AI vs Human Support at 3,200 Tickets a Month

"AI customer support is 80% cheaper" is one of the easiest claims to make and one of the hardest to defend at scale. The honest math depends on three numbers most decks skip: how often the AI actually resolves without escalation, how much human time gets spent cleaning up the cases it fails, and the brand cost of getting it wrong. This article runs a 3,200-ticket-per-month mid-market scenario through the AI vs Human Support Cost calculator and pulls the decision rule out of the result.

1. 3,200 tickets at $32/hour humans

The scenario is a mid-market SaaS at 3,200 support tickets per month. Each ticket takes a US-based human agent 9 minutes on average. Loaded cost (wages plus benefits and overhead) is $32 per hour, which sits in the band the U.S. Bureau of Labor Statistics reports for customer service representatives once benefits are added^[1]. AI-resolved tickets cost 6,500 tokens on Claude Sonnet at $3 input / $15 output^[3]. Each ticket the AI fails generates 6 minutes of human escalation overhead — context-gathering, reading the AI thread, and the actual reply.

Run those inputs through the engine and the headline split is $15,360 human-only vs $8,379.20 AI-first, a $6,980.80 monthly savings before any quality consideration. The cost per ticket falls from $4.80 to $2.62. At 68% resolution, the AI handles 2,176 tickets cleanly and 1,024 escalate to humans.

# ai-vs-human-support-cost (computed live from /engines/ai-vs-human-support-cost.js)
Engine input
  tickets_per_month     = 3200
  avg_human_minutes_per_ticket= 9
  human_hourly_cost     = 32
  ai_resolution_rate    = 68
  tokens_per_ai_resolved= 6500
  ai_input_price_per_mtok= 3
  ai_output_price_per_mtok= 15
  escalation_overhead_minutes= 6

Engine output
  humanOnlyMonthlyCost  = 15360
  costPerHumanTicket    = 4.8
  aiFirstMonthlyCost    = 8379.2
  costPerAiTicket       = 2.62
  monthlySavings        = 6980.8
  savingsPercent        = 45.45
  breakEvenTickets      = 1
  aiResolvedCount       = 2176
  escalatedCount        = 1024

2. The $15,360 human baseline

Human-only math is straightforward. 3,200 tickets × 9 minutes / 60 = 480 hours of agent time. 480 × $32 = $15,360 per month. Per-ticket cost is $4.80. This is the comparison baseline. It excludes ticket-routing software, knowledge-base maintenance, training, and supervisor overhead, which together typically add another 15-25% to true human-only support cost. Including them, the realistic human-only baseline lands closer to $17,500-$19,000 per month.

The Klaus/Zendesk benchmark reports median first-response time around 4-6 hours for SaaS and resolution time around 24-48 hours^[2]. The 9-minute figure here is the active touch time per ticket, not calendar time. A ticket that takes 36 hours to resolve still has only a few minutes of active human work on it; the rest is waiting for the customer's response. The engine prices active touch, which is what actually consumes payroll.

3. The AI-first $8,379 stack

AI-first means the AI agent attempts every ticket. If it resolves, the ticket closes. If it fails, a human picks up with full context. The engine breaks the cost into two streams:

AI cost on all 3,200 tickets:
  Tokens per ticket           6,500 (3,250 input + 3,250 output assumed)
  Cost per ticket             ($3 × 0.00325) + ($15 × 0.00325) = $0.0586
  Total AI compute            3,200 × $0.0586                   = $187.20

AI-resolved tickets (68% × 3,200 = 2,176):
  Cost                        $0 additional human time

Escalated tickets (32% × 3,200 = 1,024):
  Human time per ticket       6 escalation min + 9 resolution min = 15 min
  Total escalation hours      1,024 × 15 / 60 = 256 hours
  Escalation cost             256 × $32 = $8,192

Total AI-first monthly cost:  $187.20 + $8,192 = $8,379.20
Cost per ticket:              $8,379 / 3,200 = $2.62

Note that the engine charges humans for the full 15 minutes on escalated tickets, not 9. That is the right call: the human still does the original 9-minute resolution work, plus 6 minutes of context-gathering from the AI thread. Founders who model escalation as just "the 9 minutes" understate the cost by 67% on the escalation share, which is where the AI-vs-human math typically goes wrong.

4. The 68% resolution number

Resolution rate is the single most important variable in the model. Anthropic, Intercom, and other vendors quote 70-80% resolution rates in marketing materials; the real numbers in production usually land between 50% and 70% depending on ticket complexity and the quality of the knowledge base. The 68% used here is achievable for a well-tuned setup on a product with strong documentation.

