Big Technology Podcast

Christian Klein on why an LLM alone can't run a business

Christian Klein· CEO at SAP
·~59 min·English·Big Technology
Business StrategyAgentsAI InfrastructureOpen SourceAI Company
TL;DR

SAP's CEO argues the large language model is only the engine of enterprise AI - the value, and SAP's moat, is the surrounding layer of task-level accuracy, business data and ontology, governed access, and cost control, much of it resting on 50 years of ERP depth.

01The Value Gate

For mission-critical work, 93% is not almost there

The accuracy an agent must clear is set by the task, not the model - a customer brief tolerates a wrong figure that a financial close cannot, so the same 93% is good enough in one case and unusable in the other.

it understands today 90 with a 93% accuracy but your auditors when you release financial results they say okay sorry this time my numbers were 7% you know too high

— Christian Klein, Big Technology Podcast
Key Insight
Most debates about enterprise AI treat accuracy as a single dial you turn up. Klein's point is that the dial has a different pass mark for every task. A customer brief tolerates a wrong earnings figure; a financial close that is 7% off is not 93% done, it is unshippable. That is why the useful question is not how smart the model is, but which tasks have just crossed their own bar.
02The Mechanism

An LLM alone cannot run a warehouse

SAP's answer to accuracy is not a bigger model but a stack: the LLM for language, an AI foundation of business data and a semantic ontology for meaning, and governance for the legal and tax rules an agent must obey.

If you just use an LLM alone, there's no way that you can run a warehouse with it, that you can do a financial close with it

— Christian Klein, Big Technology Podcast
Key Insight
This reframes where the hard part of enterprise AI lives. The model is a commodity ingredient; the scarce part is the layer that tells the model what a maintenance order means, which country's tax code applies, and who is allowed to see a number. A general reader can see why SAP keeps insisting the model is necessary but not sufficient - the sufficiency is the thing they sell.
03The Moat

A prompt cannot easily rebuild 50 years of ERP

When the vibe-coding narrative hit, Klein asked his product teams which SAP products a prompt could rebuild - the answer was almost none, because an ERP encodes millions of data correlations and country-specific process knowledge that are hard to acquire.

you need to have a lot of domain knowledge

— Christian Klein, Big Technology Podcast
Key Insight
Klein is careful not to say impossible - he says not easily, and concedes that one or two small, low-context tools could be replicated. The deeper point is where the barrier sits: a prompt can lower the cost of writing code, but not the cost of acquiring decades of domain knowledge about how finance, payroll, and supply chains actually run. The moat is in the accumulated structure, not the user interface.
04The Control Point

Every outside agent comes through one governed door

SAP will let third-party agents like ChatGPT or Salesforce reach its data, but only through an agent gateway that enforces governance - while its own coworker, Joule, stays the native front end with full context.

a pure API connection pure MCP server is not the same like the ontology this the context we having on our platform

— Christian Klein, Big Technology Podcast
Key Insight
This is the strategic core of the whole conversation. A plain API or MCP connection hands over rows of data; it does not carry the meaning or the permission rules around them. Klein says SAP will offer both an open door and a gate - outside agents may reach the data, but only through a path SAP governs, so SAP keeps the control point even when a rival's agent is doing the asking.
05The Cost Discipline

The best model isn't always the right model

SAP chooses a model per agent on outcome versus cost and switches constantly - some agents are on their fifth model - because a standard or open model often does the job at a fraction of the frontier price.

it's not only important to that the agent does a great job, but it's also at what cost does the agent does a good job

— Christian Klein, Big Technology Podcast
Key Insight
Klein confirms a pattern the host brought from RAMP's data: the share of AI spend going to frontier models is sliding. For a buyer, that turns model choice into a per-task procurement decision, not a loyalty to one lab. For the frontier labs, it is the harder problem - if a cheaper model clears the task's bar, the premium for being the smartest has nothing to attach to.
06The Reckoning

When the bill rises faster than the productivity

SAP saw token spend rise enough to require limits, so it set budgets by job profile - capping most roles, leaving the top 1% of model-trainers unlimited, and adding a per-task approval to raise a limit.

it doesn't help you if some of your employees is getting 20% more productive if at the same time you know the cost going 30% up

— Christian Klein, Big Technology Podcast
Key Insight
Klein was reluctant to meter tokens because it contradicts the message to use AI freely - a tension any leader rolling this out will recognize. His resolution is telling: the limits are a backstop, but the real lever is switching to a cheaper model that clears the same bar, which in SAP's experience did more than the token limits. Cost control here is an engineering choice before it is a policy one.
07The Human Element

A workforce transformation, not just a restructuring

Klein frames the people change as a shift in the mix of skills over about a year - hiring data scientists and full-stack developers, reskilling existing staff, and needing fewer people in some roles - rather than a simple headcount cut.

there will be a workforce transformation but it's not necessarily that it's only about restructuring it's really about you know at the end of the day you need to have the right skills and the right mix of people

— Christian Klein, Big Technology Podcast
Key Insight
The load-bearing word is mix. Klein is not promising headcount stays flat - he says some roles shrink while new ones are hired. Whether a given person keeps a seat turns on reskilling, which he insists works in practice but admits not every employee will accept. So the displacement is real at the level of individual roles, even if the company's total holds or grows.
08The Policy Ask

Regulate the harm, not the tool

From the seat of Europe's largest software company, Klein argues regulators should govern AI's impact on society rather than the technology itself, warning that overlapping EU laws make it hard to compete.

you need a regulation but you should regulate more the business outcome the impact on a society and not the technology per se not the use of data

— Christian Klein, Big Technology Podcast
Key Insight
Klein grants the good intentions and still says the design is wrong. His line - regulate the outcome, not the tool - is a testable principle: judge a system by the harm it causes, not by the data it uses or the method inside it. He reads Brussels' in-progress omnibus simplification as a quiet admission that the rules were stretched too far, and ties it directly to whether a startup can get off the ground in Europe at all.