Every AI help desk promises the same thing: deflect more tickets, resolve faster, spend less. But read the pricing pages closely and a pattern emerges — the AI that is supposed to save you money is often billed in a way that quietly takes those savings back. Before you sign, it is worth understanding exactly how AI gets billed, because the model matters far more than the headline seat price.
How AI help desks actually bill for AI
There are four common models, and most vendors use some combination of them:
- Per-resolution: you pay a fee each time the AI "resolves" a conversation (Intercom Fin is the best-known example). What counts as a resolution is defined by the vendor.
- Per-ticket: you pay by ticket volume regardless of how it is handled (Gorgias bills primarily this way), so AI savings do not lower the line item.
- AI add-on tier: AI is a paid upgrade on top of the seat price (Zendesk sells Advanced AI as an add-on; Freshdesk gates Freddy AI to higher plans).
- Included per-seat: AI is bundled into the per-user price with no usage meter — the model Relay uses.
Why per-resolution pricing punishes the thing you want
Per-resolution pricing sounds fair — you only pay when the AI does work. The problem is the incentive it creates for your budget. The whole point of buying AI is to deflect more tickets. But under per-resolution billing, every deflected ticket is a charge. The better your knowledge base gets, the more the AI resolves, and the bigger your bill grows. You are effectively penalized for success, and your costs become hardest to predict in exactly the months that matter most — a product launch, an outage, or a Q4 sales spike.
The question to ask every vendor
What, precisely, counts as a billable "resolution"? A closed conversation? A deflected article view? An AI reply the customer did not follow up on? The definition determines your bill, and it varies widely between vendors.
Worked example: 2,000 tickets a month
Say your team handles 2,000 support conversations a month with 5 agents, and AI deflects or assists on 40% of them (800 conversations). Here is how the AI line item compares across models. Seat costs are illustrative entry prices; AI rates are rounded for clarity — the point is the shape of the curve, not exact vendor quotes.
Illustrative monthly AI cost at 2,000 tickets, 5 agents, 800 AI-handled
| Model | Seat cost (5 agents) | AI cost | Notes |
|---|---|---|---|
| Per-resolution (~$0.75–$0.99 ea.) | Base plan + AI | $600–$790 | Scales up as deflection improves |
| Per-ticket | N/A (volume-based) | Rises with all 2,000 | AI savings do not reduce the bill |
| AI add-on tier | Higher tier required | Tier upgrade | Often forces a plan jump |
| Relay — included per-seat | $95 (5 × $19) | $0 extra | AI bundled; no usage meter |
The crossover is stark: under usage-based models, your AI bill at 800 deflected conversations can rival or exceed your entire seat cost. Under a flat per-seat model, deflecting 800 — or 8,000 — conversations costs the same. Cost scales with your team, not with how hard the AI works.
The flat-rate, AI-included alternative
Relay bundles AI into the seat price. AI reply drafts and a free, built-in AI knowledge base are included on every plan at $19/user/month (and the free plan covers up to 2 users), with no per-resolution or per-ticket meter. One honest distinction to set expectations: Relay’s AI drafts replies for an agent to review and approve before sending — agent-assist, not unattended auto-resolution. For most teams, keeping a human in the loop is the point: it protects brand voice and avoids shipping a hallucinated answer to a customer, while still cutting response time dramatically.
If you are weighing specific tools, the head-to-head pages break down pricing line by line: Relay vs Intercom, Relay vs Zendesk, and Relay vs Gorgias.
Questions to ask any AI help desk vendor
- Is AI metered, and if so, by what unit (resolution, ticket, message)?
- What exactly counts as a billable "resolution"?
- Is the knowledge base included, or a separate product/add-on?
- Does my bill go up when deflection improves or volume spikes?
- Can I forecast next quarter’s cost from headcount alone?