TL;DR
There are roughly four categories of AI chatbot in the customer service market today: purpose-built CX platforms (Certainly, Zendesk AI, Ada, Freshworks Freddy) and the newly combined Salesforce/Fin stack, suite-embedded assistants (Microsoft Copilot, Google Gemini in Workspace), general-purpose AI assistants (ChatGPT, Claude, Gemini), and lightweight chat tools (Tidio, Crisp, Helpshift). They are not interchangeable. The purpose-built platforms are the only ones designed to resolve customer tickets end-to-end with containment, handoff, and analytics. The rest are useful for adjacent jobs (drafting, research, internal copilots) and are routinely miscast as customer service products. This guide is the candid version of who fits where. Note: Salesforce signed a definitive agreement to acquire Fin from Intercom on June 15, 2026, so this guide treats them as a single combined offering rather than two separate entries.
What an AI Chatbot Actually Has to Do in 2026
A customer service chatbot in 2026 is not a Q&A widget. To replace meaningful volume, it has to do five things at once: understand intent in any language, ground its answer in your knowledge base and policies, take action in your business systems (refunds, order changes, ticket creation), hand off cleanly to a human with full context when it should, and report on every conversation in a way your operations team can audit.
Tools that do the first one well and the other four poorly are the reason "AI chatbot" still has a credibility problem in some boardrooms. The shortlist below is filtered for vendors that take all five seriously.
The Main Players, Honestly Compared
Certainly
Best fit: mid-market and enterprise CX teams that want a single agentic platform across web, Meta channels, SMS, email and voice, with measurable containment and a short time to ROI. Certainly is built specifically for customer service and conversational commerce, runs on a multi-model architecture so you are not locked to one provider, and ships with the connectors (Zendesk, Shopify, Salesforce, MCP) that most CX stacks already use. Where it stands out is the operating model: pilot in 90 days, transparent cost per conversation, and human handoff that actually preserves context. Where it is not the answer: if you only need an internal IT copilot, this is overkill.
Zendesk AI (Resolution Bot, Advanced AI add-on)
Best fit: teams already deeply embedded in Zendesk who want the AI layer to live inside the ticket. Strong out-of-the-box ticket triage, suggested replies, and macro generation. The weakness is breadth: it is excellent inside Zendesk and limited outside it. If your channels live elsewhere, you end up paying for two platforms.
Intercom Fin (now Salesforce Fin)
Best fit: B2B SaaS support, in-product help, and SMB teams that already run Intercom. Fin set the benchmark in 2023 to 2024 for conversational quality on documentation-heavy domains. The pricing model (per resolution) aligns cost to value when the resolution criteria are transparent, though it can become expensive if your conversation mix includes frequent escalations or edge cases that do not meet the vendor's definition. Update: Salesforce signed a definitive agreement to acquire Fin from Intercom on June 15, 2026. Existing Fin customers should expect the product to be folded into the Salesforce Agentforce roadmap; new buyers should evaluate it as part of the Salesforce/Fin stack rather than as a standalone Intercom add-on.
Ada
Best fit: large consumer brands wanting a no-code build experience and strong reporting. Ada's strength is the operator UX: CX teams without engineering can build and maintain flows. The trade-off is configurability at the edges; deep custom integrations move slower than on more developer-friendly platforms.
Freshworks Freddy AI
Best fit: SMB and mid-market teams already on Freshdesk or Freshchat who want an integrated AI layer at a reasonable price. Solid for the core use cases, less ambitious than the platforms above on agentic actions and multi-channel orchestration.
Salesforce Einstein / Agentforce (now including Fin)
Best fit: enterprises with Salesforce Service Cloud as the system of record who want AI native to that stack. Powerful when the data and workflows already live in Salesforce. The planned addition of Fin's agent architecture should strengthen conversational UX, but the total cost of ownership and deployment timeline still belong in the same conversation as the rest of the Salesforce estate, not the AI line item.
Microsoft Copilot (Dynamics 365 Customer Service)
Best fit: organisations standardised on the Microsoft stack, especially in regulated industries that value the Microsoft compliance posture. Strong agent-assist features, improving customer-facing capability. The fit depends almost entirely on whether Dynamics is already your service platform.
General-purpose assistants: ChatGPT, Claude, Gemini
Best fit: internal copilots, research, drafting, training material, knowledge management, agent enablement. These are excellent tools for the humans on your CX team. They are not customer service products. Treating them as such is the single most common 2025 to 2026 procurement mistake we see.
Lightweight chat tools: Tidio, Crisp, Helpshift, Comm100
Best fit: small businesses, single-channel deployments, and teams whose volume does not yet justify a purpose-built CX platform. Genuinely useful at the low end, and honest about what they are.
