomes more valuable when AI needs trusted applications and governed data.

The evidence is less settled — and more interesting.

Reuters reported that the S&P 500 software and services index lost more than 33% between its October 2025 peak and 10 April 2026, then rebounded 33%, sold off again and partly recovered. By 26 August it was still down more than 3% for the year and more than 12% below its peak. Over the same period, second-quarter earnings growth for the sector tracked at 24.4%, well ahead of the 16.4% expected at the start of reporting season.

That is not a verdict. That is a market oscillating between disruption risk and earnings evidence.

SAP’s own Q2 2026 numbers make the bullish case, and they make it well:

Current cloud backlog reached €22.9 billion, up 27% year over year and 26% at constant currencies.

Cloud revenue reached €6.28 billion, up 22% and 24% at constant currencies.

Cloud ERP Suite revenue reached €5.53 billion, up 25% and 27% at constant currencies — roughly 88% of total cloud revenue.

Management said AI and SAP Business Data Cloud were key pillars in more than 90% of SAP’s 50 largest deals.

Those figures do not prove AI caused the deals. They do prove AI and data have moved into SAP’s largest commercial conversations.

And yet, on the day I am writing this, SAP shares fell after UBS downgraded the company and questioned how fast announced AI capabilities were actually shipping.

So let me be clear about what I am not arguing. SAP is not being displaced in its reported results. But quarterly results cannot tell us which layer of enterprise software captures the value five years from now.

That is the more important question — and it is why this sits on an independent platform rather than a vendor blog. No reseller quota depends on the answer.

WHOSE LOCK-IN WAS IT?

Ask why SAP has been so hard to replace for thirty years, and the honest answer is not that every function beat every alternative.

Part of the answer was SE80.

Thousands of companies run processes that exist nowhere else. Z-programs were written by people who retired a decade ago. User exits went into SMOD and CMOD and were never revisited. Integrations were built, changed, and documented halfway. Industry logic was encoded in configuration rather than written down. Institutional knowledge sat with employees, partners and consultancies.

That created formidable switching costs. A competitor did not merely have to offer better software. It had to reconstruct twenty years of undocumented decisions without breaking the business.

Research on ERP customisation reaches the unsurprising conclusion from the customer’s side: customisation raises maintenance and upgrade effort and increases defect risk. But the same cost that harms the customer protects the incumbent.

That distinction is the one most programmes never make explicit:

What makes SAP hard to replace is not what makes your company hard to imitate.

Some Z-code encodes genuine operating advantage. Much of it encodes history, local preference, or a workaround from 2011 that nobody has dared to challenge since. A cost can be SAP’s lock-in without ever having been your advantage.

Sorting one from the other is not a technical exercise, and it does not belong in a backlog. It is the decision the entire programme rests on — which is why we published Clean Core as a board decision rather than an architecture guideline.

CLEAN CORE TAKES THE WALLS DOWN

Now consider the architecture we are all recommending:

Public Cloud. Fit-to-standard. Clean Core. Governed master data. Released APIs. Extensions with their own lifecycle.

Every one of those is sound. I have argued for them before and would again. The alternative is a landscape that gets more expensive to change with every release, until it cannot be changed at all.

But these principles do something else as a side effect: they separate layers that a Z-heavy ECC landscape had fused together.

AI will replace SAP

Business context & differentiation — the prize. Sticky because: semantics, policies, thresholds, decision logic, judgement. What AI changes: whoever controls it decides how agents act.

Interaction & orchestration — goes first. Sticky because: screens, navigation, workflow, approvals, exception routing. What AI changes: disintermediated fastest, as agents work across applications.

System of record — slowest. Sticky because: transaction integrity, controls, localisation, audit trails. What AI changes: agents depend on it; substitution is slowest here.

A clean, standard ERP does not become replaceable overnight. It becomes understandable — and operable from above.

That is a much narrower claim than “AI will replace SAP.” It is also far more credible.

MACHINES CAN’T READ SE80

Modern agents do not need to understand an application the way an experienced consultant does. They need a reliable description of the tools they may use, the data they may read, the actions they may take, and the constraints they must respect.

Research such as ToolLLM and Gorilla shows how large language models learn to select and call large catalogues of APIs, particularly when current documentation can be retrieved at the point of use. That research does not show agents are ready to run an enterprise unsupervised — real API execution still fails in important ways.

But it establishes the mechanism: documented, stable interfaces cut the cost of teaching a machine to operate a system.

