Automation Is the Destination, Not the Starting Point
A pattern we’ve seen repeatedly in enterprise AI initiatives is an early push toward full automation using agentic workflows.
The intent is understandable. When teams examine existing workflows, it’s tempting to identify steps that appear low-cognition and assume they can be automated end-to-end through agents that plan, decide, and iterate on their own.
In practice, this framing often stalls adoption—not because the technology falls short, but because the work itself is mischaracterized.
Problem: automation-first breaks in real workflows
Many enterprise workflows look routine when described at a distance. Once operationalized, they behave very differently.
What appears deterministic usually contains judgment. What looks repetitive often hides edge cases. Humans resolve these implicitly, without articulating how or why.
When those workflows are translated into systems, that implicit knowledge is lost. Agentic workflows are then built on a simplified abstraction of the work, not the work itself.
This gap shows up quickly in production.
Why this happens
Three factors tend to dominate.
First, the long tail is larger than expected. Edge cases matter and require near-total coverage for full automation. Missing even a small fraction can be unacceptable in operational systems.
Second, humans under-describe their own work. Subject-matter experts explain the common path and official rules. They rarely capture informal heuristics, judgment calls, or cases where “this doesn’t feel right.”
Third, agentic systems are sensitive to context. LLMs that power agentic systems can trip into hallucination, which continues to be a top operational challenge. In a fully automated setup, there is no backstop when this happens—no mechanism to detect, correct, or recover.
When systems are built on these abstractions and humans are removed early, there’s nothing left to absorb reality.
A better starting point: co-pilot systems
Where possible, we’ve found co-pilot systems to offer the best combination of value and reliability, while still laying a path toward increased automation.
By co-pilot, we mean a pair-programming-style model where a human is involved either in the loop or on the loop. The human and the machine bring different strengths and weaknesses, and together act as a backstop for each other. Machines do not suffer from fatigue or attention drift; humans generalize better to novel situations and ambiguous context.
The mechanics vary—human review, confidence-based escalation, selective intervention—but the posture is consistent. Humans remain in the loop while the system learns.
This shifts failure modes from catastrophic to recoverable and learnable.
Co-pilots change the workflow itself
An important implication is that co-pilot systems are not just automation layers. They often require rethinking the workflow altogether.
Many human workflows are optimized around human limitations: careful reading, manual verification, repeated checks. When a machine enters the loop, those constraints change.
The human role moves away from execution and toward supervision and sense-making. The focus shifts from individual decisions to patterns, confidence thresholds, and aggregate signals that indicate whether the system’s behavior still makes sense.
This redesign does not emerge if AI is bolted onto existing steps. It requires asking a different question: how would this workflow be designed if a capable machine were present from day one?
What this enables over time
Keeping humans in the loop changes the trajectory of the system.
Edge cases become learning signals rather than blockers. Hidden complexity surfaces incrementally. Trust is preserved because failures are caught early.
As coverage grows, the system does more. Human effort shifts toward oversight and refinement. Automation increases, but without forcing it prematurely.
Production data will always expose more variance than evaluation environments. Designing for that reality allows agentic systems to evolve instead of breaking.
Full automation as an outcome
The systems that compound value most effectively are not those that aim for full automation from the start. They are the ones designed to adapt.
Full automation, when it happens, is an outcome of learning and redesign—not a prerequisite for progress.
That framing has proven far more durable in practice.