Table of Contents

• Why the 5-to-25-Agent Problem Is a Pricing-Unit Problem

• What Counts as a Billable Resolution (and Why Vendor Prices Are Not Comparable)

• Who Owns Whom in 2026: The Support-AI Consolidation Map

• AI Disclosure Is Now Law: EU AI Act Article 50 for Support Teams

• How We Evaluated AI Customer Service Agents for Growing Teams

• The 9 Best AI Customer Service Agents in 2026

• Platform Summary Table

• Vendor Funding and Viability Scorecard

• Benchmarks You Can Actually Compare in 2026

• How to Choose the Right Platform for Your Ticket Volume

• Implementation Checklist

• Final Verdict

Why the 5-to-25-Agent Problem Is a Pricing-Unit Problem

An AI customer service agent is software that reads a customer’s message, retrieves an answer from your approved knowledge sources, and either resolves the request end to end (including backend actions like refunds or plan changes) or hands it to a human with full context. It differs from a chatbot because it reasons over your systems rather than matching keywords to canned replies. For a team scaling from 5 agents to 25, the decision that matters is not which agent is smartest, it is which billing unit stops your headcount curve.

Here is the arithmetic that makes this urgent. A support org handling 12,000 tickets a month at a 6-minute average handle time needs roughly 13 to 15 full-time agents. Automate 60% of that volume and you avoid hiring eight of them.

But avoided headcount only counts if the automation bill stays below the salary bill. At 12,000 monthly tickets and 60% automation, you are paying for 7,200 automated units a month. At $0.49 per unit that is $3,528. At $0.99 per unit it is $7,128. At $1.50 it is $10,800, which is roughly one and a half loaded agent salaries reappearing as a software line item.

Now double the volume. Per-resolution pricing scales linearly with tickets, which is exactly what you want when ticket growth outpaces revenue growth, and exactly what you do not want when it does not. Per-seat pricing does the opposite: it punishes you for the humans you keep, then charges you again for the automation meant to replace them.

According to HubSpot’s 2026 buyer comparison, citing NICE’s “The Agentic AI CX Frontline 2026” report, the highest-performing AI agents now reach 80%+ containment on tier-one inquiries, and the same guide cites IBM research showing mature AI adopters achieve 17% higher CSAT (HubSpot). Containment at that level changes the staffing model entirely. It also means the difference between a $0.49 unit and a $1.50 unit compounds across tens of thousands of interactions a year.

This guide ranks nine platforms on that axis. If your buying question is about phone channels, per-minute economics or IVR replacement, our companion breakdown of voice-native support platforms covers that decision properly, and this page deliberately does not.

What Counts as a Billable Resolution (and Why Vendor Prices Are Not Comparable)

The four leading pricing units in this category measure four different things. An “outcome” can be triggered by silence. A “session” is a time window, not a success. A “resolution” may or may not be verified by the customer. Comparing headline per-unit prices across vendors without normalizing these definitions produces cost forecasts that are wrong by multiples, not percentages.

Intercom defines a Fin outcome as a customer confirming their issue is resolved, a customer not asking for more help after Fin responds, or Fin completing a workflow, billed once per conversation regardless of how many actions Fin takes, per Intercom’s pricing page as of 2026-08-06. That middle clause is the one to model carefully: a customer who abandons the chat in frustration and never replies produces the same billable event as a customer who says thank you.

Freshworks bills a session, which for the email AI Agent is a 72-hour window starting from the customer’s first email, with every AI reply inside that window counting as one session, sold in packs of 100 for $49 according to a July 2026 pricing analysis (eesel). That is an attempt-based unit. If the agent tries and fails, you still pay.

Zendesk bills per successfully automated resolution on a success-based model, with AI agents included on every Suite and Support plan, but publishes no per-resolution figure on its pricing page as of 2026-08-06 (Zendesk pricing). Gorgias states on its pricing page that you pay only when the AI Agent resolves a conversation, with the helpdesk priced by ticket volume from 50 to 5,000 tickets a month rather than per agent (Gorgias).

Three practical rules follow. First, ask every vendor in writing what percentage of billed units in a comparable account came from customer confirmation versus inferred success. Second, require a monthly reconciliation report that lists billable events with their trigger type. Third, negotiate a credit clause for units billed on conversations that reopened within 72 hours, because a reopened ticket is a failed resolution you paid for twice.

Fini bundles a monthly resolution allowance into each plan rather than metering from zero, and unused allowance rolls forward one month, so a seasonal spike does not immediately convert into overage. That structure exists specifically because forecast variance, not unit price, is what breaks support budgets at the 5-to-25-agent stage.

