On 18 May 2026, Zendesk changed what the word “resolved” means. It split automated outcomes into three buckets — contained resolutions, assisted escalations, and verified resolutions — and made only the third one billable. A verified resolution is a conversation an LLM evaluation confirmed was actually satisfied, assessed at the end of the session window, which defaults to 72 hours after the first message. Before that change, a customer who asked a question and then went quiet for three days counted as a resolution and drew from your allowance.
Nothing about the agents got better that week. The meter did. That is the problem this stack is built around: in every AI support deployment, the number that reaches your board deck is produced by the vendor being paid for it, and each vendor defines the billable unit differently. This stack is a composition where the resolution claim is checkable, the escalation path is not owned by the agent vendor, and you can see what the customer did after the conversation ended.
The shape
- Zendesk or Pylon is the system of record. Every conversation the agent touches lands as a ticket with the full thread attached, because the audit you will eventually run is a ticket query, not a vendor dashboard export. Zendesk is the default for consumer and mixed-channel volume. Pylon is the pick when your support actually happens in shared Slack and Teams channels with named accounts — a B2B motion where the “ticket” is a thread in the customer’s own workspace. Pylon withdrew its published seat pricing to a demo gate, so treat it as a quote-only line.
- Fin (Intercom) or Decagon is the resolution layer. One agent, resolving end to end and taking the backend action rather than answering from an article. Fin is metered at $0.99 per outcome with a 50-outcome monthly minimum; Decagon publishes no price and is reported to start around a $50,000 annual platform fee before usage. The full head-to-head is in Decagon vs Fin.
- Notion is the knowledge base, and it is a component rather than a document. Fin connects Notion as a native content source, resyncs every 24 hours, and refuses to let you edit that content inside Fin — changes must be made in Notion. That constraint is the feature. It forces one owner and one place where a refund policy changes, which is the only reason the agent’s answers stay consistent with what your humans say.
- Pendo measures what the vendor meter structurally cannot. Agent Analytics instruments the conversational surface inside your own product and connects it to behavioral analytics and session replay, reporting prompt volume, retention, suggested-prompt rate, and where conversations break down. Pendo reports measuring more than 350 agents and 2.5 million prompts a week. A vendor meter tells you a conversation ended; Pendo tells you whether the same person came back and filed a ticket 20 minutes later.
- n8n owns the escalation route and the second log. Every escalation fires a workflow that writes its own record — timestamp, intent, the agent’s confidence, the reason for handoff — into a store the agent vendor does not control. This is the cheapest layer in the stack and the one that makes a billing dispute winnable.
Named handoffs
- Policy changes in Notion → the agent’s answers change within 24 hours. One page, one owner, one resync. No parallel copy in the agent’s own editor to drift from.
- Customer opens a conversation → the agent attempts resolution → the SoR gets a ticket either way. Contained, escalated, or resolved, the thread exists in Zendesk or Pylon. An agent that resolves conversations invisibly is an agent you cannot audit.
- Agent hands off → n8n routes and logs. The workflow assigns the queue, attaches the handoff payload, and writes the independent escalation record before a human sees it.
- Session window closes → the vendor marks the outcome → n8n reconciles. Once a month, count billed resolutions against escalations logged and tickets reopened within seven days. A resolution that produced a reopen is a billing question, not a success.
- In-product agent conversation ends → Pendo carries the journey forward into retention and adoption data, which is where the honest deflection number finally shows up.
The fork that decides the cost
Both agents run on Zendesk, so this is not a migration decision. It is a decision about which meter you buy, and there is a gate worth knowing before you price anything.
Fin for Zendesk messaging requires the Sunshine API, which is included only in Zendesk Suite Professional, Enterprise, and Enterprise Plus. Suite Professional is $115 per agent per month paid yearly against Suite Team at $55. If your Fin deployment is email and contact-form tickets only, Fin can be assigned as an agent on Zendesk tickets without that gate. The moment you want it on web, mobile, or WhatsApp messaging, your helpdesk bill roughly doubles. Price that before you price the agent.
