Bounded Autonomy Is the Best Enterprise AI Can Achieve, Says SAP’s Jesper Schleimann

Key Takeaways

⇨
Bounded autonomy is the best enterprise AI can achieve, and organizations stay accountable for the AI they deploy, SAP’s Jesper Schleimann told SAPinsider Copenhagen attendees in his opening keynote.

⇨
People move from doing the work to defining intent and delegating execution to AI, he said, with task-level gains from individual AI use estimated at around 5% to 8%.

⇨
Governed context and central control of agents support trust in enterprise AI execution, and retained learning builds lasting advantage, according to SAP’s Jesper Schleimann at his SAPinsider Copenhagen keynote.

“I do not believe in the fallacy of full-scale autonomy,” said Jesper Schleimann, head of Business AI for SAP in EMEA, about introducing AI into a company’s core operating model. “The best you can achieve is what I would call bounded autonomy.”

Schleimann, author of The Intent-Driven Organization, delivered the opening keynote, “From Intent to Outcome – Autonomous Enterprise,” at SAPinsider Copenhagen on November 17. He described how work changes when people define intent and delegate execution, what it takes to trust that delegation, and how retained learning can become a long-term advantage.

“Never in the history of human have we seen a technology that is so easy to use on a personal level, yet so hard to see when we look at the P&L of enterprises and organizations,” he said as he opened the keynote. He estimated the productivity gain from individual AI use at around 5% to 8% at the task level, which he called “coffee cup level productivity.”

Delegating Work Starts With Defining Intent

Most attendees probably spend 10% to 15% of their time defining what they want to achieve, he said, and most of the day doing the work. In an intent-driven organization, about 90% of the time goes to defining intent, including outcomes, context, constraints, and the evidence of what good looks like, and the work is delegated to AI.

Factory electrification, he said, shows the pattern with new technology. Early factories ran on a single steam engine that turned a central drive shaft, with the whole floor organized around it. Swapping in an electric engine changed little. The productivity benefits took about a decade and a redesign of the floor into decentralized, specialized production lines that could use the electric engine’s flexibility.

In his view, AI is on the same path: organizations tend to introduce a new technology and do things the way they always have, and over time the way they work changes. Rethinking how an organization achieves its outcomes “is not a technology transformation. That’s an organizational transformation,” he argued.

Giving AI Autonomy One Step at a Time

Delegation is his measure of AI maturity, in his words “how much actual work you have delegated to the AI,” which raises the question of how much to hand over. He said “intelligence is already abundant, but there’s still a bottleneck”: ensuring that AI’s execution can be trusted and delivers the intent set out within the right boundaries.

He described bounded autonomy as the best that can be achieved, with AI working “within the right guardrails” and the level matched to the work. The goal, in his words, is “the appropriate level of autonomy for any given task or piece of work.” He asked leaders which autonomy zone a task belongs in, from human-led work, where people do the deciding and the doing, toward bounded autonomy.

Teams should start with humans in the loop and a low level of trust, he explained, and slowly raise it as evidence shows the AI delivers reliably. In a finance demonstration inside SAP S/4HANA, he took clearing proposals off the table together because they pointed the same direction, while intercompany reconciliation, a harder case, stayed pending release. “I’m still in control, but it’s actually guiding me,” he said.

On accountability, he put it flatly: “There’s always, always a human responsible, and this is a key concept. There will not be a situation where the human is not accountable.” He allowed that the accountable person can be hard to find, but “for any organization, you will always be accountable for the AI you deploy.”

SAP Supplies Business Context and Central Control

SAP’s platform, as Schleimann described it, addresses that trust problem by giving agents business context and controlling them from one place. What AI receives today, he said, is typically a narrow task, rule, or policy, which leaves out “the hard stuff, all the things that really make things work,” including human judgment and the consequences of a decision.

The knowledge graph, an ontology at the core of the SAP Business AI Platform, records how business objects connect, such as how a product SKU relates to a customer order and its billing details. Schleimann said it keeps a search within the relevant space, adds context, and validates the output, to reduce hallucination and support trust in execution.

Governance belongs in the plan from day one, he told attendees, and trust has to be centralized: “You can decentralize action, but you cannot decentralize trust.” SAP put that control into its AI Agent Hub, which covers which agents, SAP or non-SAP, are allowed to take transactions, what level of autonomy each has, who is ultimately accountable, and how each performs.

Retained Learning Is the Longer-Term Advantage

Every company will have access to the same technology, Schleimann said. Gains in productivity, speed, and quality are “merely first-order effects that we very quickly will have reaped the benefit from, and it will basically be a race to the bottom.” In his view the longer-term payoff is learning.

He believes building unique insights and retaining that learning is “the only true nature of competitive advantage.” SAP’s route is company memory, a direction it is illustrating and does not claim is generally available in every part today. It is a governed, persistent memory that captures what must happen in a process and why.

Schleimann closed by telling leaders to pick a few high-value outcomes, favor use cases that can scale, and set an owner, scope, evidence, and escalation path for every agent. “First, you build trust, then you grant autonomy,” he said.

What This Means for ERP Insiders

Bounded autonomy becomes the working limit for AI agents. Schleimann called it the best that can be achieved, with the level set for each task and trust raised slowly as evidence of reliable performance builds. He held that organizations stay accountable for the AI they deploy, and he measures maturity by the work an organization has delegated.

Governance becomes a design requirement for agents. Schleimann said action can be decentralized but trust cannot, which puts agent permissions, autonomy levels, and performance monitoring in the design from the first agent. In his description, governing from the start keeps ownership and accountability clear for each agent as more of them arrive.

Accumulated learning becomes the lasting advantage. Productivity gains reach every company with access to the same technology, which is why Schleimann placed competitive advantage in the insights an organization retains from its own work. SAP presented its route to that retention as a direction and does not claim that every part is generally available today.

This article was first published by SAPinsider.