Who Owns Whom in 2026: The Support-AI Consolidation Map

Two of the nine vendors most commonly ranked in this category are no longer independent companies as of 2026-08-06. Zendesk completed its acquisition of Forethought on March 26, 2026, and Salesforce signed a definitive agreement to acquire Fin, the company formerly known as Intercom, on June 15, 2026. Any 2026 buyer list that presents these as separate competitive options is describing a market that no longer exists.

Zendesk announced the Forethought deal on March 11, 2026 and closed it on March 26, 2026, an all-cash transaction described as Zendesk’s largest acquisition in nearly 20 years, reported at over $200 million, folding self-improving AI agents into the Zendesk Resolution Platform (Zendesk). Computer Weekly characterized it as a major agentic AI play giving Zendesk agents that generate, adapt and execute complex workflows across service channels (Computer Weekly). Zendesk has stated Forethought remains available to non-Zendesk customers.

Salesforce’s agreement to acquire Fin values the company at approximately $3.6 billion, with Fin folding into Agentforce and close expected before the end of Salesforce’s fiscal Q4 2027, subject to regulatory approval (Salesforce). CNBC reported that Fin’s AI agent resolves 76% of support queries without a human, and that Salesforce’s Agentforce reached $1.2 billion in annual recurring revenue in its most recent quarter, up 205% year over year (CNBC).

The buyer consequence is specific, not abstract. Fin’s most attractive property for a growing team was neutrality: it ran on top of Zendesk and Salesforce so you did not have to migrate helpdesks. Once the acquisition closes, running Fin over a Zendesk helpdesk means running a Salesforce product over a competitor’s platform, and roadmap priority for that configuration is not something either party will guarantee in writing.

What to negotiate now if your vendor might get acquired mid-contract:

• A change-of-control clause that lets you exit without penalty within 90 days of a closed acquisition.

• Price protection that fixes your per-unit rate for the full contract term regardless of ownership.

• Data portability language: full conversation history, knowledge base configuration and workflow definitions exportable in machine-readable format within 30 days of a written request.

• A written commitment on continued support for non-native helpdesks, with a named deprecation notice period of at least 12 months.

• Renewal caps, since repricing at renewal is the most common way an acquisition reaches your budget.

AI Disclosure Is Now Law: EU AI Act Article 50 for Support Teams

As of 2 August 2026, the EU AI Act’s Article 50 transparency obligations are enforceable. Providers of AI systems that interact with people must clearly inform those people they are dealing with an AI system, and AI-generated content must be marked in machine-readable, detectable form. Penalties reach 15 million euros or 3% of worldwide annual turnover, whichever is higher (Cooley).

The European Commission published its “Safer and more transparent AI” notice dated 2 August 2026 marking the start of the transparency regime (European Commission). If you serve EU end users, this applies to your support chat widget today, regardless of where your company is incorporated.

Do not over-scope it. High-risk obligations did not begin in August 2026: stand-alone Annex III systems have until 2 December 2027, and AI embedded in regulated products under Annex I has until 2 August 2028, per Norton Rose Fulbright’s Data Protection Report (Data Protection Report). A tier-one support agent answering order-status questions is almost certainly not a high-risk system.

The vendor question to ask is concrete: show me the disclosure string, where I configure it per channel, and how AI-generated content is marked in a machine-readable way. A vendor that treats this as a settings toggle you have to find yourself is offloading a regulatory obligation onto your support ops lead. Add it to your AI compliance review alongside the certification checklist rather than after it.

How We Evaluated AI Customer Service Agents for Growing Teams

Every vendor below was assessed against six criteria weighted for teams between 5 and 25 support agents, where budget headroom is limited and a bad twelve-month contract is a material mistake. Certifications and funding were verified against publisher sources dated 2026 wherever possible. Where a figure could not be sourced, this guide says “does not publicly state” rather than estimating.

1. Billing-unit honesty. We asked what triggers a charge, not what the headline rate is. A vendor billing attempts (sessions) is structurally different from one billing verified outcomes, and a vendor that counts customer silence as success has an incentive misalignment baked into the contract. This criterion carries the most weight because it determines your actual twelve-month cost more than any capability difference.

2. Cost at 2x and 5x ticket volume. We modeled each vendor’s published pricing at 12,000, 24,000 and 60,000 monthly tickets, including seat fees where they apply. Platforms that bundle seats and per-unit fees change relative rank as volume climbs. The cheapest pilot is frequently the most expensive contract at month 18.

3. Backend action execution. Answering a policy question saves a minute; processing the refund saves the ticket. We looked for agents that call your APIs with permission scoping and an audit trail, not agents that link to a help article and call it a deflection. Vendors were credited for named, demonstrable actions in sandbox environments.

