SAP Said the Autonomous Enterprise Is Here. Redwood’s CPO Says the Hard Part Starts Now.

SAP Sapphire 2026 gave the autonomous enterprise a more concrete shape.

At the event, SAP positioned agentic AI around Joule Work, 200-plus agents, Company Memory, and an Anthropic partnership bringing Claude into SAP Business AI across HR, procurement, and supply chain.

The message was clear: AI is no longer just a copilot for enterprise software. SAP wants it to act across business processes. That closes one question, but opens another.

If AI can understand enough enterprise context to recommend actions, what ensures those actions execute with the determinism, auditability, and control that businesses require?

That is the problem Charles Crouchman is focused on. As Chief Product Officer at Redwood Software, he argues that the next enterprise AI challenge is governed execution.

The Question Has Shifted From Intelligence to Execution

“When I joined Redwood in 2024, the early agentic wave was focused on the intelligence layer: which model, which framework, which copilot experience,” Crouchman says. “The question customers were asking was, ‘Can AI understand my business?’ By 2025 and into SAP Sapphire 2026, the question has moved to, ‘Can AI actually execute inside my business?'”

Redwood has spent more than 30 years building workload automation for enterprise processes where timing, sequence, and auditability are built into the design, including financial close, MRP runs, billing cycles, and supply chain orchestration. The agentic orchestration framing is new, but the underlying automation problem is not.

“Joining as CPO meant I was inheriting something real to build on,” Crouchman says.

That inheritance now has a different product context. Agentic AI does not remove the deterministic business logic behind enterprise automation. It creates pressure to make that logic available to agents without loosening control over how work is executed.

“Our roadmap has shifted accordingly from extending our workload automation platform toward a full agentic orchestration platform — with MCP server support already released, A2A multi-agent orchestration in tech preview and Agentic Studio and Agentic Workflows on the near-term roadmap,” Crouchman says. “The opportunity has gotten larger and more urgent since I joined.”

Why Agentic Failures Don’t Announce Themselves

The execution gap starts when AI moves to production. “In a demo, the AI performs beautifully on a clean, isolated scenario,” he says. “A Joule agent recommends an action. The action executes. The outcome is correct. Everyone in the room is impressed.”

Production works differently. A global financial close or production planning cycle may involve thousands of steps across ERP, cloud applications, and legacy systems. Those steps are connected by timing, dependencies, approvals, exceptions, and downstream data requirements that do not always show up in a controlled demonstration.

Crouchman’s concern is that probabilistic AI introduces variability into that chain. The individual recommendation may look reasonable. The first action may succeed. And the process may continue long enough for the problem to move downstream.

“The system doesn’t fail loudly,” he says. “It continues. The inconsistencies accumulate.”

That is what makes the execution gap difficult to manage. A failed job can trigger an alert. A broken integration can stop a workflow. An agentic process that produces a plausible but inconsistent action may not stop the chain at all. At that point, the productivity gain that AI was supposed to create can turn into remediation work.

This is the risk that becomes more important as SAP customers move from isolated AI use cases toward multi-agent environments. SAP’s 200-plus agents show how quickly the operating model is expanding; the production question is where that control should sit before agentic actions reach systems of record.

Where SAP Ends and Redwood Begins

SAP has positioned AI Agent Hub as a command center for managing SAP and non-SAP agents, while Company Memory is designed to give agents access to process knowledge, policies, procedures, and decision traces. The Anthropic partnership also gives Claude a direct connection into SAP Business AI Platform through MCP.

Crouchman does not argue that SAP lacks a governance layer. He places Redwood at a narrower handoff point: where agentic intent has to become controlled workload execution across ERP, cloud applications, and legacy systems.

“SAP’s native architecture — Joule Work, Joule Studio, AI Agent Hub, accumulated business logic — is an intelligence layer,” he says. “It allows users to state a desired outcome in natural language, coordinates the right combination of agents and workflows to pursue that outcome, and gives developers a model-agnostic canvas for building agentic applications. That is valuable innovation.”

