Putting It into Practice

2026

Operate and Improve

Run agents as accountable transformation products with ownership, monitoring, incidents and controlled evolution.

An agent that has moved into daily project work is a production service, even if it has no direct access to the productive S/4HANA system.

Establish ownership

Each agent needs a business or transformation product owner, a technical owner and a control owner. Together they manage value, performance, cost, risk and lifecycle decisions.

Ownership should cover model and prompt changes, skills, knowledge sources, tools, permissions, policies, evaluations and user communications.

Monitor what matters

Operational monitoring should combine:

  • availability, latency and cost.
  • task success and human override.
  • evidence quality and unsupported claims.
  • policy violations and attempted unauthorized actions.
  • performance by process, language, country and project phase.
  • model, knowledge and tool versions behind each material output.

Aggregate quality can hide weak performance in a critical minority of cases.

Manage change

Every material change should trigger proportionate regression evaluation. This includes a new model, revised system instruction, changed skill pack, new knowledge source, altered permission or updated SAP content.

Release records should make it possible to reproduce important outputs and understand why behavior changed.

Prepare for incidents

Agent incidents include more than outages. They may involve an incorrect recommendation, disclosure of restricted information, inappropriate action, systematic bias or loss of provenance.

Define detection, containment, notification, investigation and recovery. The organization must be able to reduce autonomy or disable a capability quickly.

Improve through feedback

Human corrections are valuable only when they become structured learning. Capture the type of failure, relevant context, reviewer rationale and approved outcome. Use this to improve knowledge, rules, prompts, tools and evaluation sets.

The mature operating model is not self-running. It is continuously governed and deliberately improved.