The Death of Per-Seat Pricing

For twenty years, support software charged by the seat. You paid for each agent who logged into the system, whether they handled 5 tickets or 500. Zendesk, Freshdesk, Intercom, Help Scout: all seat-based.

That model made sense when humans did all the work. It doesn't make sense when AI does most of it.

If an AI bot resolves 70% of your tickets automatically, why are you paying per-seat for agents who only touch 30%? The economics broke. Vendors noticed.

The Shift to Per-Resolution

Starting in 2024, the major players began rolling out per-resolution pricing for their AI features. The pitch: you only pay when the AI successfully handles a customer interaction. No resolution, no charge.

Here's what that actually looks like across the market:

The 10x spread between the cheapest and most expensive option tells you something important: these products are not doing the same thing.

What You're Actually Paying For

The $1.50-2.00 tools (Sierra, Zendesk, Salesforce) are running large language models on every conversation. The AI reads your knowledge base, generates a natural-language response, handles follow-up questions, and tries to resolve the issue conversationally. That's expensive inference. GPT-4 class models cost real money per token, and a multi-turn support conversation burns through tokens fast.

The $0.99 tier (Intercom Fin) does the same thing but has been at this longer and has better unit economics from scale.

The $0.20-0.30 tier (Supp) takes a fundamentally different approach. Instead of generating freeform conversation, it classifies the customer's intent using a purpose-built model (not a general LLM) and triggers a predetermined action. The model is small, fast (100-200ms), deterministic, and cheap to run. No token costs. No hallucination risk.

The trade-off is clear: LLM-based tools give you conversational flexibility. Classification-based tools give you speed, predictability, auditability, and dramatically lower cost. For the 70% of support tickets that follow known patterns, classification handles them perfectly. For the 30% that need creative problem-solving, you need a human anyway.

When Per-Resolution Makes Sense

Per-resolution pricing is great when your AI resolution rate is high and predictable. If the AI resolves 70% of incoming tickets and you know your monthly volume, you can forecast costs precisely.

At 5,000 tickets/month with a 70% AI resolution rate (3,500 AI resolutions):

That's a $2,700-$4,750 monthly difference. Over a year, you're looking at $32,000-$57,000 in savings.

When It Doesn't Make Sense

Per-resolution pricing hurts when your volume is unpredictable. A product launch might spike tickets 5x for a week. A service outage floods your queue. Suddenly your AI bill quadruples because the same "is this down?" question gets resolved 2,000 times.

Some vendors offer committed volume discounts (Zendesk's $1.50 committed vs $2.00 PAYG is a good example). If you can predict your volume within 20%, committed pricing saves money. If you can't, you're either overpaying on committed volume you don't use or getting hit with PAYG rates on spikes.

Supp avoids this problem partly through price. Even a 5x spike at $0.20-0.30/resolution is manageable. A 5x spike at $1.50-2.00 can blow a monthly budget in days.

The Hidden Catch: What Counts as a "Resolution"

This is where vendors get creative.

Some count a resolution when the AI responds and the customer doesn't reply within a time window (often 24-72 hours). Customer gave up and called your phone line instead? That's a "resolution." Customer found the answer themselves after the bot gave a useless response? Resolution.

Others count it when the ticket is formally closed. Slightly better, but auto-close rules can inflate numbers.

Ask your vendor these questions before signing:

• Does a customer who abandons the conversation count as resolved?
• If the customer contacts you again about the same issue within 48 hours, is that a new resolution charge?
• Are "I don't know, let me transfer you" responses charged?
• What's the average resolution rate across your customer base (not the cherry-picked case study)?

The Real Question

Per-resolution pricing aligns vendor incentives better than per-seat. The vendor only gets paid when their AI actually works. That's progress.

But the price range matters enormously. At $0.20-0.30 per resolution, AI support is almost free at scale. At $1.50-2.00, it's cheaper than humans but still a serious line item. At $0.99, it's the middle ground that most mid-market companies have settled on.

Your choice depends on what you need the AI to do. If you need freeform conversation with nuanced responses, you're paying LLM prices. If you need fast, accurate classification and automated actions on known intents, you can pay a fraction of that. Most support queues are 60-70% known intents. Start there.