Why Deskpro doesn’t charge for AI resolutions

If you’re currently using or evaluating AI-powered help desk software, there’s a good chance you’ve encountered resolution-based pricing. Well-known vendors including Zendesk, Help Scout, and Fin (Intercom) offer AI tools that are priced per successful resolution. It’s most frequently used with AI chatbots, with the vendor defining success as the chatbot solving a customer issue without human agent involvement (the exact success criteria varies by vendor and can be nebulous).

The case these vendors make is that the cost of a successful AI resolution is significantly less than the cost of a ticket that requires human assistance. Kayako, a vendor selling an autonomous chatbot called Kay, claims the average cost of a human-resolved chat interaction is $4.60 compared to just $1 for a Kay-resolved interaction.

It’s an appealing marketing claim, but it overlooks the fact that resolution-based pricing is significantly more expensive than the actual cost of AI token usage and, in the case of most vendors, represents an additional monthly cost on top of a seat-based help desk plan. It’s a billing model that can be unpredictable, especially during peak volume periods or when your organization's definition of success differs from the vendor’s. And, when a vendor defines an issue as resolved but a customer ends up coming back to a human agent later, you pay for the AI fee and the human agent’s time.

At Deskpro, we’ve made the choice not to use resolution-based AI pricing because we want our customers to use our AI help desk features on their terms, without having to guess what their monthly bill will be or dispute the definition of a resolution.

We also made a conscious decision to build our product in a way that allows our customers to bring their own AI. If they have an existing commercial relationship with an LLM provider, why should they not be able to leverage that investment to drive our AI help desk capabilities? If our customers don’t have their own agreement with an LLM vendor or simply don’t want to use it, they can use Deskpro’s managed connection. In either case, you only pay for the tokens you use. You end up paying, on average, a fraction of a cent per completed AI action compared to the $1-2 you would pay with resolution-based pricing (more on that later).

How resolution-based pricing works (and where it falls short)

Resolution-based pricing stems from a deflection-forward approach to customer support and makes the most sense for organizations that field a lot of simple, repetitive requests. The idea is that a customer chats with an AI chatbot, the chatbot solves the customer’s issue without any human intervention, and the vendor charges a set amount for that successful interaction.

That set amount is significantly marked up from the actual cost of AI token usage. The argument the vendor makes is that you’re only paying for success, and that because success is based on deflecting tickets that would otherwise go to a human agent, you’re saving money on labor. If every AI chat actually resolved the customer’s issue, that would be true. But in reality, what a vendor records as a success isn’t always a success for your business or customer.

Consider a customer who asks a question about a confusing bill they received and gets a response from the chatbot that isn’t helpful. They may abandon the chat rather than escalate it, then submit the same question through another channel (phone, email, or contact form), where it gets picked up and resolved by a human agent. In cases like this, your business essentially double-pays for the interaction.

Even when the resolution is real, the savings only materialize if deflection reduces what you spend on labor, and labor doesn’t scale in fractions. Take a two-person team handling 1900 tickets a month, as discussed in this Reddit thread:

A single agent working at a pace of 30-35 tickets a day can take on roughly 650-760 tickets a month, which means a chatbot would need to deflect at least 60% of the total ticket volume before the remaining workload would fit within one agent’s capacity. Below that threshold, you keep paying two full salaries plus the vendor’s per-resolution fees.

In other words, if you can’t successfully deflect 60% or more of your tickets, you may not break even on a resolution-based chatbot.

How Deskpro’s token-based AI pricing works

Deskpro’s approach to AI is built around transparency and flexibility. You can choose to power our AI features with our default model (OpenAI’s GPT) or any other model your organization is already using. If you use our default model, you pay for your token usage through service credits that you load into your account. This means you’re paying what the AI model provider (in this case, OpenAI) charges per token–there’s no markup from Deskpro.

If you bring your own model, you’ll keep paying your model provider–not Deskpro–for token usage. There’s no fee to connect your own model, and you can switch providers any time you want, meaning you can shop around for the model that best meets your needs and budget.

In both cases, the cost for AI to complete a help desk task (responding to a customer via chat, summarizing a ticket thread, suggesting a reply to a human agent, and so on) is a fraction of the cost of other major vendors’ resolution-based rates.

Doing the math: Resolution-based vs. token-based costs

Token-based usage costs mean you’re paying when AI does anything, not just when it autonomously “resolves” an issue according to a vendor’s criteria. That means you’re paying for more actions, but you’re also paying at a much, much lower rate. To illustrate this, I’ve pulled together a comparison of several major help desk vendor’s AI pricing.

To get the Deskpro cost per AI-resolved issue, I averaged the cost of a singular token (segmented by input and output costs) across 10 popular frontier models from OpenAI, Claude, and Gemini. My Engineering team took a very conservative approach to estimating the number of tokens consumed by a simple question that would generate an AI chatbot response, broken out by input and output. I used these numbers to get a total estimated cost of the AI interaction: $0.0067. A number that’s over 100x less than the cost of an AI resolution from the least expensive of the major resolution-based vendors.^[1]

A more complex query would obviously burn more tokens. But 100x as many?

AI that puts your team in control

Deskpro’s choice not to charge per AI resolution is a very intentional one. My belief is that AI should enhance human-led customer support, not replace it. Resolution-based AI pricing is a model that rewards deflection, requiring teams to automate the majority of chat interactions just to get to a breakeven point. Don’t get me wrong: there will be cases where this makes sense, especially in support organizations that handle a high volume of straightforward, repetitive queries. But many support teams–especially those in B2B, regulated industries, and other spaces that handle more complex customer or client issues–need human agents for their empathy and problem-solving skills. In these cases the best use of AI is to assist, not replace.

Token-based pricing works well for teams that use AI consistently as an assistant. They may use an AI chatbot with a clear escalation path to human agents, but they also use behind-the-scenes AI features like intelligent ticket triage, ticket summarization, suggested replies, and knowledge base article generation. These features keep human agents in the driver’s seat while allowing them to work more efficiently, freeing up time to focus on the complex issues where they’re needed most.

The model also works really well with Deskpro’s bring-your-own-AI option. If our customer already has a commercial relationship with, say Google for Gemini, then they are only paying for token usage, not resolution. It’s central to our overall approach to providing customer-centric help desk software: we want you to set up your help desk on your own terms, in the optimal environment and with the AI models and commercial structures that you decide are best for your organization.

If you would like to learn more about Deskpro’s AI features or pricing model, please reach out to us.

[1] For those that are interested in a breakdown of the math: I started with the average input and output costs for a single token across 10 popular LLMs. That came out to $0.0014925 for an input token and $0.00524 for an output token. I estimated that a simple AI chatbot interaction, in which the chatbot generates a response based on a knowledge base article, would require approximately 3000 input tokens and 2000 output tokens.

Input cost: 3,000 tokens × $0.0000004975 = $0.0014925

Output cost: 2,000 tokens × $0.00000262 = $0.00524

Total: $0.0014925 + $0.00524 = $0.0067 per interaction