Enterprise operations teams have an evidence problem. They are asked to grant autonomy to AI agents that have no track record in their specific workflows. The result is binary: either the agent is blocked from deployment entirely, or it is pushed into production without controls and breaks under real conditions.
The fix is a phased model in which responsibility grows only after real results. It has four steps: understand the work (Observe); the system suggests what to do but does not take action (Shadow mode); a person reviews and approves each action (HITL); then routine, proven actions run automatically within agreed limits (Autopilot). No phase is skipped. No authority is assumed.
Understand the work Observe
The constraint map is the deliverable. The agent has read-only access to operational data: tickets, logs, handoff records, exception queues. It traces the workflow end to end, identifies where work stalls, and quantifies the cost of each bottleneck. It makes zero recommendations. It triggers zero system changes.
Observe produces a diagnostic report containing: queue depths at each workflow stage, exception rates over 30 days, average time-to-resolution per category, and the cost of manual intervention per incident. This report gives the operations team and the engineering team a shared, data-backed picture of the current state. Before either side discusses automation, they agree on what the workflow actually looks like.
What the buyer sees: A constraint map and a diagnostic report. Queue depths at each workflow stage, exception rates over 30 days, cost of manual intervention per incident, and average resolution time by category. Zero automation recommendations.
Suggest, don’t act Shadow
Over two to four weeks, the scorecard accumulates. The agent generates recommendations for each decision point in the workflow but does not execute any of them. Every recommendation is compared against the decision the human operator actually made, and the comparison is scored for alignment and outcome.
The scorecard surfaces three things: the overall match rate between agent and operator, a divergence analysis showing where and why the agent chose differently, and flagged edge cases where the agent and the operator disagreed on high-impact decisions. When the agent recommends a different action and the outcome is better, that is evidence to escalate. When the recommendation performs worse, the agent's decision model gets adjusted before it gains more authority.
What the buyer sees: A recommendation scorecard with match rate, divergence analysis by decision category, and edge-case flags. Each divergence includes the agent's reasoning, the operator's reasoning, and the actual outcome.
Your team decides Human approval
The approval queue is the interface. Each item contains the agent's recommended action, the evidence supporting it, the predicted outcome, and the agreed recovery path where the connected system supports one. The designated operator reviews the queue and approves, modifies, or rejects each action. Approval rules are defined before HITL begins: which action types the agent can propose, the escalation path for edge cases, and the fallback if the approver is unavailable.
HITL tests the agent's decision-making under real conditions. Actual consequences, real time pressure, edge cases that did not appear during Shadow. It also gives the operations team direct experience with the agent's reasoning. Operators who have already reviewed the Shadow scorecard know the agent's strengths and weak spots before they approve their first action.
What the buyer sees: An approval queue where each proposed action includes the agent's reasoning, the supporting data, the predicted outcome, and the configured recovery path. Approver decisions are recorded according to the audit requirements agreed for the engagement.
Automate what has proved safe Controlled automation
Controlled automation runs inside defined boundaries. The boundaries are explicit, documented, and derived from the evidence accumulated across Observe, Shadow, and human-approved action. Outside those boundaries, the configured response is to stop and escalate to a human operator. Monitoring and the response to a breached threshold, including whether demotion is automatic or operator-controlled, are agreed for each deployment.
Supervision shifts from per-action approval to boundary enforcement and performance monitoring. The system operates within its agreed limits and records the actions, boundary checks, and escalations required by the customer's audit design.
What the buyer sees: The monitoring and evidence views agreed for the deployment, such as action logs, error rates by boundary, exception budget consumption, and responsibility changes. Recovery depends on the connected system and the procedure configured for each action type.
What happens when you skip phases
Most failed AI agent deployments follow the same pattern. The agent is deployed without evidence from Shadow or HITL. It produces inconsistent results in production. Correct on routine cases, wrong on edge cases that were never tested. The operations team loses confidence after the second or third incident. Support escalations increase. Within weeks, the project is shelved and the team reverts to manual processing.
The cause is the same every time: the agent was given authority it had not earned. Skipping the Observe phase means the team has no shared baseline to measure against. Skipping Shadow means there is no scorecard to predict where the agent will fail. Skipping HITL means the first time a human sees the agent's reasoning is after it has already taken a wrong action in production. Each phase exists to prevent a specific category of failure.
Questions to ask before choosing a platform
If you are evaluating AI agents for operations, these questions separate Proxima's approach from RPA vendors, self-serve tooling, and consulting engagements:
- Does the agent build a constraint map before it automates anything? RPA tools record clicks. Self-serve platforms skip diagnosis entirely. Proxima maps the workflow and quantifies bottlenecks before any automation is discussed.
- Can you compare agent recommendations against human decisions for weeks before the agent takes a single action? Most platforms offer a toggle between manual and automatic mode with no scorecard in between. Proxima runs a scored Shadow phase with divergence analysis.
- Does the approval queue show the agent's reasoning, evidence, and predicted outcome for each action? Consulting firms deliver slide decks. Self-serve tools show raw model outputs. Proxima structures each recommendation so the approver can evaluate it in under 60 seconds.
- What happens when an action category crosses its agreed error threshold? Confirm whether the response is an automatic stop, a move back to human approval, or an operator-controlled change, and test that path before controlled automation begins.
- Is the full evidence trail, covering Observe report, Shadow scorecard and HITL logs, available to your team at all times? If the vendor owns the evidence, you are trusting the vendor. If your team owns the evidence, you are trusting the data. Proxima gives your team full access to every phase output.