AI agents don't replace workflows. They expose the ones that were broken.

Most conversations about AI agents start with the model: which one, how capable, how expensive. In practice, the first serious obstacle is often the workflow the agent is supposed to join.

When a team asks for an agent to handle a process, they may be describing work that already depends on people compensating for something — missing data, unclear ownership, exceptions nobody documented. The agent inherits these problems, but introduces risks of its own.

The process you think you have

On paper, a typical approval flow looks linear: a request arrives, it is checked, and it is approved or rejected.

Ask the people doing it and a different picture may appear. Requests arrive in three formats. Some need a phone call to clarify. A particular colleague knows which suppliers require additional verification.

None of that knowledge is necessarily visible in the system. It lives in habits, inboxes and memory.

An agent operating against an incomplete process specification may make incorrect decisions, escalate too frequently or act without sufficient context. And even with a well-defined process, AI models can misunderstand instructions or produce unreliable outputs.

The problem is therefore not simply making the agent more capable. It is designing a process in which its decisions can be evaluated and controlled.

An AI agent can reveal where a process lacks clarity. It cannot, by itself, decide what the business rules should be.

Where the real work is

The useful first step is not prompt design. It is mapping where humans intervene today, and why.

These interventions often fall into four categories:

• —Missing information — the input does not contain what the decision requires.
• —Ambiguous rules — a policy exists, but exceptional cases are resolved informally.
• —Broken handovers — work stalls between teams or systems without clear ownership.
• —Trust gaps — people repeat checks because the underlying information has previously been unreliable.

Each category requires a different response.

Missing information may call for better data collection. Ambiguous rules require business decisions. Broken handovers may need integration or workflow redesign. Trust gaps demand better validation, traceability and accountability.

None of these problems is solved simply by selecting a more powerful model.

Automation as a diagnostic

Treating an initial agent deployment as a controlled diagnostic changes the objective.

Instead of asking how much work the agent can immediately replace, ask what information, permissions and rules it needs to operate reliably.

A practical sequence is:

1. 01Map the existing process. Document inputs, decisions, exceptions and responsible people.
2. 02Run in observation mode. Evaluate the agent alongside the existing workflow without granting unnecessary authority.
3. 03Record uncertainty and failure. Identify missing context, incorrect decisions, unnecessary escalations and model-specific errors.
4. 04Improve the system. Address data quality, process rules and integration gaps.
5. 05Expand gradually. Increase automation only when evaluation results, safeguards and operational monitoring justify it.

The distinction between process failures and model failures matters. Both must be measured.

What responsible autonomy requires

Moving from an assistant that recommends actions to an agent that executes them changes the risk profile.

The architecture should define which actions require human approval, which systems the agent can access, and how failures are contained.

Useful safeguards include least-privilege permissions, validation of tool inputs, bounded execution, audit trails, human escalation and mechanisms to stop or reverse actions where possible.

An agent should not receive unrestricted access simply because it performs well in a demonstration.

What this means in practice

The long-term value of AI agents depends on the quality of the systems around them.

Clear ownership, reliable data, well-defined interfaces and measurable operating rules make automation more useful — and often improve the business even before AI is introduced.

The model matters. But the workflow, architecture and safeguards determine whether the model can contribute reliably.