Best Tools for AI Support Agents
We ranked Intercom Fin, Zendesk AI, Freddy AI, Help Scout AI and Decagon on one question: can they actually stop a send, refund or close before it happens?
Key takeaways
- -The permission boundary that matters for a support agent is draft-vs-send, not read-vs-write. Reading a ticket and drafting a reply is low risk. Sending it, refunding it, or closing the account is a different category.
- -Most of the named platforms draft well. None of them, as documented in November 2026, enforces a hard, action-level permission boundary between drafting and sending that holds regardless of which client drives the call.
- -Pricing has shifted almost entirely to outcome-based billing (Intercom $0.99/resolution, Zendesk $1.50-$2.00/resolution, Freshdesk per session, Help Scout $0.75/resolution, Decagon sales-quoted). None of these numbers tells you what the agent is allowed to do.
- -Swytchcode is not a helpdesk. It is the execution gate in front of whichever support platform and LLM you use, enforcing draft-vs-send as policy at the point the call actually runs.
The permission that actually matters
Every AI support agent vendor talks about permissions. Most of that talk is about data access: which tickets the agent can read, which knowledge base it can search, which CRM fields it can see. That's worth controlling, but it isn't what decides whether a bad day turns into an incident.
The boundary that matters is draft versus send. An agent that reads a customer's ticket history and writes a draft reply is low risk, even if it's wrong, because a person reads it before anything happens. The same agent sending that reply, issuing a refund, or closing an account is a different risk category entirely, even though all three might hit the same API and the same object.
We checked which of the named AI support agent tools enforce that line as something closer to a permission than a setting, against each vendor's current docs and third-party coverage in November 2026. Drafting is table stakes everywhere. The gap between "can draft" and "can be stopped from sending" is wider than the marketing suggests.
How we scored these
We looked at five named platforms: Intercom Fin, Zendesk's AI agent features, Freshdesk Freddy AI, Help Scout AI, and Decagon. For each one we checked current pricing, and whether draft-only behavior and send/execute behavior are separated by something closer to an access control (a boundary the agent cannot cross on its own, independent of which interface is driving it) or by a UI convention (a toggle, a recommended rollout practice, a step you have to remember to add).
We're citing what each vendor's own documentation and credible third-party analysis state as of November 2026. Where a source didn't document a control, we say so rather than assuming it doesn't exist. This isn't a claim about api access control in the abstract. It's specifically about whether sending, refunding or closing can be gated the same way reading and drafting already are.
1. Intercom Fin
Fin bills $0.99 per resolution outcome (a resolution, a procedure handoff, or a disqualification), and $9.99 per qualification. A hypothetical 10,000 conversations at 50% resolution works out to 5,000 billed outcomes, or $4,950 a month before seats. Copilot, the agent-facing drafting tool, is a separate $35 per agent per month add-on.
Fin's own customer-facing replies aren't described as going through an internal-note-only review step by default. Copilot drafts for human agents to review and send, but that's a different product surface from Fin itself answering a customer directly.
Intercom does offer a real control worth naming: a "Loop in teammate" step you can add inside a Fin Procedure. When Fin reaches that step, it waits for a teammate to respond (or for a timeout) before continuing. That's a genuine pause, and Intercom's own docs describe refunds and account closures as example use cases for it.
What it isn't, based on Intercom's documentation, is an access control that lives outside the procedure. The gate is positional: it only exists where you placed it in that specific procedure's flow, and it isn't described as a boundary scoped to an action type (send, refund, close) independent of how that procedure was built. A procedure built without that step, or a call routed through a different client entirely, has no equivalent to fall back on.
2. Zendesk AI agents
Zendesk's AI agents bill $1.50 per resolution on a committed pack, or $2.00 pay-as-you-go, the same rate across Support Team, Suite Team and Suite Professional. Since May 2026, only Verified Resolutions bill; assisted escalations and contained resolutions are free. Since January 2026, overage auto-bills with no cap, grace period, or prior-month warning, and the only way to stop it is pausing the AI agent entirely.
Zendesk's own documentation describes agents being authorized to run discrete actions, such as updating a ticket field, looking up an order, or triggering a refund flow, without describing an approval threshold attached to those actions. Dialogues give more deterministic control than generative procedures, but that's a design choice you make, not a boundary the platform enforces on your behalf.
The drafting story is split across two products that don't stay cleanly separated. Copilot drafts replies for human agents, but with Auto Assist turned on, it goes on to take follow-up actions in Zendesk and connected systems on its own. The standard rollout advice (run the agent as a suggestion engine for two to four weeks before granting autonomy) is sound, but it's advice, not a setting Zendesk enforces.
