The Case for Making Your AI Agents Less Autonomous: What Banking's Hardest Workflows Can Teach Engineering Leaders

The dominant instinct when deploying AI agents is to push autonomy as far as the model will allow. More automation means fewer humans in the loop, lower operational cost, and faster throughput. That logic holds in low-stakes environments. In regulated industries, where a single incorrect decision can trigger a compliance breach, a credit loss, or a regulatory investigation, it tends to produce systems that are impressive in demos and unreliable in production.

Companion piece to our broader work on agent architecture and operational readiness. See What Long-Running Agents Expose About Engineering Team Readiness for a practical analysis of the workflow gaps that surface when teams move from short-context AI assistance to longer autonomous agent horizons.

Most Production Agents Are Not Actually Agents

Before debating how much autonomy to give an AI agent, it is worth being precise about what most deployed systems actually are. A recent paper from CMU draws a sharp line between what the authors call "agentic" systems, whose competence resides in engineered workflows and external scaffolding, and genuinely "agentive" systems, whose capabilities arise from internalized goal structures, identity, and self-directed learning (Xing et al., arXiv 2026).

By that standard, the vast majority of enterprise AI deployments are scaffolding with a language model at the center. The model does not hold goals, it receives them. It does not regulate its own behaviour, it follows a prompt chain assembled by engineers.

This is not a criticism. Scaffolded systems can deliver substantial production value. But calling them autonomous agents creates a category error that leads engineering teams to grant them decision rights they are not architecturally equipped to exercise responsibly.

Why Banking Workflows Expose the Autonomy Problem First

Financial services workflows are useful test cases because the error cost is asymmetric. A false positive in a fraud detection pipeline may inconvenience a customer. A false negative may expose the institution to financial crime liability. A miscalculated credit decision affects a borrower's financial position and the bank's risk book simultaneously.

In these environments, the question is not whether AI can make the decision. Often it can, with reasonable accuracy. The question is who absorbs the cost when it is wrong, and whether the system can produce an auditable account of why it decided what it decided.

Full autonomy collapses that accountability chain. When a model acts without a human checkpoint, the institution owns the outcome without a clear record of the reasoning that produced it. That is an unacceptable position for any workflow that sits under regulatory scrutiny.

Calibrating Autonomy as an Engineering Decision

Treating autonomy level as a design variable, rather than a default setting, changes how teams architect agent systems. The practical approach is to map each step in a workflow against two dimensions: the reversibility of the action and the frequency with which edge cases appear.

High-Reversibility, Low-Edge-Case Steps

These are strong candidates for full automation. Document classification, data extraction, routine status updates, and threshold-based routing decisions can generally be handed to an agent without a human checkpoint, provided the output feeds into a step that a human will eventually review.

Low-Reversibility or High-Edge-Case Steps

These require a human in the loop, not as a fallback but as a designed part of the workflow. Credit decisions, sanctions screening escalations, and exception handling in payment processing all belong here. The agent's role is to prepare a recommendation, surface the relevant evidence, and present the decision in a form that makes human judgment faster and more consistent, not to replace it.

Converting Human Decisions Into Repeatable Rules

One of the underappreciated benefits of constrained agent design is what it reveals about the decision logic that humans apply. When an agent is required to present a recommendation for human approval, each approval or rejection becomes a data point. Over time, those data points expose the implicit rules that experienced staff apply intuitively.

That process of externalizing tacit judgment is valuable independently of the AI system. It produces documented decision criteria that can be reviewed, challenged, and improved. It also creates the training signal needed to gradually extend automation to cases that were previously too ambiguous to automate safely.

The trajectory this enables is more durable than attempting full automation from the outset. Teams start with high human involvement, identify the decision patterns that are consistent and well-defined, encode those as rules the agent can apply autonomously, and retain human review for the cases that remain genuinely ambiguous.

The Organisational Conditions That Make This Work

Constrained autonomy only delivers value if the human review layer is genuinely functional. A human checkpoint that is treated as a rubber stamp provides none of the accountability benefits and all of the latency costs.

This requires that reviewers have enough context to make a real decision, not just approve what the model recommended. It requires that the agent's output be structured for reviewability, presenting evidence and reasoning rather than just a conclusion. And it requires that disagreements between human reviewers and agent recommendations are logged and analysed, because those disagreements are where the system learns.

The organisational discipline required here is not trivial. But it is the same discipline that makes any high-stakes operational process reliable. AI agents do not change what good process design requires. They make the absence of it more visible, more quickly.