The pitch for AI in support usually goes like this: "AI handles the easy stuff, humans handle the hard stuff, everyone wins." And that's roughly correct. But "easy" and "hard" aren't as obvious as they seem.

"What are your business hours?" is easy. "I want a refund" looks easy but might be complex depending on the circumstances. "I'm trying to integrate your API with a legacy Oracle system running on-prem in a regulated environment" is obviously hard. But what about "I'm frustrated with your product"? Is that easy (template response) or hard (churn risk that needs a personal touch)?

The hybrid model works. Most teams that implement it automate 40 to 70% of their volume. But drawing the line well is what separates a good implementation from a bad one.

What AI Should Handle

Start with the obvious: questions that have one correct answer, every time.

Password resets. Order status lookups. Business hours. Return policy. Shipping times. Pricing. These are information retrieval queries. The customer asks, the answer exists, AI delivers it. No judgment needed.

Next level: transactional requests with clear rules. "Skip my next subscription box" (if it's before the cutoff, skip it). "Change my email address" (update the field). "Downgrade my plan" (process the downgrade). These require action, but the action follows a rule. If [condition], then [action].

AI handles these faster and cheaper than humans. A password reset that takes a human agent 3 minutes takes AI 5 seconds. At $10 per human-handled ticket vs $0.30 per AI-handled ticket, the economics are obvious. And the customer prefers the faster experience.

The threshold for AI automation: the response is always the same (or follows a clear decision tree), no emotional judgment is needed, the risk of a wrong answer is low, and the customer doesn't need to feel heard by a person.

What Humans Should Handle

Complex complaints with emotional context. "Your product ruined my presentation in front of 200 people." AI can classify this as a complaint. It cannot grasp the humiliation, the career implications, or the appropriate level of compensation. A human can.

Edge cases that don't fit any category. The customer's situation involves two policies that conflict, or a scenario nobody anticipated, or a request that requires creative problem-solving. AI classifies it as "other" and routes it. Humans figure it out.

High-value accounts. If a customer represents $50,000/year in revenue, their billing question might technically be automatable, but the relationship value justifies a human touch. Some interactions are about maintaining the relationship, not resolving the ticket.

Legal and compliance matters. Anything involving data deletion requests, regulatory inquiries, legal threats, or liability should go to a human, period. AI should classify and route these immediately, but a human with authority makes the decisions.

Emotionally charged situations. Grief, anger, fear, desperation. A customer whose business is down because of your outage isn't just reporting a bug. They're scared. AI can classify the urgency. A human provides the reassurance.

Drawing the Line: A Decision Framework

For each ticket category, ask four questions:

1. Is the answer always the same? If yes, automate. If it varies case-by-case, keep it human.

1. Is emotional intelligence needed? If the customer is likely upset, scared, or confused, route to a human. If they just need information, AI is fine.

1. What's the cost of a wrong answer? If a wrong answer is mildly inconvenient (wrong article link), AI is fine. If a wrong answer has financial or legal consequences (wrong refund amount, wrong compliance advice), use a human.

1. Is this a retention opportunity? If the customer is considering leaving, a human can save them. AI can't (well, it can offer a scripted retention offer, but it can't have a genuine conversation about what would make them stay).

If the answer to all four leans toward automation, automate. If any one of them leans toward human, use a human. When in doubt, err toward human. A human handling a simple ticket is slightly wasteful. AI mishandling a complex ticket is damaging.

The Handoff Is Everything

The worst hybrid implementations nail the AI part and botch the handoff. Customer explains their issue to AI. AI realizes it can't handle it. Customer gets connected to a human. Human says: "How can I help you today?"

The customer just explained the problem. They're now re-explaining it. This is the single most frustrating experience in modern customer support.

Good handoffs include: the full conversation history (so the human can read what was already said), the AI's classification (so the human knows the category and priority), any data the AI already gathered (order number, account details, error codes), and a suggested next step based on similar past tickets.

The human should start with: "I can see you're having an issue with [X] and you've already tried [Y]. Let me look into this." The customer feels heard. The human has context. The resolution is faster.

Supp's classification provides this automatically. When a ticket escalates from AI to human, the agent sees the intent classification, the priority score, and the full message history. No cold transfers. No re-explaining.

Realistic Automation Percentages

Don't believe vendors who promise 80% or 90% automation. Here's what's realistic:

Teams with simple products (one product, clear policies, limited use cases): 60 to 70% automation. Most queries are FAQ-level.

Teams with moderate complexity (multiple products, varied policies, some edge cases): 40 to 55% automation. The long tail of edge cases keeps humans busy.

Teams with high complexity (enterprise software, regulated industries, technical products): 25 to 40% automation. Lots of questions require context, judgment, or specialized knowledge.

Teams with high emotional stakes (healthcare, financial services, crisis support): 15 to 30% automation. Most interactions need a human presence regardless of technical complexity.

These percentages represent full automation (no human involvement). Partial automation (AI classifies and pre-fills, human reviews and sends) can touch another 20 to 30% of volume. So even in complex environments, AI assists on 50 to 70% of tickets even if it only fully resolves 30%.

The Team Impact

When you automate 40 to 60% of ticket volume, your human agents don't just do "less work." They do different work.

The easy, repetitive tickets disappear. What's left is harder, more complex, more emotionally demanding. Agent work becomes more interesting but also more draining.

Plan for this. Agents handling only complex tickets need more breaks, more support, and more recognition. The job got harder, even though the volume went down.

Also: don't immediately cut headcount. Use the freed capacity for proactive support (reaching out to at-risk accounts), quality improvements (better responses, faster resolutions), coverage expansion (longer hours, faster response times), and documentation (building the knowledge base that reduces future ticket volume). These investments pay back more than the headcount savings.