Not All Support Interactions Are Equal

An AI-automated support interaction costs roughly $0.02-0.50 depending on the tool and complexity. A human agent interaction costs $15-22 in fully loaded labor (salary + benefits + tools + management overhead + office space).

That's a 30-1000x cost difference. But it doesn't mean you should automate everything.

Where the $0.02 Interaction Wins

"What are your hours?" doesn't need a human. Neither does "how do I reset my password" or "where's my order" or "do you offer free shipping?" These are factual questions with factual answers. The customer wants information, not empathy.

For these interactions, speed matters more than warmth. An instant automated response at $0.02 is better than a 20-minute wait for a human to type the same answer at $22. The customer is happier, you're cheaper, everyone wins.

Roughly 40-60% of support messages in most businesses fall into this category. That's where the savings are.

Where the $22 Interaction Earns Its Keep

A customer writes: "I've been a loyal customer for 3 years and this is the second time you've double-charged me. I'm considering switching to [competitor]. What are you going to do about it?"

This message needs a human. Not because the AI can't understand it (a classifier would correctly identify it as a billing dispute). But because the customer is testing whether you care. They want someone to acknowledge the history, take responsibility, and make it right.

An AI response here, even a good one, risks turning a frustrated loyal customer into a former customer. The $22 interaction doesn't just resolve the ticket. It saves a customer worth potentially thousands in lifetime value.

The Blended Cost Math

The smart approach is blending. Let AI handle the $0.02-worthy interactions (factual, simple, repetitive) and route the $22-worthy ones to humans (emotional, complex, high-stakes).

Example: a team getting 1,000 tickets/month.

All human: 1,000 x $18 (average) = $18,000/month.

50% automation: 500 x $0.30 (AI) + 500 x $18 (human) = $150 + $9,000 = $9,150/month.

That's $8,850/month in savings, or $106,200/year. And your human agents are spending their time on the conversations that actually matter, not typing "our hours are 9-5" for the hundredth time.

How to Decide What Gets Automated

Ask two questions about each support category:

Does this require judgment or just facts? "What's your return policy" is facts. "I want to return this but I'm past the window, can you make an exception" is judgment.

What happens if the AI gets it slightly wrong? If the AI gives slightly wrong business hours, the customer shows up at the wrong time and is mildly annoyed. If the AI gives the wrong refund amount, you have a financial error and an angry customer.

Low judgment + low risk = automate. High judgment or high risk = human.

The Gray Zone

Some interactions are borderline. A customer asks "can I cancel my subscription?" That's simple enough to automate (yes, here's how). But a customer asking to cancel might actually want to be talked out of it. A human agent can ask "is there something specific that's not working?" and save the account.

For these gray-zone interactions, one approach works well: let AI handle the first response (confirm the intent, collect any needed information) and route to a human for the decision. The AI does the $0.02 intake work. The human does the $22 retention work. Total cost: maybe $12 instead of $22, and the customer doesn't wait in a queue for the initial response.

The Real Savings Aren't Just Money

When AI handles 500 simple tickets, your human agents don't just cost less. They do better work. Instead of answering "what are your hours" between complex cases, they spend all their time on conversations that need their skills.

Agent satisfaction goes up (they're doing interesting work instead of rote work). Resolution quality goes up (they're not rushing through complex issues to clear a queue). Customer satisfaction on human-handled tickets goes up.

The $0.02 interaction makes the $22 interaction more effective. That's the part the cost calculator doesn't capture.