How much does AI customer service cost? Pricing models and how to compare proposals
By Bernardo Gontijo, founder of NeftPublished on Updated on 7 min read
Summary
The cost of AI customer service depends on the pricing model: per user, per use (minute, message or credit), subscription with an allowance or per outcome. An October 2026 study of 75 AI offerings found that subscriptions with allowances were most common and outcome-based pricing appeared only in customer service. To compare proposals, calculate the cost per genuinely resolved interaction, not the unit price.
How do providers charge for AI customer service?
There are five main models, and many proposals combine two or three. The classification below was adapted from From Tokens to Credits to Outcomes, published on arXiv on October 22, 2026 by Pratik Gupta, who analyzed the pricing pages of 75 AI offerings captured on September 16, 2026.
| Model | How it charges | Advantage | Buyer risk |
|---|---|---|---|
| Per user (seat) | Fixed monthly fee per agent | Predictable | Same charge whether AI handles little or a lot |
| User with usage allowance | Seat with included usage and separately billed overage | Predictable up to the allowance | Expensive overage during peak months |
| Subscription with allowance | Monthly plan with a usage limit, often in credits | Known limit | Credits expire or run out before month-end |
| Usage-based | Per minute, message or credit consumed | Pay only for what you use | Long conversations cost more, even without resolving anything |
| Outcome-based | Per interaction deemed resolved | Price linked to value delivered | Depends on who defines and counts resolution |
Which AI pricing model is most common in 2026?
Subscription with an allowance. In Gupta's study, it appeared in 16 of the 29 AI-native companies with public prices (55%). The figure has an important limitation: 23 of the 52 AI-native companies in the sample publish no prices, so the actual market proportion is unknown.
Outcome-based pricing was rare and concentrated. In the sample, all offerings priced by outcome (Intercom Fin, Zendesk AI, Salesforce Agentforce and Sierra) sell to customer service, and no offering outside customer service charged this way.
“Outcome-based pricing is real but limited, and exists precisely where the outcome can be contracted: observable and auditable enough to serve as a billing event.”
The author discloses employment at Stripe, a payments and billing platform, and excluded it from the sample. Coding was performed by one person, with independent review of 20% of rows.
What counts as a resolved interaction in outcome-based pricing?
Whatever the contract says, and the definition is often broader than “the customer's problem was solved.” On Intercom's pricing page, consulted on October 26, 2026, Fin costs US$0.99 per outcome with a minimum monthly commitment, and an outcome is counted when:
- the customer confirms the problem was solved;
- the customer requests no further help after the AI response;
- the AI completes a workflow (procedure), including handoffs.
The second criterion counts a customer who gave up and left as an outcome. The third can include conversations that ended with the team. Billing occurs once per conversation, even if multiple questions are answered.
This does not make the model bad, but it shows that “billed outcome” and “problem solved” are different measures. Before signing, ask for a written definition and a per-conversation report that lets you verify the count.
How much does AI actually resolve in customer service?
Less than sales materials often promise. In episode 15 of SaaStr's The Agents, published on October 2, 2026, the hosts discuss an evaluation published by Gorgias, a customer service AI provider, comparing its own AI with competitors in support tests.
According to the report, the median automation rate among evaluated providers was 48%, versus the 70% to 100% commonly advertised. The best result was around 79%, with lower response quality.
“48%. That is the real number. Everyone says 70, 80, 100, but those numbers come with asterisks.”
Treat the figure as an indication, not a benchmark. The data comes from an evaluation by a provider competing with those evaluated, and the host states that SaaStr was Gorgias's first investor.
Automating tickets is not the same as reducing work. In a SaaStr AI Day talk published in July 2026, Marty Kausas, cofounder and CEO of B2B support platform Pylon, described a company with around 5,000 employees whose AI deflected about 50% of tickets without changing team size. The tickets resolved autonomously were the easiest; the difficult ones stayed with people.
How do you calculate the real cost of an AI proposal?