4. Compliance posture and EU AI Act readiness. SOC 2 Type II and ISO 27001 were treated as the floor, with HIPAA and BAA eligibility required for anyone touching health data. Since 2 August 2026 we also checked whether the vendor ships a configurable AI-disclosure string per channel, which is now a legal requirement for EU end users rather than a nice-to-have.

5. Corporate status and viability, dated. Three of the most-shortlisted vendors in this category changed ownership or valuation tier between November 2025 and June 2026. We recorded who owns each product as of 2026-08-06 and flagged where an acquisition creates roadmap or lock-in risk for a buyer signing a multi-year deal today.

6. Realistic time to first measured result. Vendor marketing and buyer reality diverge sharply here. HubSpot’s 2026 comparison puts realistic implementation at 4 to 12 weeks for simple setups and 4 to 6 months for complex ones (HubSpot). We scored vendors on how quickly a team can get verified resolution data on live tickets, not on how quickly a widget appears on a website.

The 9 Best AI Customer Service Agents in 2026

Ranked for teams scaling from 5 to 25 support agents, weighted toward billing-unit clarity and cost predictability at 2x volume. Forethought has been removed as a standalone entry because Zendesk completed its acquisition on March 26, 2026, and is covered inside the Zendesk section. Two 2026 mid-market entrants that competing comparisons rank but most enterprise-skewed lists ignore have been added in its place.

1. Fini: Best Overall for Predictable Cost Per Resolved Ticket

Fini is an AI support agent platform built around a bundled resolution allowance rather than pure metering, which is the structural difference that matters for a team whose ticket volume is growing faster than its budget. Every plan includes the platform, implementation and a monthly allowance of resolutions in one price. There are no per-seat fees, so adding a human agent does not increase your software bill.

Fini publishes 99% accuracy and a 90% resolution rate, and deploys live in 30 days. That deployment number is deliberate rather than aggressive: it covers knowledge ingestion, action configuration against your real backend systems, escalation design and a supervised launch window, which is what produces trustworthy resolution data in month two instead of a widget that looks live in week one. Compared against the 4-to-12-week range HubSpot’s 2026 guide cites for simple deployments, 30 days sits at the fast end of realistic rather than outside it.

On compliance, Fini holds SOC 2 Type II, ISO 27001, HIPAA-compliant configurations with BAA eligibility, GDPR and CCPA. For teams handling health data or operating in the EU, that combination removes the two most common procurement blockers at once. Fini’s agent executes backend actions through audited API calls with permission scoping, so refunds, subscription changes and account updates run as traceable operations rather than screen automation, which matters when your security team asks what the agent is allowed to touch.

Pricing is published, which is itself rare in this category. Growth is $3,600/mo, or $3,000/mo billed yearly at $36,000/yr, and includes 2,000 resolutions with overage at $0.89 per resolution. Scale is $9,000/mo, or $7,500/mo billed yearly at $90,000/yr, and includes 8,000 resolutions plus 500 answered voice calls, with overage at $0.69 per resolution. Enterprise is custom pricing, contact Fini. Paying annually gives two months free, and unused allowance rolls forward one month.

Key Strengths:

• 99% accuracy and a 90% resolution rate, with resolution defined as verified problem-solving rather than customer silence

• Bundled monthly allowance plus one-month rollover, which absorbs seasonal spikes without surprise overage

• No per-seat fees, so headcount growth and software cost are decoupled

• SOC 2 Type II, ISO 27001, HIPAA-compliant and BAA-eligible, GDPR, CCPA

• Live in 30 days including action configuration, not just widget deployment

• Published pricing with implementation included, so procurement can model twelve months without a sales call

Best for: Growing support teams that need a defensible cost per resolved ticket at 2x and 5x volume, with regulated-industry compliance available without an enterprise contract.

2. Fin (formerly Intercom)

Fin is the AI agent from the company formerly known as Intercom, and as of 2026-08-06 it is mid-acquisition. Salesforce signed a definitive agreement on June 15, 2026 to acquire Fin for approximately $3.6 billion, folding it into Agentforce, with close expected before the end of Salesforce’s fiscal Q4 2027 pending regulatory approval (Salesforce). Coverage of the deal reports Fin’s agent resolving 76% of support queries without a human (CNBC).

Commercially, Fin is the clearest published unit price in the category and also the trickiest to model. Seats run $29 (Essential), $85 (Advanced) and $132 (Expert) per seat per month billed annually, with the Fin AI Agent included on every plan, and Fin itself bills “From $0.99 per Fin outcome” with minimum monthly commitments, per intercom.com/pricing as of 2026-08-06. An outcome fires when a customer confirms resolution, when they do not ask for further help after Fin responds, or when Fin completes a workflow, and lead qualification outcomes are priced separately.