An agent cannot read a Z-program written in 2009 with no documentation, no test coverage and three implicit dependencies. It can read a released OData service with a published contract.

Clean APIs are therefore not evidence that SAP will be replaced. They are evidence that the cost of building an intelligent layer above SAP is falling.

The first economic pressure may not land on the ledger. It may land on the application surface: user interfaces, workflow licences, navigation, reporting, routine exception handling, and whole categories of extension. Substitution then moves downward one modular capability at a time.

SAP ISN’T ASLEEP

The strongest counterargument is that SAP is not waiting for any of this to happen.

At Sapphire 2026, SAP brought Business Technology Platform, Business Data Cloud and Business AI together as the SAP Business AI Platform. At its centre sits the Knowledge Graph — a structured map of business entities, processes and relationships. SAP describes Business Data Cloud as the governed “knowledge core” from which applications and agents draw business context.

That is not a defensive product update. That is SAP moving deliberately toward the control point above the application.

If SAP owns the trusted semantic layer, the agent platform, the permissions, the process context and the route back into the system of record, then standardisation below does not commoditise SAP at all.

It deepens SAP’s role.

SAP also holds assets an AI-native entrant cannot reproduce quickly: installed transaction systems, regulatory localisation, proven controls, audit trails, customer relationships, and a global implementation ecosystem.

An agent may execute steps in a financial close. It cannot absorb management’s accountability for that close. SAP does not hold that legal responsibility either — what SAP holds is the installed control environment through which accountable people operate. That is expensive to recreate, and it is real.

Strong position. It also creates a new customer risk, and this is the sentence I would put in front of a board:

Clean Core removes code lock-in and can replace it with context lock-in.

If your definitions, knowledge graph, policies, agent permissions and decision history become inseparable from one platform, the dependency has not gone away. It has moved upward — to a layer likely to be worth considerably more than the Z-code you just retired.

You did not escape SE80. You replaced it with something you understand less well.

Where Clean Core actually takes you

Clean Core moves you rightward on its own. Only a deliberate second decision moves you up. This is the picture behind Decisions 04 and 06 in the Executive Clarity Decision Library.

IT’S A RACE, NOT A FUNERAL

The long-term question is not whether AI makes SAP stronger or weaker. It is this:

Can SAP become the authoritative AI + process + data layer before open models, cross-vendor agents and customer-owned semantics make parts of ERP execution interchangeable?

Public Cloud accelerates both outcomes at once. It gives SAP the consistency to deploy AI across thousands of customers. It also makes the underlying application easier for other agents to understand and operate.

For customers, the more useful question is not “Will SAP win?” It is:

Will we keep strategic and commercial leverage whichever architecture wins?

Four dimensions decide that. Each maps to a decision we have already published, because none of them can be settled inside a project plan.

  1. Differentiation, not deviation

Which processes and decisions genuinely improve your economics, customer promise or risk position — and which are merely non-standard?

Do not keep complexity because it is yours. Keep a capability only when you can name the advantage it creates, its owner, and the value worth protecting. “It’s in a Z-table” is not an advantage.

→ Clean Core: what leaves the core, and what earns its place.

  1. Semantic ownership, not data location

Where is the authoritative definition of customer, supplier, product, margin, service risk and materiality? Can those definitions, relationships and lineage be exported and reused outside the vendor’s platform?

Owning the records is not enough. In an agentic architecture, the strategic asset is the meaning attached to them.

→ Data & Business AI: what foundation intelligent automation actually assumes.

  1. Decision rights, not extension count

What may an agent read, recommend, write or execute? Which thresholds, policies and exception rules stay yours? Which system produces the evidence an auditor or regulator will accept?

The boundary that matters is no longer “in the core or outside it.” It is the boundary between advice, action and accountability.

→ BTP & extensibility: where differentiation is allowed to live.

  1. Institutional capability, not documentation volume

Can your people explain why the process works as it does, change it deliberately, and challenge the recommendation of an agent or a vendor?

Documentation alone protects nothing. Ungoverned documentation makes knowledge easier for everyone to copy. The asset is controlled, current, machine-readable knowledge tied to named owners, access rights and decision history.

→ Transformation governance: who carries these decisions through delivery.

WHY THE RECRUITER STILL BOTHERS ME

Cloud backlog is forward-looking for revenue, but it is built from contracts already signed. It says little about where scarce capability is forming.

Talent is not a clean forecasting instrument either. People follow compensation, fashion, geography and opportunity as well as belief. One recruiter’s sentence stays an anecdote, and I will not dress it up as more.