Divide the total monthly AI cost by the number of genuinely resolved interactions. Consider an interaction resolved when the customer does not contact the company again for the same reason within a defined period, such as seven days.
| Hypothetical example (outcome-based proposal) | Value |
|---|---|
| Conversations handled by AI during the month | 10,000 |
| Price per outcome | BRL 5.00 |
| Outcomes counted by the provider | 6,000 |
| Monthly cost | BRL 30,000 |
| Interactions with no repeat contact within 7 days (own measurement) | 4,800 |
| Cost per genuinely resolved interaction | BRL 6.25 |
The same calculation applies to other models. With usage-based pricing, cost increases with conversation length, so test with real conversations, not short demo questions. For subscriptions with allowances, simulate the busiest month of the year.
What should you ask a provider before buying customer service AI?
- What is the billing unit (user, minute, message, credit or outcome), and exactly what triggers each charge?
- For outcome-based pricing: does a customer who stops responding count as resolved? What about a handoff to the team?
- Is there a minimum monthly commitment, an allowance and an overage price? Do credits expire?
- Can I set a spending limit or receive an alert before exceeding the budget?
- Is there a per-conversation report to audit the provider's count?
- Can prices change during the contract? With how much notice?
- For public agencies and utilities: does the model allow a predictable maximum contract value?
What does the research change when buying customer service AI?
| What the source shows | What changes in procurement | Limitation |
|---|---|---|
| Gupta: subscriptions with allowances are most common among companies publishing prices | Simulate the peak month and overage price | 23 of 52 companies publish no prices |
| Gupta: outcome-based pricing appeared only in customer service | Require a written definition of outcome | Sample of 75 offerings, captured in September 2026 |
| SaaStr/Gorgias: median automation of 48% | Plan conservatively and measure in the pilot | Provider evaluation; host is an investor |
| Pylon: 50% of tickets deflected, same team size | Measure team time, not just tickets | Single-case report, no published data |
Frequently asked questions
- How much does customer service AI cost?
- It depends on the pricing model and volume. Providers charge per user, per use (minute, message or credit), subscription with allowance or outcome. To compare proposals, calculate the cost per genuinely resolved interaction.
- What is outcome-based pricing for customer service AI?
- It is a model where the company pays per interaction deemed resolved rather than per user or minute. The critical point is the definition: some providers count a customer who requests no more help after an AI response as an outcome.
- What is AI's real resolution rate in customer service?
- It varies by operation. A comparative evaluation cited by SaaStr in October 2026 found median automation of 48% among providers, below the 70% to 100% commonly advertised. Measure in your own pilot.
- Per-minute or outcome-based pricing: which is better?
- Per-minute pricing is easier to audit and predict but increases with long conversations. Outcome-based pricing connects price to value but depends on the outcome definition. Compare both using cost per resolved interaction.
- Does AI reduce customer service team size?
- Not always. Cases AI resolves autonomously tend to be the simplest. In a case reported by Pylon in 2026, AI deflected around 50% of tickets while team size remained unchanged.
Sources
Videos, studies and documents used in this article. Preprints are versions published by their authors before peer review. This article was researched and written with AI assistance; figures, quotes and dates were checked against original texts and video transcripts on October 26, 2026.
Research paper (preprint)
From Tokens to Credits to Outcomes: Pricing AI Across Models, Applications, and Cloud Platforms (opens in a new tab)Pratik Gupta (arXiv 2610.30701, version 1) · published on
Pricing pages from 75 AI offerings captured on September 16, 2026. The author works at Stripe, which was excluded from the sample.
Video
Agent Pricing is Chaos. Here's What We're Seeing From the Buyer Side | The Agents #015 (opens in a new tab)SaaStr AI · published on
Report on Gorgias's comparative evaluation: median automation of 48%. SaaStr discloses an investment in Gorgias.
Video
SaaStr AI Day: The New Agentic Support Playbook (opens in a new tab)Marty Kausas (Pylon), SaaStr AI channel · published on
Case with around 50% of tickets deflected and no change in team size. Pylon sponsors SaaStr.
Reference page
Intercom: pricing (opens in a new tab)Intercom · accessed on
Fin price per outcome and criteria for counting an outcome.
Read more
Want to serve your customers by video?
Take the free assessment or try Neft for 15 days, with no credit card.


