Intercom Fin vs Zendesk AI vs a Custom AI Agent: What Each Can and Cannot Do (2026)

Four vendors sell a ready-made AI support agent that US buyers search for by name: Intercom (Fin), Zendesk (AI agents), Salesforce (Agentforce) and Gorgias (AI Agent). On 11 October 2026 we opened 21 pages on those four companies' own sites, a mix of pricing pages, product pages and help articles. This guide sets what we read next to a fifth option, an agent built for you.

Several of the comparison pages we read line these products up by price per ticket. That does not work, because the four bill in four different units and a custom build is priced as a project. So this guide compares them on what each can do: the help desk it runs on, the channels it answers, the systems it can act on, and what it needs from you.

The general choice between building and buying, with builders such as n8n, Make and Copilot Studio, is in our build versus buy guide for AI agents. We do not repeat it here.

We are FactoryJet, an AI services company that designs, builds, implements and supports custom AI agents and the integrations behind them. FactoryJet was founded in 2014 by Bhavesh Barot and has served more than 500 businesses. A custom build is what we sell, so this page also says plainly where a rented agent is the better choice.

Fin, Zendesk AI, Agentforce, Gorgias and a custom agent at a glance

Start with the column on the right. The best first choice is usually the agent that belongs to the place your tickets already live.

What each vendor charges, and what counts as a billable result

Read the unit before the price. An outcome, an automated resolution, an action and an automated interaction are four different things, and each vendor writes its own definition.

Four details in those definitions change what a bill looks like.

• Fin bills an assumed resolution. Intercom's outcomes article counts a resolution when the customer confirms the answer helped, and also when the customer leaves without asking for more. A handoff counts too if you wrote it into a Procedure. Intercom says it does not bill when a customer asks for a human or a Procedure fails, and that it deducts the charge if the customer comes back to the same conversation for more help.
• Zendesk changed how it counts on 18 May 2026. Its help center now describes three resolution tiers. An assisted escalation (the agent helped, a person finished) and a contained resolution (the customer asked for nothing more, but the model check did not confirm the fix) do not count against your allowance. A verified resolution, which passed the check, does. Accounts on the earlier platform stay there until they move, so ask Zendesk which one yours is on.
• Agentforce meters work done. Each action draws Flex Credits, and the worked examples on the Salesforce pricing page price one standard action at $0.10. We found nothing on that page that ties the charge to a resolved conversation. The $2 per conversation option is the alternative for customer-facing agents.
• Gorgias gates actions by plan. Its plan comparison table shows Actions and brand voice controls on Basic, Pro and Advanced, and a dash under Starter.

The headline resolution rates are not measured the same way

Each vendor leads with a number. Intercom says Fin resolves 76 percent of customer queries on average. Zendesk's pricing page says its AI agents resolve up to 80 percent or more of service issues. Gorgias shows 48 to 53 percent of conversations automated across five industries. One is an average, one is a ceiling and one is a set of industry figures, and each vendor counts a resolution its own way.

The Gorgias page is the most useful of the three, because it breaks the number down by type of conversation. For fashion and apparel brands over the last 90 days it shows 82 percent of product questions closed by the AI Agent, and 31 percent of billing and account conversations. Questions answered from content close at a high rate. Conversations that need a billing or account system close at a much lower one. That gap is where connectors, custom actions and custom builds earn their place.

What a custom build costs to commission

A custom agent is priced as a project. As a market reference, development firm ProductCrafters puts initial development of a customer service AI agent at $10,000 to $50,000, and at $25,000 to $50,000 for agents that handle refunds, access order data and escalate complex issues (page modified 20 May 2026). That is a market range from a named firm and no FactoryJet price. FactoryJet quotes the build and the ongoing support as fixed prices in writing after a short scoping call, and we keep managing the servers, the AI, the APIs and the maintenance so your team does not have to.

Which channels each agent answers, and which systems it can act on

All four rented agents can now take actions, and none of them is limited to quoting help articles. So the old split between a chatbot and an agent no longer separates them from a custom build. Two narrower questions do: whether the vendor's builder can reach the system, and whether it can express your rule.