The distinction, in his view, is what happens next. When an agent recommends a financial action, flags a supply chain exception, or initiates a process that touches systems of record, the work still has to execute with accuracy.

Redwood’s layer begins at that handoff point, Crouchman explains. “Through our MCP server, agents built in Joule Studio or any other framework can invoke existing RunMyJobs workflows, jobs and enterprise connectors as governed tools — with hard deterministic guardrails applied before outputs reach systems of record, not after,” he says.

In lower-risk workflows, SAP-native capabilities may be enough. In more complex hybrid estates, Crouchman sees a different requirement: an execution layer that constrains agentic action before it changes the business record.

Solving the Cold-Start Problem With Existing Automation

Many SAP customers have years of automation logic already embedded across job schedules, finance workflows, ERP integrations, data movements, and exception handling. Crouchman describes this as the cold-start problem.

“The biggest hidden cost of deploying agentic AI is teaching the agent what your business actually does,” he says. “The decades of mission-critical logic encoded in existing automation cannot be reconstructed by AI prompting.”

That is where Redwood’s approach becomes most specific. RunMyJobs, Redwood’s workload automation and orchestration platform, gives agents a governed path into existing workflows, turning the automation layer into reusable execution logic, rather than leaving each new agent to relearn or rebuild what the business has already encoded.

The value is most concrete in hybrid SAP estates, where business processes often span SAP S/4HANA, older ERP environments, cloud applications, and non-SAP systems. In those environments, the automation library becomes a record of how work moves across the enterprise: the dependencies and controls agents would otherwise have to relearn.

With RunMyJobs, Crouchman says, “the work is already done. Existing workflows become governed tools that agents can invoke immediately, acting correctly from their first interaction.”

The value proposition is investment protection as much as AI enablement. If the automation layer already encodes timing, dependencies, approvals, and exception paths, agentic AI creates a new reason to expose that logic safely rather than replace it.

When SAP-Native Capabilities Are Enough — and When They Are Not

Crouchman’s decision framework starts with a practical question: what happens if the process gets the answer wrong?

SAP-native capabilities may be enough for workflows where the risk is limited, the process stays mostly inside SAP, or the action can be reviewed before it changes a business record. Crouchman draws the line around processes where timing, sequence, auditability, and cross-system execution are part of the operating requirement.

He points to four conditions: long-running, high-volume work across ERP, cloud, and legacy systems; regulatory or audit pressure; multi-agent environments where domain agents may compete over shared data or resources; and existing automation logic that should become available to agents as governed execution logic.

“Take your three or four most critical business processes — financial close, MRP, order-to-cash, supply chain — and ask whether you would accept a probabilistic ‘most likely correct’ output from an AI agent operating inside those processes without a deterministic guardrail,” he says. “If the answer is no, that’s where Redwood fits.”

Redwood’s case is strongest where autonomous action meets high-volume, regulated, or cross-system work. Crouchman says the company is building around that requirement, with MCP server support already available in RunMyJobs, A2A multi-agent orchestration in tech preview, and Agentic Studio and Agentic Workflows on the near-term roadmap.

The autonomous enterprise may be here. The execution model is still being decided.

What This Means for SAPinsiders

• Hybrid estates are the real proof point. The strongest validation for the governed execution argument will come from production deployments, not demos. Enterprises operating complex hybrid SAP landscapes are where the execution gap either materializes as described or it doesn’t.
• The cold-start problem is an operational challenge. Existing RunMyJobs automation libraries encode years of exception handling, dependency mapping, and error resolution. Exposing those workflows to agents as governed tools can transfer knowledge about where enterprise data is most likely to break.
• The probabilistic test clarifies deployment risk. Crouchman’s decision framework gives CIOs a practical way to identify where agentic AI needs stronger controls before execution. That makes autonomous AI adoption more disciplined by matching governance requirements to process risk, rather than applying one model everywhere.