AI Customer Support Buyer's Checklist: Task Completion vs. Ticket Deflection

Answering a question and finishing a task are not the same product

Most AI customer support tools now claim some flavor of "action-taking," and the phrase has stopped meaning much on its own. A tool that fires a Slack notification when a ticket comes in is technically taking an action. So is a tool that processes a refund, updates a Shopify order, and confirms the change back to the customer without anyone on your team touching it. Those are different products solving different problems, and the gap between them is where a lot of AI support budgets get spent on the wrong tier.

This guide is narrower than a general AI support roundup on purpose. It looks at three vendors, XBert, Decagon, and Sierra, specifically on the question of whether the agent completes a task inside your systems or hands a summary to a human who then does the actual work. Broader AI deflection tools by price and use case, Intercom Fin, Zendesk AI, Tidio Lyro, Ada, Forethought, and others, are covered elsewhere on this site; this piece goes deeper on the task-completion question specifically.

One thing worth saying up front: these three vendors are not natural peers. Decagon and Sierra are both built around resolving an existing support interaction end to end. XBert is built around a different job, being the first point of contact on a call or a text, and it turns out that distinction matters more than any feature checklist. The rest of this guide explains why, and what to verify before you sign with any of them.

Five actions worth testing before you believe a demo

A sales demo is built to show the agent succeeding. Before you sign anything, ask to see these five specific actions attempted on data that resembles your own, not the vendor's canned example:

Refund or credit processing: ask the agent to process a partial refund on a specific order, then check whether the correct amount lands in your payment processor, not just a chat message saying "refund issued." The amount should come from your billing system's calculation, never from the model doing arithmetic on the fly.

Subscription or plan change: downgrade or cancel a test account mid-cycle and confirm the proration, or the exit fee if your contracts have one, matches what your billing system would actually charge.

Identity verification before a balance-affecting action: confirm the agent asks for something beyond a name and email before it touches money, and that a failed verification actually blocks the action instead of just logging a warning.

A write-back to your CRM or ticketing system: have the agent update a field, a case status, or a customer note, then open the record yourself in Salesforce, HubSpot, Zendesk, or whatever you run, and confirm it is actually there, not just referenced in the AI's own conversation log.

A confirmed loop back to the customer: the agent should state what it did and give the customer a way to verify it, an order number, a confirmation email, a case number, not just "you're all set."

Vendors that can do all five on your own data in a live session are in a different category from vendors that can describe all five in a sales deck. That gap is the entire point of this checklist.

Decagon: agent operating procedures as the action layer

Decagon's approach to action-taking is built around what it calls Agent Operating Procedures, plain-language workflows that a support team writes and maintains without engineering involvement, while developers separately control the underlying integrations and guardrails [Decagon, "From SOPs to Agent Operating Procedures," checked 2026-09-23]. Decagon describes this as translating a business's existing standard operating procedures into a format the agent can follow step by step, with pre-built templates for common cases like refund processing and account verification.

The actions themselves run through integrations with Stripe, Shopify, and Salesforce: Decagon states its agents can process refunds, update orders, verify identity, and create tickets without escalating to a human, and that it deliberately executes the sensitive validation steps, refund math, identity checks, in code rather than leaving the calculation to the model [Decagon, "What AI customer support agents actually do," checked 2026-09-23]. For Salesforce specifically, Decagon can access the full Customer 360 record, create or update cases with the full conversation attached, and modify opportunity stages. It has no native helpdesk of its own, so a human-facing system like Zendesk or Salesforce Service Cloud stays in place alongside it for the cases that still need a person; anything outside its pre-built connectors (a custom ERP, an older ticketing system) needs additional engineering work to wire up [Fin AI's Decagon pricing guide, checked 2026-09-23].

