AI in customer service: the 2026 buyer's guide

2026 is the first year where AI is a baseline expectation for customer-support tooling rather than a premium add-on. Every help-desk vendor now ships some version of an AI Copilot for agents and an AI Bot for customers. The pricing model and the depth of grounding are where they differ — and where most buyers get burned.

The two pricing models that matter

There are two fundamentally different ways AI gets billed in this category, and they produce wildly different invoices at scale.

Per-resolution / per-answer / per-session

This is the legacy SaaS approach. The vendor charges you each time the AI hits a hard "resolved" or "answered" outcome. Zendesk publishes per-resolution pricing of $1.50 to $2.00. Intercom Fin is $0.99 per resolution. Freshdesk Freddy AI Agent is $0.10 per session. Front charges per AI Answer.

The math gets ugly fast. A team handling 5,000 AI conversations a month at $1.50 each pays $7,500 — just for the AI line item, on top of the seat licence. Worse, the bill is unpredictable: a viral spike doubles your AI cost the same month it doubles your CX load.

Per-conversation pass-through

The newer model meters AI by total conversation volume regardless of outcome, at near-cost rates. Deskwoot charges $0.01 to $0.03 per conversation. The same 5,000 conversations cost $50 to $150 in AI fees, two orders of magnitude lower than the per-resolution alternatives. Bills become predictable; finance can model them like any other usage-based SaaS expense.

Why grounding is the make-or-break feature

An AI Copilot is only as good as the context it has. A vendor that simply pipes your conversation to GPT or Claude with no grounding will produce confident-sounding answers that miss your specific terminology, product, pricing, or policy.

The vendors that work in 2026 use what's called retrieval-augmented generation — the AI reads your help center articles, training documents, prior tickets, and product wiki before drafting a reply. The output reads like your team wrote it because the model's context window contains your team's actual writing.

When you evaluate vendors, run the same test conversation through each Copilot and compare. The grounded ones produce answers your team can ship with one edit. The ungrounded ones produce paragraphs that need rewriting.

Five questions to ask every vendor

• What's the AI billed per? Per resolution, per session, or per conversation? Get it in writing before signing.
• What's the model? If it's "our proprietary AI," ask what foundation model it's built on. Most vendors are reselling Claude, GPT-4o, or Gemini under a wrapper.
• How does grounding work? Knowledge base only, or also training documents, prior tickets, and your product wiki? The breadth matters.
• What happens to deflected conversations? Does the AI Bot hand off cleanly to a human, or do customers get stuck? Test the failure path.
• What's the prompt-injection posture? A 2026-vintage AI Bot needs guardrails against malicious customer input. Ask for the policy and run an attack test.

What to actually compare on a side-by-side

Run a 7-day pilot with two vendors. Use the same 100 real customer conversations through each. Measure: time to first draft, edit distance from draft to sent reply, deflection rate (AI Bot only), customer CSAT on AI-handled conversations, and total AI cost incurred. The vendor that wins on three of those five wins the bake-off.

The one trap to avoid

Don't sign a multi-year contract priced on resolutions until you've benchmarked the actual resolution count for your traffic. Vendors often quote based on a usage estimate that turns out to be 3-4x lower than reality once you're live. The CFO conversation is much easier on a per-conversation pass-through plan because the worst case is bounded by your total volume, which is forecast-able.

Where Deskwoot fits

Deskwoot ships the per-conversation model — $0.01 to $0.03 per conversation as pass-through, AI Copilot included in every paid plan, AI Bot Fynn grounded in your help center plus the Training Hub (PDFs, FAQ pairs, website crawls). The free trial includes the full AI stack so you can run the bake-off before deciding.