Short answer: AI customer service software is sold in two shapes. One is an AI agent bundled into the help desk you already run (Zendesk, Freshdesk, Intercom, Salesforce). The other is a standalone AI agent platform that sits in front of whatever help desk you own (Ada, Sierra, Decagon). The features overlap more than the price tags suggest. What actually separates them is the billing unit, and no two vendors define that unit the same way.
That matters more than any feature checklist. One vendor bills $0.99 per resolved conversation, another bills $2.00 per conversation whether it resolved or not, a third bills sessions in blocks of 100, and a fourth bills credits per action so a single customer question can be charged four times. Run the same 10,000 tickets through those four models and the invoices are not close.
Last updated September 2026. Every price below was checked against the vendor's own pricing page or official documentation on September 1, 2026.
The two categories that share one search term
Roundup posts list Intercom next to Sierra as if you would evaluate them side by side. In practice you rarely do, because they attach to your support stack in different places and get bought by different budgets.
| AI inside your help desk | Standalone AI agent platform | |
|---|---|---|
| Examples | Zendesk AI, Freshdesk Freddy, Intercom Fin, Salesforce Agentforce | Ada, Sierra, Decagon |
| Where it sits | Native inside the ticketing tool | In front of your existing help desk, via API |
| Who buys it | Support leader, on the existing help desk contract | VP Support or COO, usually a separate procurement |
| Setup effort | Days, points at your existing knowledge base | Weeks, includes integration and custom actions |
| Published pricing | Yes, on the pricing page | No, quote only |
| Typical entry cost | Hundreds of dollars a month | Six figures a year, per third party benchmarks |
The practical rule: if your ticket volume is under roughly 5,000 a month and your knowledge base is decent, the AI in your current help desk will do most of what the standalone platforms do, at a fraction of the commitment. Standalone platforms earn their price when you need custom actions against your own systems, such as issuing a refund, changing a shipping address, or rebooking an appointment without a human ever touching it.
Verified pricing, September 2026
These are the numbers the vendors publish themselves. Where a vendor publishes nothing, the row says so rather than repeating a figure from a directory.
| Vendor | Human seat cost | How the AI is billed |
|---|---|---|
| Intercom | Essential $29, Advanced $85, Expert $132 per seat per month | Fin AI Agent from $0.99 per outcome. Included on all plans. |
| Zendesk | Support Team $19, Suite Team $55, Suite Professional $115 per agent per month, billed yearly | Automated resolutions. You pay only when the AI resolves without escalating to a human. Zendesk does not publish the per resolution rate. Copilot add on is $50 per agent per month billed yearly. |
| Freshdesk | Growth $19, Pro $55, Enterprise $89 per agent per month, billed annually | Freddy AI Agent: first 500 sessions included, then $49 per 100 sessions. Freddy AI Copilot add on $29 per agent per month billed annually. |
| Salesforce Agentforce | Requires a Salesforce edition | $2.00 per conversation, or Flex Credits at $500 per 100,000 credits. One action consumes 20 credits, so $0.10 per action. |
| Ada, Sierra, Decagon | Not applicable, sits over your help desk | Quote only. None of the three publishes a rate. Treat any number you find in a comparison post as unverified. |
Two things in that table are worth pausing on. Salesforce's own documentation lists the conversation price and the Flex Credit packs and does not list a per user Agentforce price, even though several software directories publish one. And the third party figures floating around for Sierra and Decagon contradict each other from post to post, which is the usual sign that nobody has rechecked them. Ask for a written rate card instead of trusting either.
What counts as a resolution is the whole negotiation
Every consumption priced vendor uses a word like resolution, outcome, conversation, or session. They are not synonyms, and the difference is most of your bill.
| Billing unit | What triggers a charge | Who it favors |
|---|---|---|
| Resolution with no escalation | Only when the customer leaves satisfied and no human touched it | You. The vendor absorbs failed attempts. |
| Outcome | Vendor defined. Usually a resolved query, but read the definition. | Depends entirely on the contract wording |
| Conversation | Any conversation the AI enters, resolved or not | The vendor. You pay for failures too. |
| Session | A time boxed window of activity, often sold in blocks | The vendor, if your traffic is spiky |
| Credits per action | Every backend step. One question can be several actions. | The vendor, and it is the hardest to forecast |
Take 10,000 AI handled conversations a month with a 60 percent success rate. On a resolution model with no escalation you pay for 6,000. On a per conversation model you pay for all 10,000, including the 4,000 the AI failed and handed to a person, so you pay twice for those tickets. On a credit model, if the average query fires three actions, you are billed 30,000 actions. Same traffic, three very different invoices.
Ask for the definition in writing, get it into the order form, and ask what happens on a reopened ticket. A customer who comes back the next day about the same problem is often billed as a second resolution.
