Two tickets arrive at the same time.

Ticket A: "What's my tracking number?" An AI classifier reads the message, looks up the order by email, and responds with the tracking number and delivery estimate. Total elapsed time: 4 seconds. Total cost: about $0.02 in compute, $0.30 if you're paying Supp's per-resolution rate.

Ticket B: "I was billed for a year when I signed up for monthly, and now my bank says I'm overdrawn. I need this fixed today." A human agent reads the message, checks the billing system, sees a plan migration error, consults with the billing team, processes a partial refund and a plan correction, writes a detailed response explaining what happened, and follows up the next day to confirm the refund posted. Total elapsed time: 35 minutes spread across 2 days. Total cost: about $22 in loaded labor.

Both are "support tickets" in your metrics. Your cost-per-ticket number averages them together and produces something like $11. That average is useless because it describes neither ticket accurately.

Why Averages Hide Everything

If 60% of your tickets are Ticket A (simple, automatable) and 40% are Ticket B (complex, human-required), your blended cost per ticket is:

(0.6 × $0.30) + (0.4 × $22) = $0.18 + $8.80 = $8.98 per ticket.

That $8.98 average tells you nothing actionable. You can't reduce the cost of Ticket A much below $0.30 (it's already near-zero). And you can't automate Ticket B (it requires judgment, system access, cross-team coordination, and empathy).

The actionable insight isn't the average. It's the ratio. What percentage of your tickets are in each bucket? And can you shift tickets from the expensive bucket to the cheap bucket?

The Cost Spectrum

Support tickets don't exist in two buckets. They exist on a spectrum:

Tier 0 ($0.01 to $0.05): Self-service. Customer finds the answer without contacting you. The cost is your hosting and search infrastructure.

Tier 1 ($0.20 to $0.50): AI-resolved. Customer contacts you, AI classifies and responds. No human involvement. Password resets, order status, business hours, pricing.

Tier 2 ($3 to $8): AI-assisted, human-resolved. AI classifies and gathers context. Human reads the pre-classified ticket with context attached and resolves in 5 to 10 minutes. Feature questions, moderate billing issues, basic troubleshooting.

Tier 3 ($15 to $30): Human-resolved, complex. Multi-touch, multi-day. Escalations, complaints, technical debugging, account recovery. Requires judgment, empathy, and system access.

Tier 4 ($50 to $100+): Crisis. Security incidents, legal threats, VIP accounts, issues requiring cross-departmental coordination. Involves senior staff and sometimes external parties.

Shifting the Ratio

The goal of support optimization isn't to reduce the average cost per ticket. It's to move as many tickets as possible from higher tiers to lower tiers.

Every Tier 3 ticket you can turn into a Tier 2 ticket (by giving AI the context-gathering job) saves $10. Every Tier 2 ticket you can turn into a Tier 1 ticket (by automating the response entirely) saves $5. Every Tier 1 ticket you can turn into Tier 0 (by improving self-service) saves $0.30.

The biggest ROI is in the Tier 2 to Tier 1 shift. These are tickets that have standard answers but currently need a human because the human verifies the answer, accesses a system, or provides a personal touch.

AI classification makes this shift possible. When Supp classifies a message as "order status inquiry," it can automatically look up the order (via webhook or integration) and respond with the tracking info. What was a 5-minute Tier 2 ticket becomes a 3-second Tier 1 ticket.

Real Example

A company with 1,000 tickets/month at this distribution:

Tier 0: 0% (no self-service currently) Tier 1: 0% (no AI currently) Tier 2: 60% (600 tickets at $5 each = $3,000/month) Tier 3: 35% (350 tickets at $22 each = $7,700/month) Tier 4: 5% (50 tickets at $60 each = $3,000/month)

Total: $13,700/month.

After implementing AI classification and self-service:

Tier 0: 15% (150 tickets at $0.02 each = $3/month) Tier 1: 35% (350 tickets at $0.30 each = $105/month) Tier 2: 25% (250 tickets at $5 each = $1,250/month) Tier 3: 22% (220 tickets at $22 each = $4,840/month) Tier 4: 3% (30 tickets at $60 each = $1,800/month)

Total: $7,998/month. Savings: $5,702/month ($68,424/year).

The ticket count didn't change. The distribution did. That's the lever.

Stop Reporting Averages

If you report "cost per ticket" as a single number, you're hiding the most important information.

Report cost per ticket by tier. Show the distribution. Track how the ratio shifts over time. "We moved 15% of our volume from Tier 2 to Tier 1 this quarter, saving $X/month" is a meaningful metric. "Our average cost per ticket dropped from $11 to $8" is a number that obscures more than it reveals.

Supp's analytics show intent distribution, which maps directly to tier distribution. If 35% of classified intents are in categories you've configured for auto-resolution (Tier 1), and 15% are deflected to self-service (Tier 0), you can calculate your real cost at each tier and track the shift.

The 2-cent interaction and the $22 interaction are both support. But they need different measurement, different optimization strategies, and different staffing. Stop averaging them. Start separating them.