AI agents for customer support 2026: Sierra, Decagon, Fin, Ada

If you want a published price and a fast start, Intercom Fin is the easiest of the four to try: $0.99 per outcome, and it works on top of your current helpdesk. If you are a large enterprise with complex workflows across voice and chat, Sierra and Decagon are the heavyweight builds, sold through sales. Ada sits in between, with broad channel coverage and a long track record in automated resolution.

The bigger decision is not the vendor. It is how you define a resolution, what you measure next to it, and how slowly you roll out.

Key takeaways

• Fin is the only one of the four with public pricing: $0.99 per outcome (as of October 2026), with a 50 outcome monthly minimum on non-Intercom helpdesks.
• Sierra uses outcome-based pricing and says that if a conversation is unresolved, in most cases there is no charge. Rates are not published.
• Decagon and Ada price through demos. Expect a negotiated contract, not a self-serve plan.
• Resolution rate is the headline metric, but pair it with CSAT and recontact rate or you will pay for "resolutions" that annoyed customers.
• Roll out one intent and one channel at a time, with human handoff working from day one.

How do Sierra, Decagon, Fin and Ada compare?

All details are from vendor sites as of October 2026.

How does per-resolution pricing work?

Per-resolution pricing means you pay when the AI agent closes a customer's issue, not per seat and not per message. The appeal is simple: if the agent fails, you should not pay for it.

The catch is the definition. Read it before you sign.

Intercom is the clearest example because it publishes its rules. Fin charges $0.99 per outcome. A resolution counts when "no further help is requested after Fin's last answer." That covers two cases: the customer confirms it helped ("Ok thanks"), or the customer leaves without asking for more. Intercom calls the second an assumed resolution.

What Fin does not bill:

• Conversations escalated to a human under default behavior or workspace rules.
• A customer who asks for a human at any point, or a Procedure that fails to finish.
• Greetings with no real answer.
• A clarifying question the customer never answers, which closes as abandoned.

Fin also bills a few other outcome types, including procedure handoffs and routing. Qualification of sales leads costs $9.99. Used standalone on another helpdesk, there is a minimum of 50 outcomes per month.

Sierra frames it more broadly. It defines outcomes as "tangible business impacts, such as a resolved support conversation, a saved cancellation, an upsell, a cross-sell." That means your contract can tie price to results other than a closed ticket, which is useful if retention matters more to you than deflection.

What to look for in any outcome contract:

1. Assumed vs confirmed. Assumed resolutions are where cost and reality drift apart. A customer who gives up looks the same as one who got help. Ask how the vendor tells them apart and whether you can audit a sample.
2. Recontact window. If the same customer comes back about the same issue within a day or two, was the first one really resolved? Negotiate a clawback or exclusion.
3. Minimums and overages. Know the floor and what happens in a seasonal spike.
4. Your data. Make sure you can export conversation-level outcome labels so your own team can check the bill.

What should support leaders measure?

Measure resolution rate, but never alone.

• Resolution rate. The share of conversations the agent closes without a human. Vendors publish strong numbers. Ada, for example, cites an 84% automated resolution rate on its site, and Decagon lists customer results like Chime at 70% chat and voice resolution. Treat these as what is possible, not what you will get in month one.
• CSAT on AI conversations. Survey AI-handled conversations separately from human ones. If resolution goes up and CSAT drops, the agent is closing tickets customers did not want closed.
• Recontact rate. Same customer, same issue, within 48 to 72 hours. This is your best check on assumed resolutions.
• Escalation quality. When the agent hands off, does the human get full context, or does the customer repeat everything? Sample transcripts weekly.
• Cost per resolution. Compare the AI price per outcome against your fully loaded human cost per ticket for the same intent, not your average ticket.
• Containment by intent. A 60% overall rate can hide 95% on order status and 10% on billing disputes. Break it down.

How do you roll out an AI support agent?

A careful rollout beats a big launch. Here is a sequence that works for most teams.

1. Clean the knowledge base first. Every vendor here builds from your docs, SOPs and policies. Stale articles produce confident wrong answers. Fix the top 20 articles behind your highest-volume intents.
2. Pick two or three intents. Start with high volume, low risk, clear policy: order status, password resets, plan changes. Leave refunds and disputes for later.
3. Wire up actions, not just answers. The value is in doing things: checking an order, changing an address. Decagon's AOPs, Sierra's Agent Studio, Ada's Playbooks and Fin's Procedures all exist for this. Connect the systems before launch.
4. Test before customers see it. Decagon offers simulations at scale, and Ada's Performance Center runs tests across languages. Whatever the tool, run real past conversations through it and grade the answers.
5. Make handoff work on day one. Customers should reach a human easily, with context passed over. A blocked handoff is the fastest way to tank CSAT.
6. Launch on one channel at a small share of traffic. Chat is usually easiest. Hold voice until chat is stable.
7. Review weekly. Read failed and assumed-resolved transcripts, update content and procedures, then widen intents and traffic.

Which one should you pick?

• Pick Intercom Fin if you want public pricing, a quick pilot and no helpdesk migration. It runs standalone on Salesforce, HubSpot, Freshworks, Front, Gorgias, Zoho and others, with no seats needed.
• Pick Sierra if you are a large consumer brand that wants one agent across chat, SMS, WhatsApp, email and voice, and you want price tied to outcomes like saved cancellations. Its homepage lists customers such as Rocket Mortgage, Gap Inc., SiriusXM, Uber, Vanguard and Wayfair.
• Pick Decagon if your team wants to write and test agent behavior in natural language, with strong QA tooling like simulations, A/B tests and always-on QA. It lists customers including American Airlines, Duolingo, Delta and Square.
• Pick Ada if channel breadth and enterprise helpdesk integrations matter most. It covers voice, email, chat, WhatsApp, SMS, Instagram and in-app, and connects to Salesforce, Zendesk, ServiceNow and Genesys.

If you already run Zendesk or Salesforce, also look at their native agents before adding a separate vendor. Zendesk includes AI agents in every Suite and Support plan, priced on successful outcomes with an included allowance, and Salesforce Agentforce offers service agents for cases, orders and troubleshooting. Fewer systems means fewer handoff gaps.

Bottom line

All four can resolve a real share of your volume. The winner for your team is the one whose resolution definition you trust, whose testing tools your team will actually use, and whose handoff keeps customers happy. Start small, measure CSAT and recontact next to resolution rate, and only widen scope when the numbers hold.