AI Support Agents 2026: Fin, Decagon, Sierra, Ada, Cresta

Every CX leader I’ve talked to this quarter has the same problem: the board asked them to “deploy AI in support” by year-end, the budget showed up, and now they have to actually pick a vendor. Eighteen months ago this was a maybe. Now it’s a deadline.

The good news is the market finally settled. The two dozen “AI for support” startups from 2024 collapsed into five real options that show up in nearly every enterprise RFP I see: Intercom Fin, Decagon, Sierra, Ada, and Cresta. Everyone else is either an honorable mention or a feature inside one of those five.

The bad news is that picking between them is harder than it looks, because they’re not actually competing for the same job.

The five that actually matter

Here’s the rough shape of the market in mid-2026, before we get into the details:

Intercom Fin — the helpdesk-native default. $0.99 per resolution. The path of least resistance if you’re already on Intercom or starting from scratch.
Decagon — mid-market through enterprise, technical-team friendly, customers like Notion, Eventbrite, Duolingo, Hertz, Bilt. $131M Series C in mid-2025 at a $1.5B valuation.
Sierra — Bret Taylor’s enterprise-only platform. Last valued at $10B. $150M ARR in January 2026. White-glove implementations for ADT, Cigna, SiriusXM, SoFi, Discord, Rivian.
Ada — the no-code-friendly mid-market option that non-technical CX teams can actually run themselves.
Cresta — the voice-first pick, with deep contact-center heritage and real-time agent-assist alongside autonomous voice.

Companies you’ll see in honorable-mentions territory: Forethought (helpdesk-agnostic), Kustomer (Meta-owned), Salesforce Agentforce Service, Zendesk’s own AI Agents, Maven AGI for technical/dev-tools support, Lorikeet for fintech, Lyro by Tidio for SMB, Espressive for employee IT. They’re all real products. None of them are showing up in the deals I see at the top end.

Why deflection rates jumped — and what “deflection” actually means

Around mid-2024, vendor-quoted deflection rates lived in the 30–40% range. By Q2 2026, the strongest vendors are claiming 65–75%. That’s not just better models. It’s the combination of bigger context windows, real tool-calling, and the fact that vendors finally stopped shipping “answer-from-the-knowledge-base” bots and started shipping agents that can actually update an order or refund a charge.

Watch the words, though. Three different vendors will tell you they hit “70% resolution” and mean three different things:

Deflection — the customer didn’t escalate. (Could mean they gave up.)
Contained — the conversation ended in the agent without a human touching it.
End-to-end resolved — the agent took an action that resolved the actual underlying issue.

The third is the only one that matters, and the cleanest public benchmark we have is Fin’s third-party-audited number around 67% on real customer transcripts. Klarna’s much-quoted “AI did the work of 700 agents” line from Sebastian Siemiatkowski is real, but it’s a single internal case study, not a benchmark — and Klarna walked back parts of it in early 2025.

When a vendor quotes you a number, ask: which of the three? On what dataset? Cherry-picked customer or representative? You’ll filter out half the marketing in one question.

Pricing models, and why they don’t compare

Each of the five charges differently, and the differences are bigger than the dollar amounts:

Per-resolution — Intercom Fin at $0.99 per successful resolution, with a 50/month minimum. You also need at least one paid Intercom seat (Essential starts around $29/month) and the AI Copilot for human agents is an extra $35/seat/month. Predictable when volumes are stable. Painful when a viral incident triples your tickets overnight.

Per-conversation — Ada and several mid-market players. Cheaper per unit but charges fire whether the conversation resolved or not, which means you pay for failures.

Per-agent-seat plus AI add-on — the Zendesk and Salesforce Agentforce model. Looks familiar to procurement. Hides the actual unit economics.

Annual contract value, no public pricing — Sierra. Floor sits at the $150k+ ACV range; large enterprise deals run substantially higher. You’re paying for white-glove buildout as much as the software.

Hybrid — Decagon and Cresta tend to land on custom enterprise contracts that mix platform fees with usage. Decagon deals I’ve seen anchor in the $50k–$300k+ ARR range depending on volume.

The trap with per-resolution: it looks cheap at $0.99 until you do the math at 500k tickets/year. That’s $495k/year if every ticket resolves, plus seats, plus Copilot. Suddenly the “expensive” enterprise contract isn’t the expensive one.

The separate-helpdesk tax nobody puts in the spreadsheet

Here’s the thing buyers miss most often. Decagon, Sierra, Ada, and Cresta all sit on top of an existing helpdesk. They don’t replace Zendesk, Salesforce Service Cloud, or Intercom. They plug into them.

That means your real cost is helpdesk seats + AI vendor + integration work + ongoing maintenance of the connection. If you’re on Zendesk Suite at $115/agent/month with 200 agents, you’re already at $276k/year before the AI line item shows up. Adding Decagon or Sierra on top doesn’t reduce that — at least not in year one, while you’re still measuring whether deflection is real.

Fin is the exception, sort of. It runs natively inside Intercom, so if you’re already on Intercom you don’t pay an integration tax. If you’re not, switching helpdesks to chase Fin is a much bigger project than just buying an AI agent.

This is the single biggest reason I tell people not to optimize on “AI vendor cost” alone. The AI vendor is rarely the largest line in the total cost of ownership for the first 24 months.

Channel coverage: voice is still hard

Most vendors will sell you a glossy “omnichannel” pitch. The reality, channel by channel:

Chat and email — every vendor on the shortlist does this well in 2026. Resolution rates above 60% are normal on knowledge-base-answerable questions. Where vendors differ is on the long tail and on action-taking (refunds, address changes, subscription cancellations).