3. Freshdesk Freddy AI
Freshdesk doesn't price Freddy AI per resolution. It bills per session: 500 complimentary sessions on paid plans, then $49 per 100 sessions ($0.49 each). A chat session runs 24 hours from its start; an email session runs 72 hours from the customer's first message. Copilot, the agent-facing assistant, is a $29-per-agent-per-month add-on on Pro and Enterprise.
Freddy's AI Agent can process refunds and run connector "skills," like updating a Shopify return, and Freshdesk's documentation doesn't describe a confirmation step attached to those actions. A handoff to a human bills the same as a fully automated session, which tells you something about how Freshdesk thinks about the cost of escalation, but nothing about how it thinks about the cost of an action going wrong.
Copilot drafts replies, adjusts tone and summarizes threads, which is useful, but whether a drafted suggestion requires explicit agent sign-off before it goes out, or can be configured to skip that step, isn't documented in what we reviewed. We'd call this a gap worth asking Freshdesk about directly rather than assuming either way.
4. Help Scout AI
Help Scout charges $0.75 per billable resolution for AI Answers, with a three-month trial for new accounts. A session with multiple follow-up questions counts once, as long as the customer doesn't keep asking after the final answer.
Architecturally, Help Scout splits the two jobs more cleanly than the others on this list. AI Answers operates inside the self-serve Beacon widget, answering customers directly. Inbox Assistant is a separate tool that helps a human agent draft, summarize and edit a reply inside the Inbox, where a person still has to hit send. That split is a genuinely useful default: it keeps the mailbox-reply path, the one with real consequences if it goes to the wrong customer, in human hands by design.
That said, it's an architectural default, not a declared permission boundary. Third-party integrations that connect to Help Scout's reply API note that sending "cannot be recalled" once triggered, and that there's no built-in draft mode on the send action itself; any draft-only safeguard around it is something you configure in whatever's calling the API, not something Help Scout's send endpoint enforces on its own.
5. Decagon
Decagon doesn't publish list pricing. It sells on a per-conversation or per-resolution ("outcome-based") meter, quoted through sales, with unspecified volume discounts, which fits its positioning as an enterprise-grade, fully autonomous support agent rather than a self-serve add-on.
Decagon's own content describes "Agent Operating Procedures" and escalation policies: teams define rules for identity verification, refund thresholds, and when a conversation routes to a human. Some topics (legal disputes, fraud claims, regulated financial decisions, medical questions) route to a human unconditionally, by design.
What we couldn't verify from Decagon's own public documentation is whether these procedures are enforced as a platform-level control the agent cannot act outside of, or whether they're configuration and governance practice, similar to Intercom's and Zendesk's approach. Decagon's own glossary content describes them in governance language ("documented as formal configuration, version-controlled, and reviewed regularly"), which reads like a recommended practice rather than an enforced permission system. We'd confirm that distinction directly with Decagon before relying on it for an irreversible action.
Side-by-side comparison
| Tool | Draft-only mode? | Hard permission boundary or UI convention? | Pricing (as of Nov 2026) |
|---|---|---|---|
| Intercom Fin | Copilot drafts for agents; Fin itself can reply directly | Configured step in a Procedure (positional, not scoped to action type) | $0.99/resolution outcome, $9.99/qualification, Copilot $35/agent/mo |
| Zendesk AI agents | Copilot drafts; Auto Assist can act automatically | Authorized actions with no documented approval threshold | $1.50/resolution (committed) or $2.00 (PAYG) |
| Freshdesk Freddy AI | Copilot drafts for agents | No documented confirmation step on refund/action skills | Per session: 500 free, then $49/100 sessions |
| Help Scout AI | Inbox Assistant drafts for a human to send | Clean architectural split (Beacon vs. Inbox), not a declared access control | $0.75 per billable resolution |
| Decagon | Agent Operating Procedures define escalation/thresholds | Governance-framed configuration; platform enforcement unconfirmed | Sales-quoted, per-conversation or per-resolution |
Where Swytchcode fits (and where it doesn't)
Swytchcode is not a helpdesk, not a support platform, and not a replacement for Fin, Zendesk AI, Freddy AI, Help Scout AI or Decagon. All five are good at the part they're built for: understanding a ticket, finding the right answer, producing a reply. That's genuinely hard, and we're not trying to re-solve it.
