Autopilot vs Copilot
The Levels of Autonomy for AI Customer Support
The choice is not AI or no AI. It is how much the AI is allowed to do on its own, and on which tickets. Here is the five-level framework for deciding, and for graduating from copilot to autopilot safely.
By Palak Dalal Bhatia, CEO & Co-founder · Last updated June 7, 2026
What Are the Levels of Autonomy in AI Customer Support?
The model borrows a discipline from self-driving cars, where SAE International defines six levels of driving automation. The value is not the cars. It is the rigor of naming exactly how much the machine does, exactly what the human is responsible for, and exactly when control changes hands. Support needs that same rigor, because a password reset and a billing dispute on a regulated account do not belong at the same level of autonomy, even inside the same company.
Copilot vs Autopilot: The Core Distinction
Strip away the marketing and one line separates the bottom half of the ladder from the top half. A copilot suggests, and a human acts. An autopilot acts, and a human supervises.
A copilot sits next to your agent. It drafts a reply, summarizes the thread, surfaces the right knowledge article, and suggests the next step. The agent edits if needed and clicks send. The human is in the loop on every response. That is what IrisAgent's agent assistdoes, and it is the right starting point because a bad suggestion gets caught before it ever reaches a customer.
An autopilot sits in front of your customer. It reads the message, retrieves a grounded answer, decides whether it is confident enough, and sends the response itself. That is what IrisAgent's autopilotdoes, and it is where real deflection and cost reduction come from, because the human is removed from the high-volume, low-complexity work entirely.
The reason this matters: copilot and autopilot are two different operating models. A copilot makes your existing agents faster. An autopilot changes how many agents you need and what they do all day. Confusing the two is how teams either underinvest or overreach.
The Five Levels, Explained
The level is chosen per intent. A mature queue runs as a portfolio, with some intents at L4 and others still at L1.
L0: Manual
No AI touches the ticket. A deliberate choice for a narrow set of tickets where a wrong answer is catastrophic and volume is low enough for humans to carry, not a default for the whole operation.
L1: Copilot (suggest-only)
The AI assists the agent but never speaks to the customer. It drafts, summarizes, and recommends; the agent approves everything. The safest first step: it speeds up the team and generates the data you need to graduate.
L2: Supervised autopilot
High-confidence answers on low-risk intents auto-send. Anything uncertain or sensitive is drafted and held for approval. This is where deflection starts to show up, and where most scaling teams should target first.
L3: Conditional autonomy
Within defined domains the AI resolves end to end with no human in the loop, including taking actions like issuing a refund or tracking an order. Outside those domains it escalates. This is where managed resolution operates.
L4: Full autonomy
The AI resolves the bulk of incoming volume across channels, and humans handle only exceptions: the novel, the high-empathy, the high-value. Not no humans, but humans doing only the work that needs human judgment.
How to Graduate Safely from Copilot to Autopilot
The mistake that kills autonomy rollouts is treating the level as a setting you flip rather than a rank you earn. Here is the playbook.
Start at L1 on a real queue
Put the copilot in front of agents on live tickets. Let it draft and let agents edit. What they accept, rewrite, and discard is your accuracy signal.
Pick the first intent to promote
Choose one high-volume, low-risk, policy-bounded intent: order status, password reset, basic billing FAQ. Common enough to matter, safe enough that a rare miss is recoverable.
Move that intent to L2
Let the AI auto-send only above a confidence threshold you set; everything below drafts for review. Use confidence-based routing so the line is enforced automatically.
Encode the guardrails
Write down what the AI may and may not do per intent, then machine-enforce it with smart operating procedures rather than tribal knowledge.
Measure for two weeks, then widen
Watch resolution rate, CSAT on AI-handled tickets, and escalation accuracy. If they hold, promote the next intent. If CSAT drops more than ten points, pull back.
Repeat toward L3 and L4
Autonomy spreads one proven intent at a time. The queue graduates as a portfolio, each intent measured separately.
What Each Level Requires Under the Hood
Autonomy is only as safe as the architecture beneath it. The higher you climb, the more these four foundations have to be in place.
Common Mistakes When Deploying Autonomous Support AI
How IrisAgent Supports Each Autonomy Level
IrisAgent runs the whole ladder rather than locking you into one rung. Agent assist covers L1, supervised autopilot covers L2, and managed resolution covers L3 and L4, all on the same grounded foundation so the experience stays consistent as you graduate. Because every answer is validated against your own content, raising the autonomy level does not raise the hallucination risk: the system escalates instead of guessing.
Deployments like Dropbox (160,000-plus tickets managed with AI) and Zuora (10x faster resolution) run higher autonomy on repeatable intents while keeping humans on the exceptions. And because IrisAgent does not charge per resolution, moving an intent up the ladder lowers your cost instead of raising your bill.
Go Deeper on the Agentic AI Stack
operations
Any questions?
We got you.