The second rule is simpler: run one meter. Zendesk bills $1.50 per committed automated resolution and $2.00 pay-as-you-go, with allowances of 5, 10, or 15 ARs per agent per month depending on plan. Fin bills $0.99 per outcome. If Fin does the resolving, your Zendesk AR allowance sits unused and you have paid for a meter you are not reading. Pick the agent, then buy the plan that agent needs — not both.
Cost reality
A B2B SaaS support org with 12 seats, 6,000 conversations a month, and a realistic 40% resolution rate — 2,400 resolutions a month:
- Zendesk-native agent: Suite Professional at $115 × 12 seats = $16,560/year. Allowance is 10 ARs per agent per month, so 120 covered and 2,280 billed at $1.50 = $41,040/year. Total ≈ $57,600/year.
- Fin on Zendesk: the same $16,560 helpdesk floor if you need messaging, plus 2,400 × $0.99 = $28,512/year. Total ≈ $45,000/year, and the Zendesk allowance goes unused.
- Decagon: the reported ~$50,000 platform fee lands before a single resolution, so it competes on unit price only once volume is well past this example.
- n8n Pro: €60/month for 10,000 executions, roughly €720/year, which is a rounding error against the agent line.
- Pendo: MAU-priced and quote-only above the free tier. Get the number before you assume the measurement layer is free.
The unpriced line is the knowledge base. Someone has to own it. Budget a named part-time owner; a stale KB is the single most reliable way to convert a $0.99 outcome into a reopened ticket you also pay a human to handle.
Variations and when to swap
- Pylon instead of Zendesk when more than about half your volume arrives in shared Slack or Teams channels. Do not swap for a consumer-facing web and mobile motion — you would be buying an account-centric model for anonymous traffic.
- Drop Pendo if your agent lives only in a web widget and email, with no in-product surface. There is no journey for it to instrument, and the vendor meter plus the n8n log is enough. Keep it the moment the agent is embedded in your application.
- Swap n8n for native helpdesk triggers at low escalation volume. The rule: keep n8n once you need the log to live somewhere the agent vendor cannot rewrite, or once routing depends on data from a third system.
What this stack does not replace
- It is not a support strategy. It measures and routes; it does not decide which intents an agent should be allowed to resolve. That scoping work is the ticket-deflection agent template.
- It is not a customer success stack. Resolution is not retention — see the customer retention stack.
- It does not write your knowledge base. Every layer here assumes content that is already correct.
- It does not replace human escalation capacity. If the agent resolves 40%, your humans still handle 60%, and they now get the harder 60%.
Watch-outs, each with a guard
- Every vendor’s billable unit is different, and the definition moves. Zendesk’s changed in May 2026; Decagon’s is negotiated per contract. Guard: get the resolution definition in writing before signing, and re-read it at renewal.
- Containment is not resolution. A customer who gives up is contained. Guard: track reopens within seven days and CSAT on agent-handled conversations as the paired metric — see CSAT.
- The 72-hour session window delays your feedback loop. A bad prompt change takes three days to show up in the meter. Guard: watch the n8n escalation rate, which moves within the hour, as your leading indicator.
- A resync gap makes the agent confidently wrong. Notion resyncs every 24 hours. Guard: force a manual resync on any policy change that affects money or entitlements, and hold the change until it lands.
- Escalations that arrive without context are worse than no agent. Guard: make the handoff payload a schema n8n validates, and treat a failed validation as a routing failure. When one goes badly, run the escalation RCA.
Match rules
Right pick when: you run 2,000 conversations a month or more, a real share of them are repeatable tier-1 intents, and someone is going to be asked to defend the deflection number — a board, a CFO, or a renewal conversation. It fits best at 10 to 50 support seats, where the volume justifies the meter and there is no analytics team to build one for you.
Wrong pick when: you are under a few hundred tickets a month, where a help center and a human clear the queue and the agent adds policy maintenance that never pays back. Also wrong when your volume is overwhelmingly bespoke technical investigation rather than repeatable questions, and when nobody can be named as the knowledge-base owner — that gap breaks the stack no matter which agent you buy.
If you can only do one thing: stand up the n8n escalation log before you turn the agent on. It costs €60 a month, it takes an afternoon, and it is the only record in this stack that your agent vendor did not produce.