Updated September 14, 2026

The best AI chatbot platforms for customer service in 2026 can change something in your systems, not just answer. What the market still calls an AI chatbot platform, we would call an AI agent. Twelve are compared here on pricing model, deployment, actions and how they count a resolution.

Key takeaways

• Only 4 of the 12 platforms publish a per-unit price; the other 8 quote custom contracts, and third-party estimates for those run from $30,000 to $300,000+ a year.

• Pricing has split into three models: per resolution (Fin, Zendesk, Gorgias), per conversation or action (Agentforce, Ada, Decagon), and custom enterprise (Sierra, Kore.ai, Cognigy).

• Every vendor that publishes a resolution definition counts a customer who simply stopped replying as resolved.

• Two platforms on the list offer on-premises or air-gapped deployment as a documented option; the rest are cloud-first.

• We build one of these platforms (CXOS). It is included below with the same columns as everyone else, and with the cases where it is not the right choice.

How we chose and scored the twelve

We picked platforms that a customer service leader in 2026 will actually meet in an evaluation: the ones that appear repeatedly in enterprise shortlists, analyst comparisons and competitor pricing pages. We excluded pure IT help-desk tools and general chatbot builders with no customer-service focus.

Each platform is scored on the four things that decide whether the purchase works, rather than on feature count:

1. Actions. Can it write to your systems (create a return, update an address, file a claim), or only read from them?

2. Pricing model. Per resolution, per conversation, per action, or custom; and whether the rate is published.

3. Deployment. Cloud only, or also private cloud, on-premises, and bring-your-own-model.

4. Measurement. What the vendor counts as a resolution, and whether you can verify it outside their dashboard.

All vendor figures are as of September 2026 and come from the vendors' own pricing pages or, where none exists, from named third-party procurement data. Estimates are marked as estimates. Rates change; check before you quote them internally.

Caption: The two columns that matter most are pricing model and deployment. Feature lists converge; who carries the cost of failure and where your data sits do not.

The published-price group: Fin, Zendesk, Gorgias, Freshdesk

Four platforms tell you the price before you talk to sales, which makes them the easiest to budget and the easiest to compare. They also share a structure: the AI is billed per unit of work on top of a seat or plan fee, so the bill grows with volume and with the vendor's own definition of the unit.

Intercom Fin is the clearest of the group. Intercom's Fin pricing page lists $0.99 per outcome with a 50-outcome monthly minimum. An outcome includes a customer confirming the answer helped or leaving without asking for more, and other billable outcomes exist at the same rate. Fin works on non-Intercom helpdesks too. It is the benchmark everyone else is priced against.

Zendesk AI Agents bill $1.50 per automated resolution on committed volume and $2.00 pay-as-you-go, after a small allowance included per agent seat, with the Copilot add-on at $50 per agent per month on top. A resolution is a request handled without human escalation. If you are already on Zendesk, the integration cost is close to zero, which is the real argument.

Gorgias charges roughly $0.90 per AI resolution on annual plans, with plan tiers priced by ticket volume. It is built for Shopify stores, with order data wired in; outside that world it is the wrong tool.

Freshdesk's Freddy AI is priced per session at $0.10, the cheapest published unit on the list. It suits Freshworks customers with simple, high-volume questions, and its ceiling shows quickly on multi-step requests.

The custom-contract group: Sierra, Decagon, Ada, Kore.ai, Cognigy

Five platforms publish no rate card. That is not a criticism; it reflects contracts that bundle implementation, volume tiers and outcome guarantees. It does mean your first real number arrives after a discovery call, and that third-party procurement data is the only public reference.

Sierra sells outcome-based pricing and says so in its own blog; the rate is negotiated per customer. Third-party estimates put year-one contracts at $150,000 and above. Sierra's engineers build alongside the customer, and its published research (τ-bench and successors) sets the tone for agent evaluation across the category.

Decagon quotes deployments per conversation or per resolution; its own product blog says most customers choose per conversation, for predictable invoices. Public reports describe an annual platform fee plus usage. Decagon also documents an air-gapped deployment option, which is rare in this group.

Ada is quote-based. Procurement data from Vendr puts the median contract near $70,000 a year, with reported per-interaction rates between $1.00 and $3.50 depending on tier and volume. Ada's strength is multi-channel scale; its cost curve is the thing to model carefully.

Kore.ai and Cognigy are the contact-centre platforms of the list: voice, IVR, and dozens of channels, with enterprise deployment options and custom pricing. Estimates for Cognigy start around $150,000 a year. Both are strongest where a large voice operation already exists.

What "resolution" means on each platform

Every platform on this list reports a resolution rate, and the number is not comparable across vendors because the definition is not. In every published definition we could find, a customer who stops replying after the AI's last message is counted as resolved. Customers also go quiet when the answer was wrong.

This matters twice. It matters for the headline metric, where a platform can look better by counting silence more generously. And it matters for the bill under per-resolution pricing, where the vendor's dashboard is the meter and the party that gets paid defines the unit.

The number we recommend asking every vendor for instead is the share of conversations that ended with a completed transaction in your system: a return booked, an address changed, a claim filed, a ticket opened with the right fields. A transaction either exists in your order or policy system or it doesn't, and neither party controls that log. A Turkish mobility operator running CXOS reports the agent handling 80% of its support volume; the proof sits in the operator's own ticketing system, not in ours, and that is the standard we would hold any vendor on this list to.

Caption: Read this table before the pricing table. A cheap unit with a loose definition costs more than an expensive unit with a strict one.

Where CXOS fits, and where it doesn't

We are on this list, so here is the same scrutiny applied to us. CXOS is priced as a platform fee plus a fee per completed transaction, with no charge for conversations handed to a person. It runs in four deployment modes, including on-premises and on your own model keys, and it ships with a forward deployed engineer inside the customer's workflow from discovery to handover.

It is the wrong choice in three cases. If you want a self-serve trial with a credit card, Fin, Gorgias or Tidio will get you live faster. If your operation is voice-first with an existing IVR estate, Kore.ai or Cognigy are built for that. And if your team is already inside Zendesk or Salesforce and the AI only needs to answer, not act, the native add-on will be cheaper to run than any separate platform, including ours.

It is the right choice when the agent has to complete transactions in systems a regulator will ask about, in Turkish and other languages, with the data inside a perimeter you define. A Turkish insurer runs document automation on our platform at 90% straight-through, up from 70%; a fashion retailer with 40,000+ products runs product questions, order tracking and returns in production. Those are the workloads this platform is shaped for.

How to choose a chatbot platform in 2026

Start from the unit you will be billed on and the record you will be able to check, not from the feature grid. Ask each vendor for their written definition of a resolution, for the deployment modes they will put in a contract, and for a production example where the agent changed a record in a customer's own system. Then run the same five conversations through every shortlisted platform against your real knowledge base.

Three filters remove most of the list quickly. Regulated data or on-premises requirement: Decagon, CXOS, and the enterprise tiers of Kore.ai and Cognigy remain. Shopify store under a few thousand tickets a month: Gorgias or Tidio. Already on Zendesk, Salesforce, Intercom or Freshworks with a mostly informational workload: start with the native option and measure before you add a platform.

For the full evaluation, use the 27 questions in our RFP template and the 12 questions that separate an agent from a chatbot; both link below.

Related guides

• AI in Customer Service: The Complete Guide (2026)