Decagon's customer list leans toward companies with high-volume, repeatable support flows: Notion, Substack, Eventbrite, Duolingo, Webflow, Rippling, Bilt, Curology, and ClassPass are named case studies, with ClassPass reporting a 10x increase in deflection alongside 24/7 chat coverage [Decagon case studies, checked 2026-09-23]. The company has scaled fast: a $250 million Series D in January 2026 valued it at $4.5 billion, up from a $1.5 billion valuation just seven months earlier, and Sacra estimates Decagon's annualized revenue crossed $100 million by July 2026, up from $44 million at the end of 2025 [Bloomberg, January 28, 2026; Sacra Decagon profile, checked 2026-09-23]. On G2, Decagon holds a 4.9 out of 5 rating, though from a small base of 18 reviews, with reviewers citing implementation ease and hands-on vendor support most often [G2, checked 2026-09-23].

Sierra: co-built agents with PCI-compliant payment actions

Sierra takes a more hands-on implementation approach than Decagon's self-authored procedures: the company's team works directly with a customer to build the initial branded agent rather than handing over a self-serve configuration dashboard. Agents handle conversations across voice, chat, email, and WhatsApp in 59 languages, and the action layer covers processing a return, changing a subscription, and verifying identity inside the conversation itself [Sacra Sierra profile, checked 2026-09-23].

The clearest recent expansion of what Sierra's agents can do landed in April 2026, when the company added PCI-compliant payment handling, letting an agent process a refund, update a subscription, or check an account balance without routing the interaction to a human for the financial step [LeadrPro coverage of Sierra's funding round, checked 2026-09-23]. That is a meaningful gap to close for a company evaluating whether an AI agent can actually touch billing rather than just discuss it.

Sierra's customer roster skews toward large, brand-sensitive consumer companies: named accounts include WeightWatchers, SiriusXM, Sonos, ADT, Chime, Cigna, Nordstrom, Nubank, Ramp, and Rivian, and the company states it serves more than 40% of the Fortune 50. Published case studies report automation rates of 65%+ at Minted, 74% at Casper, 70%+ at Chime, and 90% at Ramp [Sacra Sierra profile; eesel AI Sierra reviews roundup, checked 2026-09-23]. The company, founded in 2023 by Bret Taylor and Clay Bavor, raised a $950 million round in May 2026 led by Tiger Global and GV at a post-money valuation above $15 billion, with reported ARR climbing from roughly $150 million in February 2026 to around $200 million by mid-2026 [Sacra; LeadrPro, checked 2026-09-23]. G2 shows a 4.1 out of 5 rating from 17 reviews, again a small sample; reviewers describe a clean interface once teams are past the initial ramp-up, with complaints centered on early-stage complexity and the agent losing context in longer conversations [G2 Sierra reviews, checked 2026-09-23].

XBert, Decagon, and Sierra compared

Read the pricing column carefully. XBert's number comes straight from Nextiva's own pricing page. Decagon and Sierra publish nothing; every figure attributed to them above is a third-party estimate built from public reporting and pricing-analysis sites, not a vendor-confirmed number, and estimates for the same vendor vary by tens of thousands of dollars depending on the source. Treat the table as a starting point for a real quote, not a number you can budget against directly.

Where XBert fits, and where it does not

XBert is Nextiva's own product, and it is the one entry in this comparison built around a different core job than the other two. Nextiva markets and documents it as an AI receptionist and voice/chat assistant: it answers inbound calls, texts, and web chats around the clock, qualifies leads, books appointments, and hands off to a human when a conversation needs one [Nextiva XBert product page, checked 2026-09-23]. Nextiva's own material describes it connecting to CRM, scheduling, and ticketing systems to log calls, update records, and trigger follow-up actions, and checking calendars, sending confirmations, and updating CRM records when a caller makes a request [Nextiva XBert product page; Nextiva customer support tools blog, checked 2026-09-23].

