what percentage of tickets can an AI agent resolve on its own?

Most of the messages a business receives over messaging aren't complex cases: they're repetitive questions — where's my order, what size is left, what are your hours, how do I make a return. In a well-configured operation, an AI agent can take between 70% and 90% of that kind of inquiry without a human stepping in. It's the range heysteff states for its own customers, and it is deliberately a range, not a single number: each business's real percentage moves within that interval according to its own operation, not according to a marketing promise.

Beyond the repetitive-inquiry range, heysteff publishes an internal average of its own: 87% of resolutions without human intervention, measured across active workspaces. It's an average, not a guaranteed minimum for any new business — some workspaces sit above it, others below, depending on how much configuration and curation work they've invested. The exact definition of what counts as "resolved without a human" is documented in the methodology, and it's worth reading before comparing that number against another vendor's.

No vendor controls your final percentage: it's largely determined by your own configuration. Four factors weigh more than any other:

• Knowledge base quality. An agent only resolves well what it has documented. If your policies, your catalog and your FAQs are complete and up to date, the agent resolves more; if they're half-filled, it resolves less, no matter how good the AI model underneath is.
• Access to real data via tools. An agent that only chats with static text can't confirm an order's status or real stock; one connected through native integrations or a tools builder to your system can. That difference directly moves how many cases get resolved without escalating.
• Catalog complexity. A simple, standardized catalog (few variants, clear rules) is easier for AI to resolve than one with case-by-case exceptions, negotiations or highly customized products.
• Escalation rules. How many keywords you use and how strict your human-escalation rules are directly influences the percentage — overly broad rules escalate too much "just in case"; overly narrow ones let through cases that did need a person.

No AI agent resolves absolutely everything, and anyone who promises it is selling something different from what they can deliver. There will always be a percentage of cases that need human judgment: sensitive complaints, negotiations outside policy, ambiguous situations that not even the best language model should resolve on its own. An honest vendor talks in ranges and averages, not absolute guarantees, and shows you how they define "resolved" before asking you to trust the number.

Before comparing your percentage against any industry benchmark, first define what you'll count as "resolved without a human" in your own case: a conversation where the bot replied and the customer never wrote back? One where the bot explicitly closed the case? With that definition fixed, periodically review what share of conversations never touched a human agent, and cross that number with the most common escalation reasons — that's where the real opportunity to improve your knowledge base lives, not in chasing someone else's percentage.

• How Much Does AI Customer Service Cost?
• First Response Time: Why It Matters
• How We Measure Our Metrics
• AI Customer Support