Voice — meaningfully harder. Resolution rates run 15–25 points lower than chat at every vendor I’ve measured. Cresta and Sierra are the strongest here because they grew up in voice contact centers; the others mostly bolted voice onto a chat-first product. If voice is your primary channel, you should not be picking Fin or Ada as your primary agent.

SMS and WhatsApp — fine for transactional flows (delivery updates, appointment reminders). The free-form-question quality lags chat by a noticeable margin because customers write more sloppily in SMS.

Social DM — supported, mostly through Meta integrations, mostly an afterthought. If your support volume on Instagram DMs is meaningful, audit the actual integration quality before signing.

The “we do every channel” pitch is technically true at every vendor and practically misleading at most. Pick based on where 80% of your volume actually lives, not where it might live in three years.

The compliance moment is here

The August 2026 EU AI Act high-risk-system deadline is going to force a real audit cycle, especially for anyone in finance, healthcare, or regulated retail. What I’d actually check today:

SOC 2 Type II — table stakes; every vendor on the shortlist has it
HIPAA — Sierra and Decagon ship BAAs in their enterprise tier; Fin has it on higher plans; Ada and Cresta vary
PCI-DSS — needed if you take payment data inside the agent flow; verify scope, not just the badge
EU data residency — Sierra has it after the Fragment acquisition added a Paris team; Fin offers EU regions; Decagon has it in their enterprise tier; the others are improving
Audit trail of agent actions — this is the one buyers underestimate. The EU AI Act expects you to be able to show why an agent took an action against a real user. Some vendors have proper action logs; others have chat logs and a prayer.

I’d push every vendor to walk you through their action-audit story specifically. “We log conversations” is not the same answer as “we log every tool call with inputs, outputs, and the reasoning trace, and we keep it for seven years.”

Will it actually replace headcount?

This is the question every CFO asks and every salesperson dodges. The honest answer in 2026: maybe, but slower than the deck suggests.

What I see working: 30–50% reduction in tier-1 ticket volume within six months at companies that did the knowledge-base hygiene work first. That’s real money. It usually shows up as redeploying agents to higher-value work (retention, sales-assist, complex escalations) rather than firings, partly because those agents are now your training data and quality reviewers for the AI.

What I see breaking: companies that promote the agent past 50% automation in the first eight weeks. NPS craters. The bot starts deflecting questions it shouldn’t. Customer trust takes a year to rebuild. Klarna’s mid-2024 backpedal — where they quietly started rehiring human agents after the original “AI did the work of 700 people” story — is the canonical version of this mistake.

The winners are running the AI in shadow mode for four to six weeks, building golden conversation datasets from human-resolved tickets, and only promoting categories of questions to full automation as the eval scores hold up. Slower. Less viral. Works.

A decision tree that’s actually useful

I keep getting asked “which one should I pick” as if there’s one answer. There isn’t. Here’s the closest I can get:

Pre-PMF startup, <10k tickets/year — start with Fin if you’re on Intercom, or roll your own with Claude Sonnet plus a couple of MCP tools. You don’t need a vendor yet.
Series A–B SaaS, 10k–100k tickets/year, on Zendesk — Decagon if you have a CX engineering function, Ada if you don’t.
Mid-market, 100k–1M tickets/year, growing fast — Decagon or Fin. Decagon if your support involves a lot of action-taking; Fin if your volume is mostly knowledge-base-answerable.
Enterprise, regulated industry, 1M+ tickets — Sierra is the obvious pick if you have the budget and the patience for a multi-month rollout. Salesforce Agentforce Service if you’re already deep in Service Cloud and the org won’t tolerate a second vendor.
Voice-heavy operation — Cresta or Sierra. Don’t pick a chat-first tool and hope it works on phone calls.
Employee IT support, not external customer support — Espressive or Moveworks; the external-CX vendors are not optimized for your use case.

A pattern that works really well in practice: Fin for the front line on chat/email, Salesforce or Zendesk staying as the system of record, and a separate voice tool (often Cresta) for the contact center. Two-tool stacks are normal in 2026. The “one platform for everything” pitch usually means “good at one thing, mediocre at the others.”

The 12-week rollout that actually ships

If you’re starting fresh, here’s the rough cadence I see at companies that don’t blow up their NPS:

Weeks 1–2 — knowledge base hygiene sprint. Half your articles are stale. Fix them before the bot reads them.
Weeks 3–4 — golden-conversation dataset. Pull 500 representative tickets, including hard ones. This becomes your eval set.
Weeks 5–6 — shadow mode. The agent runs alongside humans, suggests answers, never sends. You measure quality without risk.
Weeks 7–8 — narrow automation. One or two ticket categories with clean knowledge and low risk. Refunds under $X, password resets, order status. Measure deflection AND CSAT, not just deflection.
Weeks 9–10 — expand by category. Don’t expand by percentage of total volume. Expand by question type, with an eval gate at each step.
Weeks 11–12 — escalation pathways and handoff QA. The handoff from bot to human is where most CSAT damage happens. Make sure context transfers cleanly and humans don’t have to ask for the customer’s name again.

Anyone selling you “live in 2 weeks” is selling you something that will need to be rebuilt in 6 months.

What I’d watch next

Three things I’d keep an eye on through Q3 2026: whether Fin v3’s multi-channel parity actually holds up in production at large customers; whether Sierra’s voice-first pivot starts pulling deals away from Cresta in voice-heavy verticals; and whether Salesforce Agentforce Service finally ships the autonomous-action quality that justifies its CRM-incumbency tax. The vendor lineup will probably stay these five through the end of the year. The pecking order inside it is going to shift a lot.

If you’re picking right now and want a single concrete next step: pull last quarter’s tickets, classify them by channel and by whether the agent would have needed to take an action versus just answer a question. The shape of that pie chart picks two of these five vendors for you, and you can run a real bake-off between those two instead of evaluating all five.

Sources: