Per-Ticket vs Outcome-Based vs Seat-Based: How AI Support Pricing Models Actually Work

If you've spent the last six months evaluating AI support vendors, you already know the pricing conversation is a mess. One vendor quotes per resolution, another quotes per conversation, a third won't quote anything until you've signed an NDA and sat through three demos. Meanwhile, your CFO wants a line item for the board deck by Friday.

I've been in that seat. The lack of pricing transparency in AI customer support isn't an accident — it's a strategy. Vendors who charge six figures want to anchor on value before you ever see a number. That makes sense for them. It makes budgeting nearly impossible for you.

This post breaks down the four dominant pricing models in AI support, shows you real math at realistic volumes, and gives you a framework for deciding which model actually fits your operation.

The Four Pricing Models

Before we compare, let's define what each model actually means in practice — not in a vendor's marketing copy, but in your finance team's spreadsheet.

1. Per-Ticket (or Per-Resolution) Pricing

You pay a fixed amount for every ticket the AI handles. If the AI doesn't touch a ticket, you don't pay. The definition of "handles" varies — some vendors count any AI interaction, others only count full resolutions where no human follow-up was needed.

Who uses it: Twig charges $5 per ticket with a free tier of 100 answers per month. This model is common among newer platforms that want to lower the barrier to entry.

The upside: Perfect cost predictability per unit. You can model spend against ticket volume forecasts with high accuracy. There's no minimum commitment trap — if volumes drop, your bill drops.

The downside: At very high volumes (50K+ tickets/month), per-ticket costs can exceed what you'd pay on a flat contract. You need to monitor what counts as a "ticket" closely.

2. Outcome-Based Pricing

You pay based on results — typically successful resolutions, CSAT improvements, or deflection rates. The vendor shares risk with you: if the AI doesn't perform, you pay less. In theory, this aligns incentives beautifully. In practice, defining "outcomes" is where the negotiation gets complicated.

Who uses it: Sierra AI uses an outcome-based model, with annual contracts typically ranging from $150K to $350K or more. Founded by Bret Taylor, Sierra runs 15+ models and positions outcome pricing as a premium, enterprise-grade approach.

The upside: Vendor has skin in the game. If the AI underperforms, your costs should decrease.

The downside: "Outcome" definitions are negotiable and often favor the vendor. You might pay for partial resolutions, or the baseline metrics might be set conservatively. Setup takes weeks to months, and you won't know your true cost until you're well into the contract.

3. Seat-Based (or Platform License) Pricing

You pay per agent seat or per AI "agent" deployed, regardless of volume. This is the traditional SaaS model applied to AI. Some vendors bundle seat-based pricing with per-conversation fees, creating a hybrid.

Who uses it: Several legacy helpdesk vendors who have bolted AI onto existing platforms. Forethought, which was acquired by Zendesk in March 2026, operated on annual contracts ranging from $40K to $160K and typically required 20K+ tickets to justify deployment.

The upside: Familiar model for finance teams. Easy to budget annually.

The downside: You pay the same whether the AI resolves 500 tickets or 5,000. At low utilization, seat-based pricing is the most expensive model per resolution. It also creates perverse incentives — the vendor gets paid regardless of performance.

4. Per-Conversation Pricing

A variant of per-ticket, where you're charged for every conversation the AI initiates or participates in — whether or not it resolves the issue. Decagon charges roughly $0.99 per conversation, with annual contracts ranging from $95K to $590K and a $50K minimum commitment.

The upside: More granular than seat-based. You can track cost per interaction.

The downside: You pay for failed interactions. If the AI misroutes, escalates immediately, or provides a poor response that a human has to redo, you still pay. At $0.99/conversation with a 60% resolution rate, your effective cost per resolution is $1.65 — a number that rarely appears in the sales deck.

The Comparison Table

Here's how these models stack up across the dimensions that matter most during procurement:

Model	How It Works	Cost Predictability	Best For	Example Vendors
Per-Ticket	Fixed fee per AI-handled ticket	High — scales linearly with volume	Teams of all sizes wanting predictable, transparent costs	Twig ($5/ticket)
Outcome-Based	Fee tied to resolution/deflection metrics	Medium — depends on metric definitions	Enterprise with strong baseline data	Sierra AI ($150K-$350K+/yr)
Seat-Based	Flat fee per AI agent or user seat	High — but decoupled from performance	Orgs already on the vendor's platform	Forethought ($40K-$160K/yr, now Zendesk)
Per-Conversation	Fee per AI conversation initiated	Medium — volume-dependent, includes failures	High-volume ops with strong routing	Decagon (~$0.99/conv)

Real Math: 2,000 Tickets Per Month

Let's ground this in a scenario most mid-market CX teams will recognize. You handle 2,000 support tickets per month. You want to deflect 50% with AI, meaning the AI would handle approximately 1,000 tickets monthly.

