Scaling Agentic AI in Enterprise ERP Processes
BLP, an ETH Zurich spin-off, is converting manual ERP workflows into autonomous agent pipelines for finance, procurement, and sales. The company reports five to ten times productivity gains at clients such as BMW, Edeka, and Roche. The discussion covers the shift from AI pilots to measurable ROI, the need for agent-ready process design, and governance models that separate IT security from business-unit application ownership.
The Agentic AI Inflection Point
BLP, an ETH Zurich spin-off founded in 2019, is converting manual ERP workflows into autonomous agent pipelines for finance, procurement, and sales. Its clients, including BMW, Edeka, and Roche, report productivity gains of five to ten times in high-volume processes such as accounts payable. The strategic lesson is clear: the model layer is now a commodity, while durable value sits in the application layer, where agents execute end-to-end business processes with audit-grade reliability.
From Pilot Chaos to ROI Discipline
Large enterprises have spent the last two years launching hundreds of AI pilots. Many achieved quick wins, but few reached production maturity. A decisive shift is emerging: by 2027, boards are expected to reject token spend without measurable top-line or bottom-line impact. Companies are now buying the ten use cases that matter and outsourcing the remaining ninety to specialized platforms. This marks the end of exploratory AI and the beginning of operational AI.
Why Build Versus Buy Is Changing
Modern tooling makes it easy to build demos, but production agents require exception handling, cost control, and continuous monitoring. In accounts payable alone, BLP breaks a single workflow into roughly 160 micro agents, each handling a specific extraction, validation, or routing task. Generic agents cover common cases, while customer-specific agents encode local context and historical patterns. For audited processes, 99 percent reliability is insufficient; the target is 100 percent, with human checks phased out only after statistical confidence is proven.
Governance and Change Are the Real Bottlenecks
Legacy processes were designed for human control, not autonomous execution. They contain implicit knowledge, redundant approvals, and undocumented exceptions. The BLP framework requires making that knowledge explicit, redesigning workflows, and assigning an Application Owner in the business unit. IT remains responsible for integration and security, while domain experts manage agent context, testing, and release. This split ownership prevents ticket backlogs and aligns automation with operational reality.
Executive Takeaway
The competitive advantage of agentic AI is no longer model access. It is the ability to reengineer processes, institutionalize governance, and manage human change. Companies that treat agents as production systems, not experiments, will capture measurable ROI before the market matures.
Key insights
-
The model layer has become a commodity, while the application layer now determines enterprise value. BLP builds end-to-end agent workflows on top of commodity LLMs rather than proprietary models. This reduces dependency on model vendors and accelerates deployment.
Impact: Enterprises can lower AI risk and cost by standardizing on commodity models. Competitive advantage shifts to process integration, governance, and domain-specific automation.
-
AI pilots are reaching a maturity wall. Companies have launched hundreds of experiments but struggle to convert them into reliable, cost-efficient production systems. The market is moving from exploration to measurable ROI.
Impact: Boards will demand proven top-line or bottom-line impact by 2027. Firms that focus on ten high-value use cases will outperform those chasing broad pilots.
-
Core financial processes require near-perfect reliability, not 99 percent accuracy. BLP decomposes accounts payable into roughly 160 micro agents and uses customer-specific context to handle exceptions. Human checks are phased out only after statistical confidence is established.
Impact: Audited workflows become viable for autonomous execution. This opens large cost-reduction opportunities in finance, procurement, and sales operations.
-
Legacy processes are not agent-ready because they encode implicit knowledge and human control points. Reengineering requires making exceptions explicit and redesigning approval chains. Domain experts must own agent context and testing.
Impact: Organizations that reengineer processes will unlock durable automation gains. Those that only layer AI on old workflows will hit reliability and adoption ceilings.
-
Governance must split between IT and business units. IT should own integration and security, while Application Owners in the business manage agent tuning, validation, and release. This prevents ticket bottlenecks and aligns automation with operational needs.
Impact: Faster iteration and clearer accountability improve agent performance. It also reduces resistance by giving domain experts ownership of the technology.
Action items
-
Select ten high-impact use cases with clear ROI and audit requirements. Prioritize processes such as accounts payable, procurement, or sales operations where volume and exception patterns are well defined. Avoid broad pilot programs that lack production criteria.
Impact: Focuses investment on measurable outcomes. Reduces pilot fatigue and accelerates path to production.
-
Assign an Application Owner in the business unit for each agent workflow. This owner should manage context, exception handling, testing, and release decisions while IT retains integration and security. Define clear escalation paths for model or context changes.
Impact: Speeds iteration and improves reliability. Aligns automation with domain expertise and reduces IT bottlenecks.
-
Reengineer legacy processes before deploying agents. Document implicit knowledge, map exceptions, and redesign approval chains for autonomous execution. Use validation sets to test every model or context change.
Impact: Raises reliability from demo level to production level. Enables audited processes to run with minimal human intervention.
-
Communicate a concrete role transition plan for affected employees. Identify specific value-adding tasks that automation will free, and involve executives as visible sponsors. Use early wins to build trust and reduce resistance.
Impact: Improves adoption and reduces sabotage risk. Turns change management into a competitive advantage.
-
Build continuous monitoring and regression testing into the agent platform. Track exception rates, cost per transaction, and reliability against validation sets. Automate retesting when models or context change.
Impact: Protects production performance as models evolve. Provides the evidence needed for board-level ROI reporting.
Quotes
“The model layer is now a commodity, while the application layer is where durable enterprise value is created.”
“By 2027, boards will no longer accept AI pilots that generate token spend without measurable return on investment.”
“The technology is no longer the problem; the real challenge is adapting people, processes, and governance.”