Skip to content
AI Service Agent Guide

Independent comparison · checked October 2026

Open-source AI customer service agents that stop before billing

An open-source AI customer service agent is a running machine rather than a resolution price, and the difference shows on the ticket the knowledge base never answered. It reads the codebase and the billing account behind the complaint, then stops for a person before it changes either.

The open-source AI Operating System

Open source · Self-host is free · Approval rules sit in the project manifest

kortix.yaml
kortix.yaml
Approval gates you set. Off until you set them.
kortix.yaml
policies:  - match: gmail.send_email    action: require_approval    conditions:      - arg: to        match: "*@example.com"

Source: https://kortix.com/docs/project/manifest

Five agents on the terms a support lead checks

Kortix is our recommendation. Prices are each vendor's own, opened 27 October 2026.

Open sourceKortix (our pick)
Price model
Credits from $50 a month, at 100 credits per dollar.
Works with your helpdesk
Any helpdesk, through a connector; 3,000+ apps and MCP.
Where it runs
Kortix Cloud, your VPC, or your own on-prem network.
Where a person approves an action
Approval gates you set. Off until you set them.
Reaches your codebase and billing
Scoped access to the knowledge, the codebase and Stripe.
Open source
Yes
Intercom Fin
Price model
$0.99 per outcome; $49 a month minimum for 50.
Works with your helpdesk
Any helpdesk, including Salesforce and Zendesk.
Where it runs
Intercom's cloud.
Where a person approves an action
A person takes over when the customer asks.
Reaches your codebase and billing
API, MCP and data integrations.
Open source
No
Zendesk AI agents
Price model
€19 to €115 per agent a month, plus €1.50 a resolution.
Works with your helpdesk
Its own helpdesk, priced per agent.
Where it runs
Zendesk's cloud.
Where a person approves an action
A human agent takes the ticket.
Reaches your codebase and billing
Apps and APIs on Zendesk data.
Open source
No
Salesforce Agentforce
Price model
$500 per 100,000 credits; an action is 20.
Works with your helpdesk
Service Cloud, with Sales and Service editions.
Where it runs
Salesforce's cloud.
Where a person approves an action
A service rep handles the escalation.
Reaches your codebase and billing
Salesforce data and flows.
Open source
No
Lindy
Price model
$49.99 a month per active user for 5,000 credits.
Works with your helpdesk
1,500+ integrations plus MCP.
Where it runs
Lindy's cloud.
Where a person approves an action
Approvals built in.
Reaches your codebase and billing
Computer use and 40+ built-in skills.
Open source
No

Prices and plans come from each vendor's pricing page, opened 27 October 2026; Zendesk bills in euros. Capabilities are described from what each vendor publishes.

Pick the agent that fits your helpdesk

Three questions, then the check to run before you sign.

  1. Where does the answer have to come from?
  2. Who signs off a refund?
  3. Where must the agent run?

0 of 3 answered

Questions a support lead asks before signing

Short answers to the questions that decide the choice.

Can an AI agent refund a customer on its own?
Yes, if the tool you connect lets it, which is why the approval rule matters more than the model. In Kortix a matching call is held for a person and runs only after that decision; approval holds default to 24 hours and approvals.ttl sets the limit.
Do we have to replace our helpdesk?
No. Fin works with any helpdesk, including Salesforce and Zendesk, and Kortix connects to the helpdesk you already run instead of replacing it.
Can Kortix run on our own infrastructure?
Yes. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.
Can an agent work a real support queue today?
Yes. Kortix's own support inbox in Plain runs on an agent with scoped access to the product's knowledge, its codebase and Stripe; it investigates, resolves what it can, and stops at a person for sensitive actions.
Does every message need a person to start it?
No. An event trigger on the support inbox, gmail.message.received in the example, starts a session when a message arrives, and the agent triages the thread from there.

Put the approval rule in the repo, then let the agent work the queue.

Start with one support inbox and one rule that stops before money moves.