The six capabilities that actually differ
| Capability | Why it matters | What to check |
|---|---|---|
| Grounding | Stops the AI inventing policy | Does it answer only from your approved content, and does it refuse when it has no source? |
| Actions | The difference between answering and resolving | Can it write to your systems (refund, cancel, update address) or only read? |
| Escalation handoff | Protects the customer experience when AI fails | Does the human agent get the full transcript and the AI's reasoning, or start cold? |
| Channel coverage | Rates often differ by channel | Is voice priced the same as chat? Salesforce charges more credits for voice actions. |
| Reporting | How you audit the invoice | Can you export the raw list of billed events, not just a dashboard total? |
| Language coverage | Quality drops fast outside English | Test your actual second language, not the demo's |
Grounding and actions are where the real gap sits. A tool that only answers from a knowledge base is a better search box. A tool that can take an action on your behalf changes your cost per ticket, because it removes the human entirely rather than shortening the human's work.
Five tests that break a demo open
Vendor demos are built on clean data. Yours is not clean. These five questions have changed buying decisions more than any feature list.
- Ask it something your knowledge base does not cover. A good agent says it does not know and offers a human. A weak one invents an answer. Do this before anything else.
- Bring your own tickets. Hand over 50 real anonymized tickets from last month, including the messy ones, and ask them to run those rather than their sample set.
- Ask for the billed event export. If the vendor cannot show you a row level export of what it would charge for, you cannot audit the invoice.
- Ask what a reopened ticket costs. The answer tells you how honest the resolution definition is.
- Ask who writes the fallback content. Most disappointing rollouts are a content problem, not a model problem. If your knowledge base is thin, fix that before you buy.
One more practical step: before you sign a consumption contract, pull your own ticket export and count how many of last quarter's conversations would have met the vendor's definition. Support platforms bury this in dashboards, so it usually means working directly with the raw export, and being able to ask your ticket data questions in plain English gets you to a number in an afternoon instead of a week. Walking into the negotiation with your own count is the single biggest piece of leverage you will have.
Mistakes that cost real money
Buying on the headline rate. A $0.99 outcome and a $2.00 conversation are not comparable numbers, and neither is comparable to $49 per 100 sessions. Model your own volume against each definition.
Replacing the help desk to get the AI. You almost never need to. Standalone platforms exist precisely so you can keep your ticketing system and add an agent in front of it.
Measuring deflection instead of resolution. A customer who gives up and leaves counts as deflected in most dashboards. That is not a win. Track first contact resolution and repeat contact rate alongside the deflection number, or you will celebrate abandonment.
Skipping quality review on AI conversations. AI handled tickets need the same sampling discipline as human ones. The same quality management tooling should score both.
What is the best AI customer service software?
There is no single best. For teams already on Zendesk or Freshdesk, the native AI agent is almost always the right first move because setup is days and the pricing is published. For high volume teams that need the AI to take actions in custom back end systems, a standalone platform such as Ada, Sierra, or Decagon is worth the procurement effort. Intercom sits in between, with published per outcome pricing and a strong agent.
What counts as a resolution in AI customer service pricing?
It depends entirely on the vendor, and it is not standardized. Zendesk bills an automated resolution only when the AI resolves the request with no escalation to a human. Intercom bills per Fin outcome from $0.99. Salesforce bills $2.00 per conversation regardless of whether it resolved. Freshdesk bills sessions in blocks of 100. Get the definition written into the order form.
Do I need to replace my help desk to add an AI agent?
No. If you run Zendesk, Freshdesk, Intercom, or Salesforce, an AI agent is already available inside the product you pay for, usually as an add on or a consumption line. If you run something without a strong native agent, standalone platforms connect over API and sit in front of your existing ticketing. Replacing a help desk to get AI is an expensive way to solve a problem that has a cheaper answer.
Can AI customer service software work with Zendesk or Freshdesk?
Yes, in two ways. Both vendors ship their own AI agents natively, so the simplest path is turning on what you already own. Independent platforms also integrate with both over API, reading your knowledge base and writing back into tickets. The thing to verify is write access: some integrations can read your content but cannot update a ticket, tag it, or close it, which leaves an agent doing cleanup behind the AI.
Is AI customer service software worth it for a small support team?
Often yes, but only on published, low commitment pricing. A three person team on Freshdesk gets 500 Freddy sessions included before paying anything extra, which is enough to find out whether the AI handles your top questions. What is not worth it for a small team is a six figure standalone contract, or any deal with a large annual minimum. Start with the AI in the tool you already pay for, measure it for a quarter, then decide.
Where this fits with the rest of your support stack
AI agents are the newest layer on top of work most teams have already started. If you have not yet mapped which requests are repetitive and rule driven, that comes first, and the broader customer service automation picture is the right place to start. If you want the conceptual background on how these systems work and where they fail, rather than what they cost, the companion guide on AI customer service chatbots and agents covers the five layers and the deflection trap in detail. And if your volume problem is really a contact volume problem, the ticket deflection math will tell you how much headroom an AI agent can realistically buy you.