Intercom Fin: an agent you can add to the help desk you already have

Fin is the AI agent made by Intercom. It runs inside the Intercom help desk, and Intercom also sells it on its own for teams on another help desk. Its integrations page lists Zendesk, Salesforce Service Cloud, HubSpot Service Hub and Freshdesk, plus any other help desk over an API. On Zendesk, Intercom says Fin connects through the API, updates tickets and custom fields, and hands off to your agents inside Zendesk.

For multi-step work Fin uses Procedures. Your support team writes the steps in plain language, adds if/else rules, uses code where a rule must be exact, and connects systems such as Stripe or Shopify through Data Connectors. Intercom says Procedures need "no engineering or custom builds". Its ecommerce example has Fin checking an order, confirming the change and issuing a refund.

Where it fits. You want to keep your help desk, your help center content is in good shape, and the systems Fin must act on have an API or an MCP server (a standard way for AI tools to call other software).

Where it stops.Every connector goes through an API or MCP. The Fin integrations FAQ says, "If your tool has an API, you can connect it to Fin." A system with no API needs one built in front of it first. On a help desk other than Intercom, a minimum of 50 outcomes a month applies.

Confirm phone pricing in writing. The Intercom pricing page lists Fin Voice at $1.99 per voice outcome through its sales team, while the Fin pricing page says voice is on a custom pricing plan for select customers.

Zendesk AI agents: included in the plan, billed on verified resolutions

Zendesk says AI agents are included in every Suite and Support plan. They answer on messaging, email and voice. You describe a policy in plain language, Zendesk generates a procedure from it, and the agent follows that procedure while it reasons through the conversation. The pricing page lists support for more than 80 languages.

To act on other systems, a Zendesk AI agent calls custom actions and API integrations. The help article on the no-code integration builder lists six ways to sign in to an API (API key, bearer token, basic auth, OAuth 2.0, AWS signature and a custom option) and says request bodies support JSON only so far. The same article says new integrations should now be created as custom actions, and that access is limited to Client Admins.

Voice needs Zendesk voice. Zendesk says voice AI agent calls carry telephony charges, a charge for the time the agent is on the call and a recording charge, and can also use automated resolutions.

Where it fits. Your team already works in Zendesk. The agent is in the plan, the allowance is in the plan, and there is one vendor to deal with.

Where it stops. Seats are priced per human agent, and the systems the agent can reach are the ones with a JSON API. For teams on another platform, Zendesk now sells Forethought AI agents, which its pricing page describes as working whether you use Zendesk or another service provider.

Salesforce Agentforce: actions built from what already lives in Salesforce

Agentforce is the agent layer of the Salesforce Platform. Salesforce says you can turn your existing workflows, prompt templates, Apex code and APIs into agent actions, with MuleSoft, Slack and Data 360 available as tools. It offers ready-made agents for service, sales, marketing and commerce, and says agents connect to your website, WhatsApp, Messenger and the CRM itself. Agentforce Voice covers phone calls.

There are three ways to pay. Flex Credits cost $500 per 100,000, at 20 credits per standard action. Conversations cost $2 each for customer-facing agents. Per-user licenses cover staff-facing use, with add-ons at $125 per user per month and Agentforce Max editions from $550. The Salesforce licenses underneath are separate. Service Cloud editions listed at $25 to $550 per user per month when we read the page.

The edition names differ between two Salesforce pages. The Service Cloud pricing page lists Starter Suite, Pro Suite, Core, Advanced and Max. The Agentforce FAQ says Foundations, the $0 way to try it with credits for the first 1,000 Agentforce for Service conversations, is an add-on for "Enterprise Edition and higher". Ask your account executive which editions qualify today.

Where it fits. Your cases, customer records and automations already live in Salesforce, and you have an admin or developer who knows flows.

Where it stops. Outside Salesforce. The agent reaches other systems through APIs and MuleSoft, which is integration work whichever tool does it. The pricing page also notes that its worked examples leave out Data 360 credits and other consumption services.

Gorgias AI Agent: built around a Shopify store

Gorgias is a help desk made for ecommerce. Its pricing page says the help desk is "never priced per agent". Plans go by ticket volume, from 50 tickets a month to 5,000.

