Two Very Different Bets

The AI support industry has split into two camps, and they're selling you very different futures.

Camp one says AI should sit next to your agents. It drafts replies, summarizes long threads, pulls order history, suggests macros. The agent stays in the loop. GitHub Copilot for support, essentially. Zendesk's Agent Copilot, Intercom's Fin AI Copilot, and Salesforce's Einstein Copilot all live here.

Camp two says AI should replace agents for specific ticket types entirely. No human reviews the response. The customer gets an answer in seconds, the ticket closes automatically, and your team never sees it. Supp, Sierra, and Ada's autonomous mode operate this way for qualifying intents.

Both camps have paying customers. Both have real results. But they're solving fundamentally different problems, and picking the wrong one wastes months and budget.

What Copilots Actually Do Well

A copilot reduces handle time. Agent still reads the ticket, still writes the reply, still clicks send. But instead of spending 90 seconds pulling up the customer's account, reading the last three conversations, and drafting a response from scratch, the copilot does that grunt work in 2 seconds.

Zendesk reports their copilot cuts first-response time by 36%. Intercom claims a 31% reduction in handle time. These numbers are plausible. If an agent handles 60 tickets per shift and each one takes 4 minutes, shaving 90 seconds off each interaction means they can handle 85 tickets instead. That's a 40% throughput increase without hiring anyone.

Copilots are also safe. If the AI suggests something wrong, the agent catches it before the customer sees it. You get AI speed with human judgment as a backstop. For regulated industries (healthcare, finance, insurance), this matters a lot. A bad automated response to a billing dispute can create legal liability. A bad draft that an agent corrects creates nothing.

The downside is obvious: you still need the agents. Per-seat costs don't drop. Headcount stays flat. You're buying productivity, not cost reduction.

What Full Automation Actually Does Well

Full automation eliminates cost per ticket for qualifying conversations. If 40% of your inbound volume falls into categories like order status, password reset, return policy questions, and account updates, automating those means 40% of your tickets cost pennies instead of dollars.

The math is straightforward. A human agent costs roughly $18-25 per hour fully loaded. At 8 tickets per hour (including wrap time, breaks, coaching), that's $2.25-3.12 per ticket. Supp charges $0.30 per resolution. Even at the low end of human cost, automation saves $1.95 per ticket. At 1,000 tickets per month, that's $1,950/month in savings on just the automated portion.

Intercom's Fin charges $0.99 per resolution, and Zendesk charges roughly $1.50 per automated resolution. Still cheaper than a human, but 3-5x more expensive than Supp. The gap matters at scale. A company handling 10,000 automatable tickets per month saves $12,000/month with Supp's pricing versus $5,100/month with Fin.

The risk with automation is customer experience. If the AI handles something it shouldn't, the customer gets a bad answer with no human to catch it. This is the Forrester prediction playing out in real time: companies rushing to automate everything damage their brand when the bot fumbles complex or emotional tickets.

The Hybrid Path Most Teams Should Take

Start with copilots for complex ticket types. Billing disputes, technical troubleshooting, cancellation requests, anything where context and empathy matter. Let agents use the AI as a research assistant while they maintain control of the conversation.

Simultaneously, identify your highest-volume, lowest-complexity intents. "Where's my order?" doesn't need a human. Neither does "What's your return policy?" or "How do I reset my password?" These are lookup tasks with deterministic answers. Automate them fully.

The split usually lands around 30-50% automation, 50-70% copilot-assisted. Companies with mature knowledge bases and clean data can push automation higher. Companies with complex products or regulated industries keep more tickets in the copilot lane.

How to Decide Which Intents to Automate

Three criteria. First, does this intent have a deterministic answer? If the answer changes based on account state, order history, or nuance, it needs at least copilot-level involvement. If it's the same answer every time (return window, pricing page link, feature availability), automate it.

Second, what's the consequence of a wrong answer? Getting a return policy wrong is annoying. Getting a medical billing question wrong could be actionable. Weight your intents by risk.

Third, what's the volume? Automating an intent that fires 5 times per month saves almost nothing. Automating one that fires 500 times per month saves real money. Start with the highest-volume, lowest-risk intents and work your way up.

Supp's classifier sorts incoming messages into 315 intents across 13 categories at $0.20 per classification. That classification step is what makes the copilot-vs-automation decision possible. You can't route tickets to the right path if you don't know what they are.

The Pricing Model Tells You Everything

If a vendor charges per seat, they're betting you'll keep humans in the loop. Their incentive is to make agents more productive, not to replace them. Copilot tools from Zendesk and Salesforce fit this model.

If a vendor charges per resolution, they're betting on automation. Their incentive is to resolve tickets without humans, because that's where the customer sees the value. Supp at $0.30 per resolution, Intercom at $0.99, Zendesk at roughly $1.50.

Neither model is inherently better. But understand what you're buying. A per-seat copilot makes your $50,000/year agent produce 40% more output. A per-resolution automation tool eliminates the need for that agent on specific ticket types entirely.

For most teams under 20 agents, the right answer is classification-first automation for simple intents, with humans handling the rest. You don't need a copilot if you don't have agents struggling with volume. You need fewer tickets reaching agents in the first place.