Feature Comparison at a Glance
| Platform | Best fit | Channel coverage | Agentic depth | Pricing model | Typical time to pilot | Architecture / lock-in |
|---|---|---|---|---|---|---|
| Certainly | Mid-market/enterprise CX teams that want one agentic platform across web, Meta, SMS, email and voice | Web, Meta, SMS, email, voice | End-to-end resolution with human handoff and analytics | Per conversation | 90 days | Multi-model: switch LLM provider without rebuilding |
| Zendesk AI | Teams already embedded in Zendesk who want AI inside the ticket | Strong inside Zendesk; limited outside the ticket | Ticket triage, suggested replies, macro generation | Seat licence + AI add-on | 30 days | Tied to the Zendesk stack |
| Ada | Large consumer brands wanting no-code build and strong reporting | Yes | No-code flows; deep custom integrations move slower | Custom | 60 to 90 days | Platform-specific |
| Freshworks Freddy AI | SMB and mid-market teams on Freshdesk or Freshchat | Yes | Core use cases; less multi-channel orchestration | Custom | 30 days | Tied to the Freshworks stack |
| Salesforce / Fin | Enterprises with Salesforce Service Cloud as the system of record | Yes via Salesforce | Strong when data and workflows live in Salesforce; Fin adds conversational UX | Custom | 90+ days | Deep Salesforce estate integration |
| Microsoft Copilot | Organisations standardised on the Microsoft stack, especially regulated industries | Yes via Dynamics | Agent-assist is strong; customer-facing capability is improving | Custom | 60 to 90 days | Deep Microsoft stack integration |
| ChatGPT / Claude / Gemini | Internal copilots, research, drafting, training, knowledge management | N/A (not a customer service channel product) | N/A (not a customer service product) | Usage-based token pricing | N/A | Model-dependent |
AI Readiness Score
How ready is your team for AI?
6 quick questions. Get a personalised score and action plan.
Try the AI Readiness Score1000+ agents deployed worldwide · 4.8 on G2
A Note on Pricing Models: Per Conversation vs. Per Resolution
Most buyers compare vendors on feature lists and miss that the pricing model shapes incentives. In 2026, two models dominate: per conversation and per resolution.
Per conversation charges for each customer interaction the AI handles, regardless of outcome. It is predictable: multiply your monthly conversation volume by the rate and you have a forecast. It also removes ambiguity. There is no debate about whether a refund, an order update, or a clarification counted as 'resolved.'
Per resolution charges only when the AI reaches a defined outcome. When the qualification criteria are transparent and shared, this aligns the vendor to real value. The risk is that 'resolution' is not a standardized unit. One vendor may count a ticket closed by the customer as a resolution; another may require the AI to complete a specific action; a third may charge only when no human touches the ticket. If the definition is vendor-controlled, your bill can move independently of your actual service cost.
Resolution-based pricing can also become expensive at scale. A customer service operation with high intent complexity, frequent escalations, or strict compliance checks may find that few interactions meet the vendor's narrow resolution criteria while most still consume model compute, retrieval, and review time. The effective cost per handled conversation can end up higher than a per-conversation rate.
Neither model is universally better. The right choice depends on whether you can audit and control the resolution definition. If your use case has a small, measurable set of end-to-end outcomes (for example, password resets, order cancellations, refund approvals) and you can verify them in your own systems, per resolution is attractive. If your conversation mix is broad, your success criteria are nuanced, or you want a forecastable cost line, per conversation is the more honest contract.
How to Choose Without Getting Sold a Demo
Five questions cut through the marketing.
1. What is your fully-loaded cost per handled conversation, and which pricing model do you use? Ask whether the vendor charges per conversation, per resolution, per seat, or by token usage. Per conversation gives predictability; per resolution aligns cost to outcomes but only if you can audit the resolution definition in your own systems. The right answer is a number that includes model tokens, retrieval, tool calls, and any human review.
2. What containment rate has a real customer hit on a channel mix like ours? Ask for a reference. "Up to 80 percent" is a marketing number. "This brand hit 62 percent on Meta and 47 percent on email in quarter two" is a real one.
3. Can we switch the underlying model? In 2026, model lock-in is a strategic risk. Multi-model platforms protect you from price changes and capability shifts. Single-model platforms do not.
4. What is the handoff to a human actually like? Get a live demo where the bot escalates mid-conversation. If the human agent does not receive the full transcript, intent, and customer history, the platform is not enterprise-ready.
5. What does a 60 to 90 day pilot look like, and what defines success? Any serious vendor will agree to a pilot with a measurable target. Vendors who push annual contracts before proving value are pricing for risk, not outcomes.
What to Do This Quarter
Shortlist three vendors using the categories above. Pick one purpose-built CX platform, one suite-embedded option if you have a clear Salesforce or Microsoft estate, and one lightweight tool as the floor. Run a parallel pilot on a single high-volume intent (returns, order status, password reset). Measure cost per handled conversation, containment, and CSAT against the same baseline. The winner will be obvious inside 60 days, and the decision will be defensible to the board.
See How Certainly Compares on Your Data
If you want a working session that puts your own ticket volume and channel mix against the numbers above, we run those weekly. No slideware, just your data and a working agent.