Fin can still run standalone on top of an existing helpdesk at the same $0.99 rate without migrating, which was the strongest argument for it in 2025. That argument now carries an asterisk. Once Salesforce owns the product, the neutral-layer-over-Zendesk story becomes a first-party Salesforce story, and a growing team signing a two-year deal today should assume roadmap priority follows the parent company.

Pros:

• 76% of support queries resolved without a human, per June 2026 acquisition coverage

• Published per-outcome pricing that procurement can model directly

• Runs standalone over an existing helpdesk without migration

• Mature analytics and AI-assisted QA in a single suite

Cons:

• Seat fees plus per-outcome fees means two cost curves growing at once

• The outcome definition counts customer silence, which can bill abandoned chats

• Salesforce ownership introduces roadmap and lock-in risk for non-Salesforce shops

• Minimum monthly commitments apply to standalone deployments

Best for: Teams already committed to Intercom, or Salesforce shops that will benefit from the Agentforce consolidation once the deal closes.

3. Decagon

Decagon sells an enterprise AI support agent centered on the Agent Operating Procedure, a construct in which support runbooks are written in natural language and the agent executes them step by step while calling APIs, so every action traces back to a written procedure. That approach appeals to teams that need auditable behavior more than conversational flair, and it makes revisions a documentation task rather than an engineering ticket.

The company’s market position changed sharply in early 2026. Decagon raised a $250 million Series D led by Coatue Management and Index Ventures at a $4.5 billion valuation, tripling its valuation in under six months, and named Avis Budget Group, Block, Deutsche Telekom, Oura Health, Affirm and Chime as customers while citing more than 100 new global enterprise customers added in the fiscal year (Decagon). Business Wire confirmed the valuation triple and the investor set, which added ChemistryVC, Definition Capital and Starwood Capital alongside a16z, Accel, Bain Capital Ventures and Ribbit (Business Wire).

Decagon does not publicly state pricing, and there is no self-serve tier, so evaluation runs through sales on annual terms. Its certification posture is not listed in its own Series D announcement, so verify it against Decagon’s trust documentation before procurement rather than assuming parity with older vendors. For a 10-person support team, the practical barrier is usually contract structure, not capability.

Pros:

• Agent Operating Procedures make behavior explicit, auditable and revisable without code

• Customer roster now spans regulated and enterprise names including Deutsche Telekom, Block and Affirm

• $250M Series D at $4.5B (January 2026) substantially reduces vendor-viability risk

• Covers chat, email and messaging channels from one procedural agent layer

Cons:

• Does not publicly state pricing; sales-led with no self-serve evaluation path

• Certification set is not published in its own announcements and needs direct verification

• Implementation quality depends on how well your team writes and maintains procedures

• Enterprise-oriented contracting makes short pilots harder for sub-25-agent teams

Best for: Mid-market and enterprise teams with the operational discipline to maintain written runbooks and the budget for annual sales-led contracts.

4. Sierra

Sierra builds branded customer-facing agents and has become the most heavily capitalized independent vendor in the category. It raised approximately $950 million in May 2026 at a $15.8 billion post-money valuation, led by GV and Tiger Global with Benchmark, Sequoia and Greenoaks participating, up from roughly $10 billion in autumn 2025, and is reported to count more than 40% of the Fortune 50 as customers (TechCrunch). It crossed $100 million in annual recurring revenue in under two years from founding, reported November 2025 (TechCrunch).

Architecturally, Sierra’s differentiator is a supervisory layer that checks agent outputs against accuracy and policy constraints before responding, paired with developer tooling for defining skills, guardrails and escalation logic. It handles voice as well as chat, though voice buyers should read the companion voice guide rather than treating this entry as a voice evaluation.

Sierra does not publicly state pricing. Contracts are outcome-based and negotiated, oriented toward large enterprise deal sizes. For a team of 5 to 25 agents, the limiting factor is procurement minimums and the length of the implementation engagement, not whether the technology can do the job.

Pros:

• $15.8B valuation and roughly $1.6B raised in total signal exceptional staying power

• Supervisory model layer checks outputs for accuracy and policy compliance before sending

• Outcome-based contracting aligns spend with resolved conversations

• Developer tooling gives engineering teams deep control over agent behavior

Cons:

• Does not publicly state pricing, so cost comparison requires a full sales cycle

• Procurement minimums are oriented toward enterprise, not 10-person support teams

• Implementation runs longer than platforms with bundled 30-day deployment

• Enterprise-first roadmap means mid-market feature requests compete with Fortune 50 accounts

Best for: Large enterprises with engineering capacity and appetite for a custom-engineered, heavily branded agent.