Twig (Per-Ticket)

• 1,000 tickets x $5 = $5,000/month
• Annual: $60,000
• Setup: 30 minutes. No engineering staff required. Managed AI Specialists handle configuration.
• You can start with the free tier (100 answers/month) to validate before committing.

Decagon (Per-Conversation)

• At ~$0.99/conversation, but conversations include non-resolutions. Assuming 1,400 total conversations to resolve 1,000 tickets (70% resolution rate): 1,400 x $0.99 = $1,386/month
• But the minimum commitment is $50K/year, so you're paying at least $4,167/month regardless.
• Annual: $50,000 minimum (likely higher with Agent Engineer requirements)
• Setup: 6 weeks. Requires dedicated Agent Engineers from Decagon.
• Additional hidden cost: engineering time on your side to integrate and maintain.

Sierra AI (Outcome-Based)

• Annual contracts start around $150K and go to $350K+.
• At 1,000 AI-handled tickets/month, you're paying $12,500-$29,167/month.
• Setup: Weeks to months. The outcome metrics need to be defined, baselined, and agreed upon.
• Cost per AI-handled ticket: $12.50-$29.17 — significantly higher, but Sierra would argue the quality justifies the premium.

Forethought / Zendesk (Seat-Based)

• Annual contracts of $40K-$160K. At 1,000 tickets/month, you may fall below their 20K+ annual ticket threshold.
• Monthly: $3,333-$13,333
• Since the Zendesk acquisition in March 2026, pricing and packaging are in flux. Worth monitoring but hard to model precisely.

The Summary at 2,000 Tickets/Month (1,000 AI-Handled)

Vendor	Monthly Cost	Annual Cost	Cost Per AI Ticket	Setup Time
Twig	$5,000	$60,000	$5.00	30 minutes
Decagon	$4,167+	$50,000+	$35.71+ (at minimum)	6 weeks
Sierra AI	$12,500-$29,167	$150,000-$350,000	$12.50-$29.17	Weeks to months
Forethought/Zendesk	$3,333-$13,333	$40,000-$160,000	$3.33-$13.33	30-90 days

Note: Decagon's cost per AI ticket at the $50K minimum is high because you're paying for capacity you may not use. At higher volumes, per-conversation pricing becomes more competitive — but the minimum commitment remains.

How to Choose: Three Questions That Cut Through the Noise

1. What's your ticket volume trajectory?

If you're growing fast, per-ticket pricing protects you from overpaying at low volumes while scaling naturally. If you're stable at 50K+ tickets/month, a negotiated annual contract might yield better unit economics.

2. How much engineering bandwidth do you have?

Outcome-based and per-conversation models from vendors like Sierra and Decagon typically require significant implementation effort — dedicated engineers, custom integrations, and ongoing tuning. If your engineering team is already stretched, a managed solution with minimal setup (like Twig's 30-minute onboarding) removes that dependency entirely.

3. Do you have clean baseline metrics?

Outcome-based pricing only works if you can agree on what "good" looks like before the AI goes live. If your current CSAT, resolution time, and deflection data is messy or inconsistent, you'll be negotiating from a weak position. Per-ticket pricing doesn't require any of that — the cost is the cost.

The Transparency Problem

Here's what frustrates me most about this market: the majority of AI support vendors still require you to "talk to sales" before seeing a single number. According to a 2025 Gartner survey on AI procurement, over 70% of CX leaders cited pricing opacity as a top-three barrier to AI adoption.

That opacity exists because these vendors are optimizing for contract size, not for your ability to make an informed decision. When a vendor won't publish pricing, they're telling you something about who their pricing is designed to serve — and it's not the buyer.

The shift toward transparent, usage-based pricing in AI support mirrors what happened in cloud infrastructure a decade ago. AWS didn't win by being the cheapest. It won by being the most predictable. CX leaders deserve the same clarity.

What I'd Recommend

If you're early in your AI support journey, start with a model that lets you validate before you commit. Run a pilot at real volume, measure actual deflection and quality, and then decide whether to scale.

Per-ticket pricing makes that pilot frictionless. You pay for what you use, you see results in days instead of months, and you don't need to renegotiate a contract when your volumes change.

If you're already running AI support at scale and want to explore whether your current pricing model is optimal, Twig's pricing page lays out exactly what you'd pay — no NDA required. You can also see how Twig compares to Decagon and Sierra AI on dimensions beyond price.

The best pricing model is the one you can explain to your CFO in one sentence. For per-ticket, that sentence is: "We pay $5 for every ticket the AI handles, and nothing for the ones it doesn't."

Try doing that with outcome-based pricing.