But the capability shift is worth watching. If experienced SAP people move toward agent engineering, knowledge graphs, data products and AI governance, that does not necessarily weaken SAP. It changes which SAP capabilities become scarce — and which transformation plans stay deliverable.

Our own skills-gap analysis already shows the contradiction: executives describe talent as a leading constraint while ranking talent attraction, retention and productivity near the bottom of their stated priorities.

So the signal is not “SAP talent is leaving.” It is:

Which combinations of process, data, architecture and AI capability are becoming harder to buy?

Boards should track that before it turns into a programme delay.

WHAT I’D DO ON MONDAY

Standardise what does not differentiate you — aggressively. The SE80 argument is not a defence of Z-code. Unreadable complexity is a recurring tax, and it protects your vendor more reliably than it protects you.

Build a differentiation register before fit-to-standard starts. Do not start from a list of custom objects. Start from the capabilities, decisions and service promises the business believes create advantage — then test which ones actually do.

Treat semantics and decision policies as governed enterprise assets. Name owners. Make definitions machine-readable. Preserve lineage. Test portability. Put export rights, interfaces and exit costs into the commercial decision, not into an appendix.

Separate the system of record from the system of action. For every agentic use case, define where the transaction commits, where controls execute, what evidence is retained, and who stays accountable.

Build the capability before the technology arrives. Codify expert knowledge under access control, develop people who can bridge process and AI, and measure skills risk as seriously as licence and implementation cost.

The uncomfortable version, in one line:

SE80 was never a strategy. It was switching friction doing the work of one. Remove it — but replace it with deliberate ownership of differentiation, semantics, decision rights and capability.

WHERE THIS ARGUMENT CONTINUES

I publish this on sapperment because the argument needs independence. It criticises a market narrative and questions a vendor’s strategic position in the same piece. That is only possible where no reseller quota depends on the conclusion.

Read the decisions. This edition touches four of the eight decisions in the Executive Clarity Decision Library — Clean Core, BTP & extensibility, Data & Business AI, and Transformation governance. They are sequenced the way a board actually meets them, connected to one another, and revised as the facts change. A feed forgets; a library revises.

Contest them. Leading Top Voices and Experts bring independent, domain-specific perspectives to questions with no single correct answer. Expertise is organised by domain rather than prominence — you find the person for the question, not the other way round. Publishing is by invitation; if you carry a practitioner perspective that deserves to outlive the feed, propose yourself as an LTV author.

Build the capability. The sapperment Talent Network is live in beta, and it is built for exactly the shift described above. Skills are published under a reference number — no names, no photos, no CV uploads. Bias has no data to act on, and specialists who are currently employed can be found without exposing a job search to their employer. Browse profiles by domain, or join the network.

Tell me where this argument is wrong. sapperment exists to make these decisions contestable before a programme makes them irreversible.

EVIDENCE AND SOURCES

Evidence note: The financial results and product descriptions below are documented claims. The argument that standardisation changes the location of competitive control is an inference. The recommendation on what customers should own is the author’s point of view. Where the three differ, I have tried to make it visible.

SAP, Q2 and Half-Year 2026 Results — backlog, cloud revenue and Cloud ERP Suite figures.

SAP Q2 2026 earnings-call transcript — management statement on AI and SAP Business Data Cloud in the 50 largest deals.

Reuters, “Momentum, leveraged trading keep software stocks on roller coaster,” 26 August 2026 — software-index path and Q2 sector earnings growth.

MarketWatch, “SAP stock slumps as UBS downgrades stock on ‘slow’ move into AI,” 26 August 2026 — same-day market counter-signal.

SAP, “SAP Unveils the Autonomous Enterprise,” 12 May 2026 and SAP, “Accelerate the Autonomous Enterprise with SAP Business Data Cloud,” 13 May 2026 — Business AI Platform, Knowledge Graph and governed business context.

Yujia Qin et al., “ToolLLM: Facilitating Large Language Models to Master 16,000+ Real-world APIs”, ICLR 2024.

Shishir G. Patil et al., “Gorilla: Large Language Model Connected with Massive APIs”, NeurIPS 2024.

Sudhaman Parthasarathy and Srinarayan Sharma, “Impact of customization over software quality in ERP projects: an empirical study”, Software Quality Journal 25 (2017), 581–598; and Eli Hustad et al., “Customizing ERP-systems: A framework to support the decision-making process”, 2023.


First published on LinkedIn. Subscribe to 10x Signal: Data & AI Trends to get new editions as they appear — or browse the sapperment library.