"Resolution" needs a strict definition. The customer's problem is solved and the customer does not come back within 14 days with the same issue. Loose definitions (anything the AI replied to is "resolved") inflate the rate by 10-15 percentage points and break the cost model. The honest measurement is reopens-within-14-days, which most ticketing systems track natively.

Run sensitivity at three resolution rates: 55%, 68%, and 80%. The engine returns:

• 55%: 1,440 escalations × 15 min × $32/hour = $11,520 + $187 AI = $11,707 total. Saving vs human baseline: $3,653 (24%).
• 68%: 1,024 escalations × 15 min × $32/hour = $8,192 + $187 = $8,379. Saving: $6,981 (45%).
• 80%: 640 escalations × 15 min × $32/hour = $5,120 + $187 = $5,307. Saving: $10,053 (65%).

The takeaway: every 10 points of resolution rate is roughly $3,300 of monthly savings at this volume. That makes resolution improvement worth more than any cost optimization on the AI side. Most of the work to lift resolution is knowledge-base quality, prompt design, and intent classification — none of it expensive, all of it requiring effort. The AI Product Margin calculator handles the broader product-level economics once support cost is settled.

5. Escalation overhead and the 60% floor

The 6 minutes of escalation overhead is the variable founders most often miss. When an AI fails, the human cannot start from scratch — they need to read what the AI said, check what the customer replied, and decide whether to fix the AI's answer or restart. That work is real, and it does not exist in a human-only world. The engine prices it explicitly.

The break-even resolution rate is the rate where AI-first cost equals human-only cost. With these inputs:

Human-only cost:           $15,360
AI-first cost at R%:       $187 + (1-R) × 3,200 × ($32 × 15/60)
                         = $187 + (1-R) × 3,200 × $8
                         = $187 + 25,600 × (1-R)
Set equal:                 $15,360 = $187 + 25,600 × (1-R)
Solve for R:               (1-R) = 15,173 / 25,600 = 0.5927
                           R = 40.7% break-even at minimum AI

The clean break-even is 40.7% — well below the 60% floor cited as conservative practice. The reason 60% is the operating floor, not 40%, is that the model excludes brand cost. A wrong AI answer that frustrates a customer costs more than the 6 minutes of escalation overhead; it costs churn risk, negative reviews, and word-of-mouth damage. The Klaus/Zendesk CSAT benchmark places top-quartile CSAT at 95%+ and median around 88%^[2]; AI-first stacks at low resolution often drop to 75-82% CSAT, which has measurable retention impact.

6. CSAT and the bad-LLM-day risk

Cost is necessary but not sufficient. Three quality risks to price in:

• Hallucinated solutions. The AI confidently asserts incorrect fixes ("clear your cache" when the bug is server-side). Customer follows the wrong fix, problem persists, frustration compounds. Mitigate with retrieval-augmented generation (RAG) on the real knowledge base, plus a refusal-to-answer pattern when confidence is low.
• Tone mismatch. AI responses default to formal, corporate, and slightly defensive. Customers reading them feel they are being handled by a bot, which kills CSAT independent of correctness. Mitigate with a tone-tuned system prompt and human-in-the-loop review for the first 1,000 tickets to calibrate.
• Bad LLM days. Model providers occasionally ship regressions. A model that was at 68% resolution drops to 52% overnight after a quiet update. Mitigate with continuous evaluation: a held-out test set of 100 representative tickets, run weekly, with an alert when resolution drops more than 5 points.

These risks do not invalidate the cost model; they justify staying north of 60% resolution. Below that, the human-only baseline is competitive on cost and superior on quality.

7. The break-even framework for solo SaaS

Three rules survive the math for solo founders running support themselves or with one part-time contractor:

• Track resolution rate weekly, not monthly. A weekly cadence catches model regressions and prompt drift before they consume a month of margin. Below 60%, pull the AI in front of escalations only (suggested replies for the human to send), not fully autonomous.
• Run the calculator at three resolution rates. 55, 65, 75. If the 55-case still saves real money against your blended cost, the AI stack holds up across bad weeks. If the 75-case is the only one that wins, the stack is fragile and one bad model release wipes the savings.
• Count escalation overhead at 1.5x to 2x the AI-resolved touch time. 6 minutes in this scenario is conservative for a product with reasonable documentation. For complex technical products it is closer to 10-12 minutes. Tune to your actual data.

The decision is not "AI or human." It is "what mix gets to 60%+ resolution at acceptable quality." Below that bar, stay human and use the AI as a draft-assistant; above it, switch to AI-first and run humans only on escalations. The agent cost engine handles the parallel question on acquisition. See the methodology page for the full derivation^[4].