The AI Agent is tied to the store. Gorgias docssay, "You get a dedicated AI Agent for each Shopify store you connect." It answers on email, chat, SMS, Instagram DMs, Facebook Messenger and WhatsApp, in more than 80 languages. It plays two roles: support agent after the purchase and shopping assistant before it.

Its actions are about orders. The product page shows refund, cancel, edit or replace an order, update an address, start a return, apply a discount, and pause or skip a subscription, from a library of more than 130 actions, plus custom actions to any API or app. Gorgias says a support team can set it up with no engineers and be live in under an hour.

Plan gating matters here. Actions and brand voice controls start on Basic. The help desk connects to Shopify and WooCommerce on every plan, to BigCommerce from Basic and to Magento from Pro.

Where it fits. A Shopify brand whose tickets are about orders, with a support lead and no developer.

Where it stops. When orders, stock or customer prices live outside Shopify, for example in an ERP that serves wholesale accounts. Ask Gorgias what the AI Agent can act on in that case before you plan around it.

A custom-built agent: for work that leaves the help desk

A custom agent is software written for your process. It can sit inside Zendesk, Gorgias or Intercom through their APIs, so your team keeps its inbox, or it can run in a channel of your own. FactoryJet designs, builds, implements and supports agents like this, and the client owns what we build.

It starts from the other end. A rented agent starts with the vendor's builder and reaches out to your systems. A custom agent starts with your systems and rules: the ERP that holds customer prices, the approval a refund needs, the warehouse feed that says where the parcel is.

It is the slower route. Intercom and Gorgias both say setup takes under an hour, and Zendesk says minutes. As a guide, a custom pilot on one narrow workflow usually takes two to four weeks, and a production agent with permissions, logging and monitoring six to twelve weeks. So build only when one of the five conditions further down holds.

When a rented agent is the right call

Rent if most of these five describe you.

1. Your tickets are common ones. Order status, returns, product questions, password resets and plan questions. These are the types the vendors built for, and the Gorgias figures above show where they close best.
2. Your team already lives in that vendor's product. A Zendesk team has an allowance of automated resolutions in its plan. A Salesforce team can reuse the flows it already built.
3. The systems the agent must touch have a modern API or a listed connector. Fin listed 60 connectors on the day we read its page, and Gorgias a library of more than 130 actions.
4. Nobody on your side wants to own software. The vendor runs the models and the hosting, and ships the updates.
5. You need to be live this month. Every vendor here offers a way to try it first: a 14-day trial at Intercom and at Zendesk, a free trial at Gorgias, and Salesforce Foundations with credits for the first 1,000 Agentforce for Service conversations.

If that is you, rent. If an off-the-shelf agent will do the job, we will tell you so.

When a custom build is the right call

Build when at least one of these five is true. Each one names something the vendor pages told us.

1. A system the agent needs has no API that a vendor builder can call. The Zendesk builder sends JSON request bodies only, and Fin connects to tools that have an API or an MCP server. An older ERP that trades EDI files, a SQL database on your own server or a supplier portal with no API needs code in between.
2. The rule crosses systems and needs approval. Take a refund above your limit. It has to wait for a manager to approve it, then post a credit note in the accounting system, then update the order and the ticket. A rented agent can do the pieces its connectors reach. When the chain runs through systems they do not reach, or the approval lives in a tool outside the help desk, holding it together with a record of who approved what is custom work.
3. The work is not a support conversation. Quote requests, purchase and sales orders and daily bookkeeping arrive by email and end in an ERP. FactoryJet has worked on Odoo, NetSuite, SAP Business One, ERPNext and custom ERPs on jobs like these. See AI agents for ERP systems.
4. You want to own the agent.With a rented agent, your procedures and settings live inside the vendor's product. With a build, the client owns what we build, which matters if you expect to change help desk or add channels later.
5. You already rent one, and the same ticket types keep reaching your team. Pull the handover report. If two or three types make up most of it, and each needs a system from point 1 or a rule from point 2, that is the piece to build.

The middle route: keep the rented agent and have the missing piece built

Renting and building are not exclusive. Every vendor in this guide lets its agent call something you supply.

• Fin accepts a custom REST API Data Connector or a custom MCP server.
• Zendesk AI agents call custom actions and API integrations.
• Agentforce turns your APIs into agent actions.
• Gorgias AI Agent lets you build a custom action to any API or app.

