The bar is on the floor, and the industry still trips on it
SuperOffice's benchmark of 1,000 companies: 62% never reply to customer service emails. Among those that reply, the average response time is 12 hours and 10 minutes, and only 20% answer fully on the first reply. Read that again — one in five first replies contains the actual answer.
So no, the AI-support question is not "can a bot replace your team". The bar is "does the first reply contain the answer". Most teams fail that with humans. That is the gap worth automating, and it takes less technology than the keynote demos imply.
What the assistant reads
The assistant reads the full thread — not the last message, the whole conversation: every channel turn, the customer's history, the internal notes, the knowledge your team wrote. Then it drafts. The draft lands in the agent's view with what it based the answer on visible. The agent edits or approves. Where a human owns the ticket, the assistant never sends on its own. Closing the loop — pressing send, setting expectations, promising a refund — is a human act, because it is a commitment made in your name.
That rule costs us some demo wow. It buys you the thing that actually matters: no draft leaves a human-owned ticket without a human's click. An assistant that drafts in seconds and cites its sources makes your 20%-fully-answer statistic a 60 or 80. An assistant that fires unreviewed replies on autopilot at your customers makes your logo a liability.
Hands, not just words
A draft that says "please contact the billing team" is a cost center. The assistant has tools: MCP tool integrations — GitHub and GitLab among the built-ins — plus webhook-based tools into your own systems. Concretely: the draft can include the real status from your order system, because the assistant called your webhook and read it. It can reference the actual issue it opened in your repo. It works inside your operation, not around it.
The same hands run internal automation: flows that watch queues, enrich tickets, route work. All of it under rules you set, all of it logged in the audit trail.
What your data does (and does not do)
- PII is masked before any model call. Card numbers, national IDs, tenant-marked sensitive fields — replaced before inference.
- Your data never trains models. Ours, or a provider's. Contractually and technically.
- You can pin inference to EU regions, or bring your own provider keys if your compliance sheet prefers it.
- Everything is in the audit trail: what was read, what was drafted, who approved, what was sent.
The meter question
The assistant consumes tokens, and tokens are metered — tiered, cheaper as volume grows. Reading a full support conversation costs a fraction of a cent; drafting a reply, about the same. A month of heavy assistant use for a small team is single-digit euros — compare that to the per-agent AI add-ons at €50/agent/month (Zendesk Copilot's published price, September 2026), which is more than we charge for the entire teammate (3.000 BDK ≈ €2.50).
We can price it that low because we are not paying for a sales team to explain the AI upsell. The menu is public. The meter is visible in-app. Your first drafts cost less than the coffee your team drinks while reviewing them.
Drafts in seconds. Sources visible. Direct answers signed as AI. That is the whole trick.the AI assistants page