"Autonomous AI" gets written about as if it's binary — either the AI does things on its own, or it doesn't. The reality of building a support agent that actually earns trust is far more granular than that.

The trust problem

The central challenge of agentic support isn't intelligence. Modern language models can read a support message, identify the issue, find the relevant knowledge base article, and draft an accurate reply in under a second. The intelligence is table stakes.

The challenge is trust — specifically, knowing when the agent's judgment is reliable enough that you don't need to check it before it acts.

A useful framework: action risk vs. action confidence

Not all agent actions carry the same risk. Updating a CRM property is low-risk and easily reversed. Sending an email to a customer is higher-risk and not reversible once opened. Creating an external ticket in Jira is somewhere in between.

Confidence is also not uniform. An agent that correctly classifies 95% of billing queries might be right only 70% of the time on ambiguous churn signals. Auto-approving based on a single confidence threshold treats these very differently-shaped decisions as equivalent.

The right model is a two-dimensional grid: action risk × action confidence. Low-risk, high-confidence actions can be automated. High-risk or low-confidence actions need a human.

What this looks like in practice

For a support intelligence agent, the matrix shakes out roughly like this:

Auto-execute (low risk, high confidence):

  • Logging a KB gap when a question can't be answered
  • Updating a customer's health score in the CRM
  • Creating a Jira ticket for a confirmed bug report
  • Sending a Slack notification to the assigned support owner

Require approval (higher risk or lower confidence):

  • Sending a draft reply to a customer
  • Triggering proactive outreach to a churn-risk account
  • Closing a ticket as resolved

Always human (high stakes, complex judgment):

  • Offering a refund or discount
  • Escalating to legal or compliance
  • Any communication about a service incident

Selective autonomy is a dial, not a switch

The approval threshold isn't fixed — it's set by the operator and adjusts over time as the agent proves its judgment. A team that starts with everything requiring approval and sees consistently accurate recommendations will, over time, move more action types into the auto-execute category.

This is what "selective autonomy" actually means: not full automation, not full human-in-the-loop, but a deliberately managed expansion of the agent's operating envelope, calibrated to observed performance on each action type separately.