What we noticed across every platform above is that the draft-vs-send boundary, when it exists at all, lives inside that platform's own configuration: a procedure step, a rollout recommendation, a split between a self-serve widget and a human inbox. Useful, but none of it is independent of the client making the call. If a different integration or a differently configured client calls the same underlying send, refund or close method, that boundary doesn't travel with it, because it was never a property of the API call. It was a property of how one particular flow was built.
That's the layer Swytchcode sits in: between whatever agent is drafting and sending, and the APIs it actually calls to do it. Drafting a reply and sending one might hit the same endpoint, but they aren't the same action in consequence, and Swytchcode's allowlist and policy evaluation treat them that way no matter which platform or model is driving the agent.
In practice, that looks like allowing draft-creation methods freely, since a draft has no blast radius, and requiring approval on anything that leaves the building:
{
"id": "support-refund-approval",
"target": ["stripe.refund.create"],
"when": { "field": "amount", "operator": ">", "value": 10000 },
"action": { "type": "REQUIRES_APPROVAL", "message": "Refunds over $100.00 need approval" },
"approval_timeout": "2h"
}{
"id": "support-send-approval",
"target": ["intercom.conversation.reply.create", "intercom.conversation.close.create"],
"when": { "field": "status", "operator": "=", "value": "customer_facing" },
"action": { "type": "REQUIRES_APPROVAL", "message": "Customer-facing replies and account closures need approval" },
"approval_timeout": "1h"
}The method names above are illustrative; run swy discover against whichever support platform you've connected to get the real canonical IDs before writing a policy against them. The point is the shape: one method a support agent can call with no friction, a different set of methods that hold for a human no matter which drafting tool or model produced the call. That holds whether the drafting side is Fin, Zendesk's agent, Freddy, Help Scout's Inbox Assistant, Decagon, or something built in-house, because the policy lives at the point of execution, not inside any one vendor's procedure builder.
Unapproved calls fail closed. If nobody approves a held refund or send within the timeout, it never runs, the same way it would if the permission had never been granted. That's a meaningfully different guarantee than a recommended rollout practice, because it doesn't depend on anyone remembering to add a review step to a specific procedure.
This also means the choice of support platform and the decision about what it's allowed to execute stop being the same decision. You can run Fin or Decagon for the parts they're good at, and still have a policy that holds a refund over a threshold, or any send to a customer, for a person, enforced at the API boundary rather than inside whichever tool drafted the message. For more on this split between drafting tools and execution policy, see our notes on connecting Intercom's API to a support agent and doing the same with Freshdesk. The general case for treating access control as a property of the call rather than the client is in api access control for AI agents, and the policy mechanics are documented at docs.swytchcode.com/policies/overview.
FAQ
Why does draft-vs-send matter more than read-vs-write for a support agent?
Read-vs-write is the wrong axis because plenty of writes are harmless and plenty of reads aren't. Writing a draft reply is a write, but nothing happens to the customer until someone sends it. Sending, refunding or closing an account are also writes, but they're irreversible the moment they execute. The risk comes from whether an action is reversible and customer-facing, not from whether it modifies a record. A policy built around read-vs-write will either block useful drafting or wave through a refund, because it answers the wrong question.
Can I just rely on my support platform's own approval or draft settings?
You can, and for a lot of teams that's a reasonable start, especially with Intercom's "Loop in teammate" step or Help Scout's split between Beacon and Inbox. The limit is scope: those controls live inside that platform's own procedure or UI, built for one specific flow. They don't automatically extend to a different integration, a different client, or a prompt change that routes around the procedure that had the approval step in it. If the consequence of getting it wrong is a duplicate refund or a wrongly closed account, it's worth having a boundary that holds independent of which client is calling the API.
Does adding Swytchcode mean giving up our current support AI?
No. Swytchcode doesn't draft replies, search your knowledge base, or summarize tickets, so it isn't competing with Fin, Zendesk AI, Freddy AI, Help Scout AI or Decagon on any of that. It sits in front of the APIs those tools (or your own agent) call to actually send a reply, issue a refund, or close an account, and it applies the same allowlist and policy checks no matter which of them is doing the drafting. You keep the support platform you've already chosen and add a layer that decides, at the point of execution, whether a given call is allowed to run on its own or needs a person first.
More content
Best Agent Frameworks for Production Use
We rank six agent frameworks used in production as of November 2026 and show why none of them handle API permissions, retries, or audit by design.
Best API Catalogs and Registries for AI Agents
We rank Backstage, Postman, Stoplight, SwaggerHub and the agent-native catalogs on one question: does the schema stay current at call time?
Best API Platforms for Fintech AI Agents
We rank Stripe, Modern Treasury, Plaid, Adyen, Increase, and Column on the two things that actually matter for fintech agents.