5. Zendesk (including Forethought AI Agents)

Zendesk is now two acquisitions deep into AI agents. It bought Ultimate in 2024 to create the Zendesk Resolution Platform, then completed its acquisition of Forethought on March 26, 2026, an all-cash deal reported at over $200 million and described as its largest acquisition in nearly 20 years (Zendesk). The Forethought technology brings self-improving agents that generate, adapt and execute complex workflows across service channels (Computer Weekly). Zendesk states Forethought AI Agents remain available to non-Zendesk customers.

The commercial model is seats plus success-based automation. Zendesk’s live pricing ladder shows Suite Team at 55 euros per agent per month and Suite Professional at 115 euros per agent per month billed annually, with Suite Enterprise + Copilot as contact-sales; AI agents are included on every Suite and Support plan and billed per successfully automated resolution (Zendesk). Zendesk publishes no per-resolution figure, so any comparison against a named competitor rate is a comparison against an unknown. Note also that the pricing page geo-redirects, so EUR and USD figures differ and your regional rate must be checked on your local page.

For teams already paying for Zendesk seats, activation friction is close to zero, and that is the honest case for choosing it. For teams not on Zendesk, you are adding a per-agent cost curve on top of a per-resolution cost curve, which is the exact structure a growing team should be trying to avoid.

Pros:

• Native to a helpdesk a very large share of support teams already run

• Two acquisitions (Ultimate 2024, Forethought 2026) have deepened agentic workflow capability

• AI agents included on every Suite and Support plan with no separate product purchase

• Mature enterprise compliance and procurement posture

Cons:

• Does not publicly state a per-automated-resolution rate, making twelve-month forecasting guesswork

• Seat fees plus resolution fees compound as headcount and volume both grow

• Pricing page geo-redirects, so published figures may not be your currency or rate

• Overlapping tooling from Ultimate and Forethought will take time to consolidate cleanly

Best for: Existing Zendesk customers who value single-vendor consolidation over transparent per-unit economics.

6. Freshworks Freddy AI

Freshworks positions Freddy as the value option, split into AI Agent for customer-facing automation, AI Copilot for human agents and AI Insights for managers. Setup is genuinely fast and no-code, which suits lean ops teams without engineering support, and grounding draws on your existing knowledge base with flow builders for common actions like order tracking.

The pricing correction matters more than anything else here. Freddy AI Agent is sold in session packs of 100 for $49, roughly $0.49 per session, not the $0.10-per-session figure that circulated in 2025 buyer guides, per a July 2026 pricing analysis (eesel). Freddy AI Copilot runs about $29 per agent per month on annual billing and about $35 monthly, and Freshdesk base ticketing runs roughly $19 per agent per month (Growth) to $89 per agent per month (Enterprise) on annual billing. Growth, Pro and Enterprise plans include 500 free trial sessions once per account.

The session definition is the catch. For the email AI Agent, a session is a 72-hour window from the customer’s first email, with all AI replies inside that window counting as one session. That means you are buying attempts rather than outcomes, so a 60% success rate turns an advertised $0.49 session into an effective $0.82 per actual resolution. Freshworks blocks automated fetches, so re-confirm all of these figures on the official pricing page before signing.

Pros:

• Lowest published entry price among vendors that publish at all, at $49 per 100 sessions

• 500 free trial sessions on paid plans lets you measure before committing

• Copilot, customer-facing agent and manager analytics bundled in one suite

• Fast no-code setup suited to teams without engineering support

Cons:

• Sessions bill attempts, not outcomes, so effective cost per resolution is meaningfully higher

• A 72-hour email session window makes usage forecasting harder than per-conversation billing

• Autonomous action execution is lighter than dedicated agent-first platforms

• Best experience assumes full commitment to the Freshworks stack

Best for: Budget-constrained Freshdesk teams that want incremental AI lift and are willing to model attempt-based billing carefully.

7. Ada

Ada is one of the longer-standing independent vendors in support automation and remains a regular shortlist name for multilingual mid-market and enterprise operations. It is worth stating plainly what this guide can and cannot verify: ada.cx blocks automated fetches, and no dated Ada source could be opened during research for this update. Every specific claim about Ada’s founding, install base, customer logos, language coverage, scoring methodology and certification posture should be verified directly with the vendor rather than taken from any comparison article, including this one.

What can be said with confidence is structural. Ada does not publicly state pricing, so evaluation runs sales-led on annual terms with no self-serve path. That places it in the same procurement category as Decagon and Sierra: capable, but requiring a full sales cycle before you can even build a cost model.

For a growing team, the practical implication is timeline. If you need a defensible twelve-month cost forecast in the next two weeks, a vendor with no published rate cannot give you one, and a discovery call will not compress into a spreadsheet. Ask specifically for a rate card in writing, the billable-unit definition, and current SOC 2 Type II and GDPR documentation with issue dates.

