According to SAP’s Michael Ameling, CPO of SAP Business Technology Platform, 81% of organizations now have a detailed AI strategy, yet only 12% to 16% actually reach AI-driven execution. The gap is not about the quality of AI-generated code. It stems from enterprises underestimating what it takes to operationalize that code: data readiness, system integration, governance, and years of maintainability inside environments that were never built to accommodate autonomous agents.
For BFSI and collections operations, this distinction between generation and operationalization is the entire game. A model that drafts a compliant collections script or generates a workflow update is solving a narrow problem. Running that same logic reliably against live account systems, transaction records, and legacy infrastructure that predates any AI initiative is a categorically different challenge, one involving latency, cost, and system load that a one-time code output never has to face. As Ameling frames it, AI amplifies an organization’s existing data and process maturity rather than substituting for it, which means a BFSI institution with fragmented data stores will see that fragmentation surface inside its AI agents, not disappear because of them.
What gets missed in most enterprise AI conversations is that governance requirements change entirely once an agent moves from producing recommendations to executing actions. Two distinct models now need accounting for: agents that inherit a user’s permissions and act on their behalf, and agents that operate under their own role-defined identity, similar to an automated employee. Both require the same underlying visibility, meaning operators can see which agents exist, what systems they can touch, and what they’re authorized to do. Equally underappreciated is that technical evals, which check whether an agent produces consistent outputs, are necessary but insufficient. Business evals, which measure whether an agent is actually moving the metric it was deployed to improve, are the harder and more often skipped layer, particularly because traditional dev-test-production validation breaks down when a model behaves differently against live data than it did against test data.
This aligns with what we see in production deployments across collections: the organizations that succeed are not the ones with the most sophisticated code generation, but the ones that treated integration and governance as prerequisites rather than afterthoughts. Operationalizing agentic AI in a regulated environment means establishing federated data access and a harmonized process layer before any agent begins executing, not after a pilot proves the model can generate a good response. It also means accepting that observability and evaluation for agents look fundamentally different from the QA cycles engineering teams have relied on for decades. The differentiator going forward will not be tooling access, since most enterprises now have comparable AI capability, but how effectively an institution encodes its own domain expertise, whether that’s a bank’s risk logic or a collections team’s escalation policy, into the systems built on top of it.