Two companies. One figured out which layer to own. One hit the wall. The IBM i operator who reads this correctly is closer to the right answer than almost anyone building from scratch.
This month, Social Capital published an 84-page primer on the agentic stack, in collaboration with Lederle Capital. Read it not as a technology overview — read it as a value capture map. Five layers. One sentence buried in the synthesis that changes how IBM i operators should think about their position:
The intelligence layer — the models everyone is racing to adopt — is becoming infrastructure. Switching costs approach zero. Capabilities converge. The layer that resists commoditization is not the one everyone is talking about.
The agentic stack has five layers. Each enables the next. But they do not accrue value equally — and understanding which layer you already own determines what to build.
The orchestration layer is where the same model becomes either far more valuable or far more expensive depending on what's underneath it. Domain knowledge encoded into agentic behavior. Business rules made machine-readable. Workflow logic that cannot be replicated without the process depth that underlies it. That is why it resists commoditization. Two companies prove this in different directions.
In 2024, Klarna deployed an AI customer service agent that handled 2.3 million conversations in its first month — equal to two-thirds of all customer service chats. The intelligence layer was real. The results were real.
Then the retreat. Complex cases, sensitive issues — routed back to humans. Not because the model failed. Because the orchestration layer wasn't fully encoded. The rules about which cases required human judgment existed in experienced agents' heads — not in machine-readable form. Klarna later refined the model further, keeping AI on routine repeatable cases and re-routing sensitive ones to humans permanently.
Shopify made every store agent-ready by default. Co-developed the Universal Commerce Protocol (UCP) with Google. Walmart, Target, Etsy, American Express, Visa, Mastercard, and Stripe signed on. The results compounded before the default activation was even fully live.
Shopify didn't build a better model. It made its domain data — product catalogs, pricing logic, merchant relationships, checkout rules — agent-callable at scale. The intelligence layer is the same frontier model everyone else uses. The orchestration layer is what nobody else has.
Klarna built the bridge without the road. Shopify built the road first and let the bridge follow. The difference is not model quality — both use frontier models. The difference is whether the organization's domain knowledge was encoded into the orchestration layer before agents were deployed on top of it.
A major IBM i organization is doing something most of its peers don't know is possible yet. Not because it acquired new technology. Because it understood which layer of the agentic stack it already owned — and built from there.
The dominant narrative says IBM i organizations need to modernize before they can deploy agentic AI. The five-layer map says the opposite. The orchestration layer — the one Klarna didn't fully encode, the one Shopify invested in first — is not built in a sprint. It is the accumulated architecture of how work actually gets done.
The rules, the exceptions, the integrations, the sequence dependencies, the validation logic. RPG programs that contain business logic. DB2 tables with referential integrity. CL procedures that encode workflow. Trigger programs that enforce rules at the data layer. Decades of production-hardened business knowledge — not in documentation, in code.
The mechanism: Project Bob — a structured approach to surfacing the business logic embedded in IBM i RPG and DB2 — makes that logic agent-accessible via MCP. The IBM i doesn't become a migration target. It becomes the intelligence layer — the encoded domain knowledge the agentic stack runs on top of.
The unnamed organization proves this is not theoretical. An account event fires in the core IBM i system. The agent evaluates immediately against rules that have governed that organization's operations for years. The right action fires — or doesn't. No campaign was built. No human assembled a list. The orchestration layer that made it possible wasn't new. It was already there.
Agents own execution. Humans own judgment. IBM i underneath all of it.
The crawl/walk/run discipline — starting contained, with humans approving everything, proving the loop before expanding autonomy — is what keeps the transformation from getting ahead of the organization's readiness to govern it. Installment 3 maps the technical path. The LoanIQ demo makes it visible.
The map is published. The proof cases are real. The unnamed organization is building. The question for every IBM i operator is not whether to move — it's whether you know what you already own.
The four questions above map the orchestration layer. The next question is whether your full agent surface — discovery, comprehension, access, action, governance — is ready for the traffic that's already arriving.
What Does Agent-Ready Actually Look Like? → Five-Layer Reference
Klarna is a consumer lender. Their AI agent touched credit-related conversations — payment arrangements, dispute resolution, hardship cases. The retreat to humans wasn't just an operational limitation. In a regulated lending context, it's a compliance and fair lending event waiting to be documented. The companion piece applies the orchestration layer argument to the sector where the wall has regulatory consequences, not just operational ones.
Read the Consumer Finance Companion → consumerfinance.ai