That is a real action-taking capability, and it is worth being specific about what it is not. Nothing in Nextiva's public product pages, help documentation, or case studies describes XBert resolving an existing support ticket end to end with an account-level financial write, issuing a refund, processing a cancellation, adjusting a subscription, the way Decagon's Stripe/Shopify/Salesforce actions or Sierra's PCI-compliant payment handling are explicitly built and marketed to do. Calling XBert a smaller Decagon or a cheaper Sierra would overstate what the sourced material supports. It is a different tool for a different moment in the customer interaction: the first point of contact, mostly by phone or text, logging what happened and triggering the next step, rather than the system that closes out a complex billing dispute already sitting in a queue.

Where that distinction cuts in XBert's favor is accessibility. At $99 a month including 100 conversations and $0.99 each after that, with no long-term contract and a 14-day free trial, XBert is reachable by a small business the same week they sign up [Nextiva XBert product page, checked 2026-09-23]. Decagon and Sierra both require a sales process, a multi-week implementation, and, per third-party estimates, a budget in the tens to low hundreds of thousands of dollars a year before you process a single refund. If your bottleneck is calls and texts going unanswered and information not making it into your CRM, XBert addresses that directly and immediately. If your bottleneck is a support queue full of tickets that need an account-level action to actually close, XBert is not the tool built for that job, and Decagon or Sierra are the two worth evaluating instead. If calls and texts going unanswered is the actual problem, our AI answering service buyer's checklist goes deeper on evaluating XBert against three other receptionist-focused tools built for that specific job.

How the three price this, and what the number does not tell you

All three vendors charge for outcomes rather than seats, but the unit each one counts is different, and that difference is where budgeting mistakes happen. XBert bills per conversation, defined specifically as a call over 30 seconds or a chat/SMS thread with three or more exchanges, at a published $0.99 rate after the included 100 per month. That definition is public and you can model your bill before you sign.

Decagon and Sierra both work on negotiated annual contracts, priced per conversation or per resolution depending on the deal, and neither publishes a rate card. Get each vendor's exact definition of a billable resolution or conversation in writing before comparing a quote to anything else, since a "resolution" that counts a customer who simply stops replying is a very different number from one that only counts a verified, completed action. Decagon has reportedly moved most of its customer base toward per-conversation (usage-based) pricing over per-resolution (outcome-based) pricing, according to third-party pricing analyses, for the same reason: it is a more predictable number to forecast against.

None of these prices include the cost of a wrong action. A confidently incorrect deflection is an annoyance; a confidently incorrect refund, subscription change, or identity verification is money out the door or a compliance problem. Before signing with any vendor that can write to your billing system, ask specifically how the vendor bills you if the agent takes an action incorrectly, whether that is a reversible event on their side or entirely your team's cleanup, and whether a paid pilot period exists where you can measure the error rate on your own tickets before the full contract kicks in.

Guardrails before you let any agent write to a live system

The moment an AI agent can change a customer's account rather than just talk about it, the product category changes from "chatbot" to "software with write access to your business," and it should be governed that way regardless of which of these three vendors you pick.

Put a confirmation gate in front of every irreversible action, refunds, cancellations, plan downgrades, before go-live, not after the first incident. Set an amount ceiling above which the action routes to a human regardless of how confident the agent is. Keep an audit log of what the agent did, on which account, and from what data, and make sure someone on your team reviews it in the first weeks rather than assuming it is working. Never let the model calculate money: a prorated refund, an early-termination fee, or a credit amount should come from your billing system's own logic and be handed to the agent to read back, not computed by the language model itself.

Verify the customer's identity before any balance-affecting action, and route anything that looks like a possible account-takeover or an open dispute to a human until you have live-traffic data showing how the agent performs on that category specifically. This applies to Decagon and Sierra directly, since both are built to take exactly this kind of action, and it applies to XBert too the moment you connect it to a CRM workflow that can trigger a downstream billing change, even if the agent itself is answering the phone rather than issuing the refund.

Frequently asked questions