Pros:

• Long-standing presence in the category with a mature no-code builder for support ops teams

• Multichannel coverage across chat, email and messaging in a single agent configuration

• Frequently shortlisted by mid-market teams with multi-language customer bases

• Sales-led model means implementation support is bundled into the engagement

Cons:

• Does not publicly state pricing, so no cost model exists before a sales cycle

• Certification and capability claims could not be independently verified from public sources in this research pass

• Annual contracting with no free tier limits low-commitment evaluation

• Advanced API-driven actions still benefit from developer involvement

Best for: Mid-market and enterprise teams with international customer bases who can absorb a sales-led evaluation timeline.

8. Gorgias

Gorgias is the ecommerce specialist of this list, built around Shopify, BigCommerce and Magento data rather than generic connectors. That vertical focus is the whole proposition: the agent reads order status, edits shipping addresses, processes returns and applies discount codes natively, which is why it excels at where-is-my-order tickets that dominate DTC ticket mix.

The commercial model is unusually well aligned for a growing merchant. Gorgias’s pricing page states that AI Agent is on every plan and you pay only when it resolves a conversation, and that the helpdesk scales from 50 to 5,000 tickets a month and is never priced per agent (Gorgias). Not being priced per agent is the single most useful property here for a team adding seasonal support staff, because your Black Friday headcount does not multiply your software bill.

The dollar figures render in an interactive component that could not be extracted during research, so Gorgias does not publish an easily citable per-resolution rate or plan price in static form. Get both in writing before signing. Outside ecommerce, the platform has limited applicability, and SaaS, fintech and services teams will find the integration catalog pointed firmly at online retail.

Pros:

• Deepest native Shopify, BigCommerce and Magento actions of any vendor on this list

• Never priced per agent, so seasonal support staffing does not raise the software bill

• Pay only when the AI Agent resolves a conversation, an outcome-aligned unit

• Ticket-volume tiers from 50 to 5,000 per month suit small and scaling merchants alike

Cons:

• Effectively unusable outside ecommerce verticals

• Plan prices and per-resolution rate are not published in extractable static form

• Ticket-tier plans plus AI resolution fees complicate forecasting across seasonal peaks

• Compliance documentation could not be verified from public sources in this research pass

Best for: DTC and ecommerce brands on Shopify or BigCommerce whose ticket mix is dominated by order status and returns.

9. Mid-Market No-Code Specialists (Zowie, Tidio Lyro, Kore.ai, Cognigy)

The slot vacated by Forethought is best filled not by one vendor but by a category that enterprise-skewed lists routinely omit: mid-market no-code agent specialists that teams under 25 agents actually shortlist. Zowie, Tidio Lyro, Kore.ai and Cognigy all appear in top-ranking 2026 comparisons for this buyer, and they compete primarily on setup speed and self-serve evaluation rather than enterprise contracting.

These platforms matter to the 5-to-25-agent buyer for one reason: they let you measure before you commit. Where Decagon, Sierra and Ada require a sales cycle before you can build a cost model, this tier usually publishes rates and lets a support ops lead run a pilot without procurement involvement. That shortens the gap between shortlist and evidence from months to weeks.

None of their pricing, certification or performance claims were verified during research for this update, so treat them as a shortlist input rather than a recommendation. Ask each one the same three questions you ask the enterprise vendors: what triggers a billable unit, what does the EU AI Act disclosure string look like and where is it configured, and what SOC 2 Type II documentation can you produce with an issue date.

Pros:

• Self-serve or low-friction evaluation without a full enterprise sales cycle

• Frequently published rates, enabling cost modeling before a vendor call

• No-code builders suited to support ops leads without engineering support

• Fast time to a first measurable pilot on live tickets

Cons:

• Pricing, certification and performance claims unverified in this research pass

• Generally lighter compliance documentation than enterprise-grade vendors

• Backend action execution depth varies widely and needs sandbox testing

• Vendor viability and funding position not established from dated public sources here

Best for: Teams under 25 agents that need a live pilot with real resolution data before committing budget to a twelve-month contract.

Platform Summary Table

The table below compares each platform on the six evaluation criteria, using only figures verified against a dated source as of 2026-08-06. Where a vendor does not publish a number, the cell says so rather than estimating. Every price should be re-checked on the vendor’s own page before contracting, since three of these vendors geo-redirect or render pricing dynamically.

Vendor Funding and Viability Scorecard

Vendor viability moved from a footnote to a procurement criterion in this category during 2026, because two of the most-shortlisted names changed owners inside six months. The figures below are stamped as of 2026-08-06 and should be re-verified before any multi-year contract. Funding scale reduces the risk of a vendor disappearing; it does not reduce the risk of a vendor being acquired and repriced.