So the build can be small. One API in front of the ERP that answers three questions (where is this order, can this item be returned, what price does this customer pay) and performs one action can be enough to start. The rented agent keeps the conversation and the reporting. For a team that already pays for Fin or Zendesk, this is usually a smaller project than a full custom agent.

FactoryJet builds these connecting pieces. EDI integration, punchout catalogs, Xero, HubSpot and WhatsApp integrations are among them, along with AI integration work on systems the vendors do not list.

Zendesk, Intercom and Gorgias alternatives: when you can keep your help desk and change the agent

The table below takes four common reasons for looking at an alternative. Two of them do not need a new help desk. Intercom sells Fin to run on Zendesk, Salesforce, HubSpot and Freshdesk, and its page for Zendesk says you can deploy it "without adding per-agent costs or upgrading to a higher platform tier". Zendesk sells Forethought AI agents for other platforms. The agent and the help desk can now be bought separately.

Zendesk vs Intercom

Both sell a help desk by the seat and an AI agent billed on results. Zendesk seats list at $19, $55 and $115 per agent per month paid yearly, and Intercom seats at $19, $85 and $132 per seat per month paid annually. Zendesk includes an allowance of automated resolutions in the plan and bills $1.50 to $2.00 beyond it. Intercom bills $0.99 per outcome from the first one. The two count a result differently, as the section on billing units shows, so the unit prices do not settle the choice. Run both on the same tickets.

Gorgias vs Zendesk

Gorgias prices by ticket volume and Zendesk by agent seat. The Gorgias AI Agent is set up per Shopify store and its action library is about orders. Zendesk AI agents are not tied to one store platform, reach other systems through custom actions, and include voice AI agents. Gorgias describes its help desk as "designed specifically for ecommerce and Shopify". Zendesk describes plans for support teams of any kind, from teams that have outgrown a shared inbox to complex contact centers.

How to choose in one afternoon

You can settle most of this with a ticket export and a spreadsheet.

1. Export the last 90 days of tickets and sort them by type.
2. Take the five biggest types. Next to each, write the systems a person opens to resolve it.
3. Mark each system one of three ways: listed connector, has an API, or neither.
4. Start a trial of the agent that belongs to your help desk, using the trial routes listed above.
5. Run the same 20 real tickets through it, including a few it should refuse.
6. Whatever sits in the "neither" column, or fails step 5, is the list to bring to a builder.

Our build versus buy guide lists the four failure cases worth adding to that test. For how a support agent connects to a help desk under the hood, read the support agent architecture guide. For store-side tools beyond support, see the best AI agents for ecommerce.

How FactoryJet works on a custom support agent

We design, build, implement and support the agent, and you own what we build. To scope one we ask for three things: read access to your help desk or an export of your last 90 days of tickets, your written return, refund and shipping policies, and a list of the systems the agent would need to check.

We can show working software on your own data before a contract. FactoryJet quotes the build and the ongoing support as fixed prices in writing after a short scoping call. After launch we keep managing the servers, the AI, the APIs and the maintenance, so your team does not have to. We work remotely with US clients and schedule calls in US business hours.

To see the work, read about our AI customer support agents, Shopify AI agents and AI agent development. For phone work, see our AI receptionist and phone answering service. For a UK business, see renting or building AI customer service in the UK.

Sources

All pages opened and read on 11 October 2026. Vendors change prices and plan names often, so check the source before you budget.

• Intercom: Intercom pricing, Fin AI Agent pricing, Fin AI Agent outcomes (help article), Fin integrations, Fin for Zendesk, Fin for any helpdesk and Fin Procedures
• Zendesk: Zendesk pricing (US page), AI agents for service, About automated resolution tiers, About the automated resolutions platform (prior to resolution tiers), About AI agents with agentic AI, About the integration builder for AI agents and About voice AI agents
• Salesforce: Agentforce pricing, How Agentforce works, Agentforce Voice and Service Cloud pricing
• Gorgias: Gorgias pricing, AI Agent product page and AI Agent explained (docs)
• ProductCrafters, AI Agent Development Cost (modified 20 May 2026), for the market range on building a customer service agent
• Our own measurement: US monthly search volumes for the named comparisons, DataForSEO, 11 October 2026