AI customer service software compared: what it really costs

AI customer service software answers routine customer questions on its own, looks up orders and bookings, and hands everything else to your team. Most help desk vendors now charge for it per conversation the AI resolves, at list prices between $0.50 and $2.00 in October 2026, on top of what you already pay for seats. The catch is that each vendor defines "resolved" differently, so the same 1,000 conversations can cost $300 with one product and $2,000 with another. This guide compares eight products on their own published prices, shows the arithmetic, and explains where the AI should and should not be trusted.

What does AI customer service software actually do?

It sits in your help desk or chat widget and works on incoming messages from chat, email and sometimes WhatsApp, social media or the phone. A good one does four things:

1. Answers questions from your help articles, policies and past replies.
2. Looks things up, such as where an order is or when a booking starts, through connections to your systems.
3. Takes simple actions under rules you set: changing a delivery address, cancelling within the allowed window, booking a slot.
4. Hands over to a person with the full conversation and a short summary, so the customer doesn't repeat themselves.

The first is the easy part. The last three are what separate a useful agent from an expensive search box, and they depend on how well it is connected and tested, which is the same whichever product you choose.

Eight AI customer service products compared

Prices below are the vendors' own published list prices, checked on 17 October 2026. They change often, so check before you buy.

What counts as a "resolution"?

This is where like-for-like comparisons fall apart. Before you compare prices, compare what triggers a charge.

The practical test: ask each vendor what you would be charged for a conversation where the AI answered, the customer went quiet, and then emailed a person the next day. The answers differ.

What does AI customer service cost at real volumes?

The chart shows the AI fee alone for 1,000 resolved conversations a month. Seats and plan fees are left out so the products compare on the same footing.

Chart

AI fee for 1,000 resolved conversations a month

• Zendesk AI agents (pay as you go)$2,000

  $1,500 if you commit to a volume in advance.

• Gorgias AI Agent (monthly contract)$1,000

  $900 on an annual contract.

• Intercom Fin$990
• Freshdesk Freddy AI Agent$833

  Billed per session, resolved or not. Assumes 6 in 10 sessions end resolved.

• Help Scout AI Answers$750
• HubSpot Customer Agent$500
• Salesforce Agentforce (Flex Credits)$300

  Billed per action. Assumes three actions per conversation.

• Custom agent on GPT-6.1 Sol (model cost only)$24

  Building, hosting and looking after it cost far more than the model.

Show the data as a table

Product	Billing unit	List price	AI fee for 1,000 resolutions
Zendesk AI agents	Automated resolution	$2.00 pay as you go, $1.50 committed	$2,000 ($1,500 committed)
Gorgias AI Agent	Automated interaction	$1.00 monthly, $0.90 annual	$1,000 ($900 annual)
Intercom Fin	Outcome	$0.99	$990
Freshdesk Freddy AI Agent	Session	$49 per 100	$833 (1,667 sessions)
Help Scout AI Answers	Resolution	$0.75	$750
HubSpot Customer Agent	Resolved conversation	$0.50 (50 credits)	$500
Salesforce Agentforce	Action (Flex Credits)	$0.10	$300 (3 actions each)
Custom agent, GPT-6.1 Sol	Tokens	$2 in, $10 out per million	$24 (model only)

The same arithmetic at three sizes of business looks like this. Two products charge for conversations rather than resolutions, so we have assumed the AI fully resolves 6 in every 10 conversations it handles; if yours does better, those two get cheaper.

Three things change the real bill:

• Seats. A five-person team on Zendesk Suite Team pays $275 a month in seats on annual billing before any AI fee; the same team on Help Scout Standard pays $125.
• Included allowances. Zendesk includes 5 to 10 automated resolutions per agent, Gorgias 30 to 530 interactions depending on plan, and HubSpot plans include credits worth 10 to 100 resolutions a month.
• The work around it. Writing the policies the AI follows, connecting it to your order or booking system, and testing it on real questions takes the same effort on every product.

The custom row is not a typo. Running a capable model costs about 2.4 cents a conversation at GPT-6.1 Sol's $2 and $10 per million tokens, assuming 8,000 tokens in and 800 out. But a custom agent has to be designed, connected, tested and kept up to date, and that is where the money goes. For a deeper look at models and their prices, see our comparison of ChatGPT, Claude, Gemini, Grok and Muse.

Which one fits your business?

What should AI handle, and what should it never touch?

The best results come from deciding this before launch, not after the first bad conversation.

Klarna is the best-known warning. In early 2024 it said its AI assistant was doing the work of 700 customer service agents; by May 2025 its chief executive told Bloomberg it was hiring people again because quality had suffered, so customers could always reach a person.

What do customers think about AI customer service?

Honestly, they are sceptical. In a Gartner survey of 5,728 customers, published in July 2024, 64% said they would prefer companies didn't use AI for customer service, and 53% said they would consider switching to a competitor if a company did. Their biggest worry was finding it harder to reach a person.

Gartner also expects that to change: in March 2025 it predicted that by 2029, agentic AI will resolve 80% of common customer service issues without a person, cutting operating costs by 30%. Both can be true. Customers don't object to a fast, correct answer from AI; they object to being trapped. The hand-over to a person is not a fallback, it is part of the product.

Rules you need to follow

• Tell people it is AI. In the European Union, Article 50 of the AI Act has required since 2 August 2026 that AI systems talking to people make clear they are AI, unless it is obvious. One line in the greeting covers it, and it is good practice everywhere else.
• You are responsible for what it says. When Air Canada's website chatbot invented a refund rule in 2024, a Canadian tribunal held the airline "responsible for all the information on its website". Your AI should answer from your actual policies, which is what a RAG knowledge base does; our guide to what RAG is explains how.
• Protect customer data. Use business plans with a data processing agreement, check where conversations are stored, and keep payment details out of chat.

How to launch AI customer service in four steps

1. Read last month's inbox. Sort a few hundred real conversations into types, and count which questions come up most and how they were answered. That list is your starting point, whatever product you choose.
2. Write the playbook. Your policies, tone, limits and hand-over rules, written down and agreed. The AI can only be as good as this.
3. Start with drafts. For the first two to four weeks, the AI drafts replies and your team sends them. You see exactly how accurate it is before any customer does.
4. Switch on gradually. Let it answer on its own for the question types it got right consistently, keep everything else on drafts, and widen the list as the numbers earn it.

What to measure

• Resolution rate, using your own definition, not just the vendor's billing count.
• Re-contact rate: how many "resolved" customers come back about the same issue within seven days. This catches silent failures.
• Satisfaction on AI conversations compared with human ones.
• Time to reach a person when the AI hands over.
• Cost per resolution, including seats and the time spent maintaining the AI's knowledge.

What we would do

For most small and medium-sized businesses, we would start with the AI built into the help desk you already use, run it on drafts for a month, and measure re-contact rate before letting it answer alone. If you don't have a help desk yet, a low per-resolution product with a free trial, such as Help Scout, is the cheapest way to learn what your customers actually ask. A custom agent earns its place later, when volumes are high, when the job spans systems no product connects, or when the AI needs to take real actions, such as the ones we describe in our AI agent examples.

Whichever route you take, the work that decides success is the same: clean policies, good connections to your systems, and a hand-over that works. That is what our AI customer service setup covers, and we build website and WhatsApp assistants and custom integrations when a product isn't enough.