Benchmarks You Can Actually Compare in 2026

Half the numbers quoted in this category are forecasts wearing the clothing of measurements. Separating them is the difference between a business case and a wish. The table below labels each widely cited figure as measured or predicted, with its date and publisher.

Note the distinction between containment and resolution, because vendors blur it deliberately. Containment is the share of contacts handled without escalating to a human; a contained conversation can still leave the customer’s problem unsolved. Resolution requires the problem to be fixed, which is why reopen rate within 72 hours belongs on your dashboard next to any containment figure.

When a vendor quotes you a benchmark, ask three questions: is it measured or projected, what population was it measured on, and what was the reopen rate. A 90% containment figure on password resets tells you nothing about billing disputes.

How to Choose the Right Platform for Your Ticket Volume

Choose on unit economics at your projected twelve-month volume, then on action depth, then on compliance, in that order. Capability differences between the top platforms have narrowed enough that pricing structure is now the dominant variable in total cost. The six steps below are ordered to fail fast and cheap.

1. Tag 90 days of tickets by automatability before you talk to a vendor. Pull your ticket export and classify every ticket as repetitive tier-one work, judgment call or edge case. If 50% or more is order status, password resets and policy questions, an outcome-priced platform pays back within two quarters. Our breakdown of tier-one ticket automation covers how to run that classification without a data team.

2. Normalize every quote into cost per verified resolution. Take each vendor’s unit price, divide by the success rate they will commit to in writing, and add seat fees where they apply. A $0.49 session at 60% success is $0.82 per actual resolution; a $0.99 outcome that includes silence-triggered billing may be higher than its sticker price too.

3. Model at 1x, 2x and 5x monthly ticket volume. Build one spreadsheet row per vendor per volume tier, including seat growth, and look at where the lines cross. Bundled-allowance models look expensive at 1x and cheap at 5x; pure per-unit models do the reverse.

4. Run a grounding stress test on every finalist. Feed each agent 50 questions your documentation answers and 20 it does not. The wrong-answer rate on the second set predicts your cleanup-ticket volume better than any accuracy claim, and structured AI red teaming on adversarial phrasing should be part of the same exercise.

5. Force a real backend action in a sandbox. Require the agent to actually process a refund, change a subscription or update an account record in your test environment, with the audit log visible. Vendors whose support tools update customer accounts through scoped API calls will show you the trace; vendors relying on screen automation will show you a video instead.

6. Check compliance against your 18-month roadmap, including EU AI Act configuration. If healthcare data, EU expansion or financial services are in scope within 18 months, require HIPAA compliance documentation, data residency options and a per-channel AI disclosure string now. Re-platforming an agent mid-growth costs far more than picking a certified vendor on day one.

Implementation Checklist

Deployment failure in this category is usually a sequencing problem, not a technology problem. Teams connect a messy knowledge base to a capable agent, get poor answers, and blame the model. The four phases below front-load the work that determines whether your resolution rate lands at 40% or 85%.

Phase 1: Pre-Purchase

• Export and tag 90 days of tickets, ranking the top 20 intents by volume and handle time

• Write down every backend action the agent must perform, with the system and API behind each

• Get each vendor’s billable-unit definition in writing, including what triggers a charge on customer silence

• Build a twelve-month cost model at 1x, 2x and 5x volume, including seat fees

• Confirm SOC 2 Type II and ISO 27001 documentation with issue dates, plus HIPAA and BAA eligibility if relevant

Phase 2: Evaluation

• Run the 50-answerable / 20-unanswerable grounding test and record wrong-answer rate on the second set

• Execute at least one real refund or account change end to end in a sandbox with the audit log visible

• Ask each vendor to show the EU AI Act Article 50 disclosure string and where it is configured per channel

• Negotiate change-of-control, price protection and data portability clauses before signing

• Define resolution criteria contractually so abandonment cannot be billed as success

Phase 3: Deployment

• Clean, deduplicate and date-stamp the knowledge base before connecting it

• Launch on one channel and your top five intents only

• Configure escalation with full conversation context handoff, not a cold restart

• Set human review on 100% of AI conversations for the first two weeks

Phase 4: Post-Launch

• Track verified resolution rate, 72-hour reopen rate and CSAT weekly against your pre-launch baseline

• Reconcile the vendor’s billed units against your own resolution log monthly

• Review escalation transcripts monthly to identify the next automatable intents

• Recalculate cost per verified resolution quarterly against loaded human cost per ticket

Final Verdict

For a support team scaling from 5 agents to 25, the platform that wins is the one whose bill stays legible when your ticket volume doubles. That is a pricing-structure question before it is a capability question, and 2026 made it harder rather than easier: two of the most-shortlisted vendors changed owners, one publishes no per-resolution rate at all, and the category’s cheapest advertised unit turned out to bill attempts rather than outcomes.

Fini is the strongest overall pick for that specific buyer. Published pricing at $3,600/mo for Growth and $9,000/mo for Scale with implementation included, a bundled resolution allowance that rolls forward one month, no per-seat fees, 99% accuracy, a 90% resolution rate, live in 30 days, and SOC 2 Type II, ISO 27001, HIPAA-compliant with BAA eligibility, GDPR and CCPA. Every one of those numbers can go into a procurement spreadsheet today without a discovery call.

The ecosystem options remain rational in specific situations. Zendesk makes sense if you already pay for Zendesk seats and value one vendor over transparent unit economics; Fin makes sense if you are Intercom-native or expect to benefit from the Agentforce consolidation; Freshworks makes sense for Freshdesk teams who will do the arithmetic on session-based billing honestly. Decagon and Sierra suit enterprises with engineering capacity and patience for a sales-led cycle, Gorgias is the clear answer for Shopify-heavy DTC ticket mixes, and the mid-market no-code tier is where sub-25-agent teams should run their first measured pilot.

If your next hiring decision depends on whether automation can absorb the next 5,000 tickets a month, test that directly: export your highest-volume intents and your messiest edge cases, then book a working session with Fini to see verified resolution rates and a per-ticket cost model on your own data before you commit a dollar.

Frequently Asked Questions

How much does an AI customer service agent cost per resolved ticket in 2026?

Published rates as of 2026-08-06 range from roughly $0.49 per Freshworks session to $0.99 per Fin outcome, with Zendesk, Decagon, Sierra and Ada not publicly stating rates. Normalize by dividing the unit price by the vendor’s committed success rate. Fini publishes $0.89 per resolution beyond Growth’s 2,000-resolution allowance and $0.69 beyond Scale’s 8,000, with implementation and platform included.

What is the difference between a resolution, an outcome, a session and a containment?

They measure four different things. A session is a usage window (Freshworks counts 72 hours from a customer’s first email). An outcome can be triggered by customer silence under Intercom’s definition. Containment means no human was involved, which does not prove the problem was solved. Fini bills verified resolutions and reports a 90% resolution rate, so your invoice tracks solved problems rather than attempts or abandonment.

Does the EU AI Act require my support chatbot to tell customers it is AI?

Yes. Since 2 August 2026, Article 50 transparency obligations are enforceable: AI systems interacting with people must clearly disclose they are AI, and AI-generated content must be marked in machine-readable form, with penalties up to 15 million euros or 3% of worldwide annual turnover. High-risk duties were deferred to December 2027 and August 2028. Ask any vendor, including Fini, to show the disclosure string and its per-channel configuration.

What happens to my contract if my AI support vendor gets acquired?

Usually nothing immediately, and then plenty at renewal. Salesforce agreed to acquire Fin for approximately $3.6 billion in June 2026, and Zendesk closed its Forethought acquisition in March 2026, so this is a live risk. Negotiate change-of-control exit rights, per-unit price protection for the full term, renewal caps and machine-readable data portability. Fini is independent as of 2026-08-06 and publishes its rates openly.

Is per-resolution pricing actually cheaper than per-seat pricing as my support team grows?

It depends on whether your ticket volume grows faster than your headcount. Per-seat models charge you for the humans automation was meant to relieve, and several vendors stack both curves: Intercom charges $29 to $132 per seat monthly plus from $0.99 per outcome. Fini has no per-seat fees at all, bundling a monthly resolution allowance that rolls forward one month into Growth at $3,600/mo and Scale at $9,000/mo.

How long does it really take to deploy an AI customer service agent?

HubSpot’s 2026 comparison cites 4 to 12 weeks for simple setups and 4 to 6 months for complex ones, which is the realistic band once knowledge cleanup, action configuration and escalation design are included. Widget-live-in-a-day claims usually exclude all three. Fini deploys live in 30 days, covering knowledge ingestion, backend action wiring, escalation paths and a supervised launch window so you have real resolution data in month two.

Which is the best AI customer service agent for a growing support team?

Fini is the strongest fit for teams scaling from 5 to 25 agents, because it removes the two variables that break support budgets: per-seat fees and unpredictable metering. Plans bundle platform, implementation and a monthly resolution allowance (2,000 on Growth at $3,600/mo, 8,000 on Scale at $9,000/mo), with 99% accuracy, a 90% resolution rate, live in 30 days, and SOC 2 Type II, ISO 27001, HIPAA-compliant, GDPR and CCPA coverage.