Integrations

Audit Zendesk AI: check every answer, and whether it held.

Connect Zendesk at the level you choose, from only the conversations its AI handled alone to every conversation. Each answer is checked against your help center as it stood that day, each resolution is followed for 7 days, and the patterns behind the failures come with a fix for you to approve.

What we read in Zendesk

You choose how much to grant when you connect, and you can change it afterwards. It is held in our application on every read and write, not only by the access token. The rung model.

  1. 1

    Your AI’s conversations

    At rung 1 we read the tickets Zendesk marks as handled by its AI agent alone, filtered before anything leaves Zendesk. A conversation the AI handed to a person becomes an ordinary ticket, so it is read from rung 2.

  2. 2

    Every conversation, read-only

    Your team’s answers are checked the same way as the AI’s.

  3. 3

    Approve and run

    A fix you approve is written back into Zendesk. A separate choice, never where a connection starts.

From Zendesk, at rung 2 and above
  • Conversations, with every message in order
  • The ratings customers left
  • Your help center articles, with every version, so an answer is checked against the article as it stood that day
  • Your saved replies
  • Company records, so a pattern can say which accounts it reaches
  • Its AI’s own resolution outcome, kept as the vendor’s claim and never as our verdict
  • Which teammate held each conversation, for coaching one person at a time and never for a ranking

Answers for the AI agent are not read or written through Zendesk yet.

This is what Zendesk’s API allows. Your account’s plan may allow less; what is uncertain is checked with a read each when you connect, and the product says which it is reading.

What Zendesk bills for its AI

What it bills
  • Since May 2026, a model reviews each conversation its AI handled alone when it ends and places it in a tier. Assisted escalations and contained resolutions do not count against your allowance, which leaves the verified ones as what is billed.
What we check

We check the resolutions Zendesk counted, and because a ticket does not say which were verified, any money figure says “at most”. Zendesk tags a ticket when its AI resolves it. The tag does not say which tier Zendesk placed it in, or whether the customer confirmed it, so we record it as their claim.

Checked 16 October 2026 against About automated resolution tiers.

Whether a resolution held

A resolution is still settling for 7 days, counted neither way. After that it held unless:

  • The conversation was reopened after your vendor counted it as resolved.
  • The same customer came back about the same thing.
  • A person on your team answered after your vendor counted it as resolved.
  • The customer said it was not fixed.
  • The AI’s answer claimed something nothing in your documentation supports.
  • The AI’s answer contradicted your documentation.
  • Your vendor’s own record says its AI did not resolve this.

This is evidence that these customers did not get the resolution billed. It is not a finding that you are owed a refund; what your vendor counts and credits is set by your contract with them.

What runs in Zendesk, and what a person does

At rung 3, once approved. Where Zendesk’s API does not offer a step, it becomes a person’s step, with the words drafted for them. Nothing reaches a customer without a person approving the exact words.

Runs there, once approved

  • Write the article into your help center
  • Raise a ticket and link the conversations
  • Add the saved reply to your helpdesk
  • Message the customers this affected · reaches customers, approved word for word
  • Reopen the conversation for a person, with the evidence attached
  • Leave a handoff note for the person who inherited it
  • Write to this one customer about their conversation · reaches customers, approved word for word
  • Check the saved reply now exists
  • Ask the people affected whether it worked, through your helpdesk · reaches customers, approved word for word

A person’s step, drafted for them

  • Give the AI agent a precise answer

Zendesk AI reporting, QA tools, and FunctioningHQ

What a team on Zendesk already has, or would look at next. Each claim about another product is from its own public pages, linked, on the date shown. Where they are stronger, we say so.

Built into Zendesk

Zendesk’s automated resolutions

Since May 2026, a model reviews each AI conversation that ended without a person, when it ends, and places it in a tier. Assisted escalations and contained resolutions do not count against your allowance, which leaves verified resolutions as the billed tier.

Where it is stronger
  • It runs inside Zendesk on every conversation its AI closed alone, and decides what you are billed.
  • Nothing to connect.
Where we differ
  • What we read from a ticket does not say which resolutions Zendesk verified, so we keep its resolution as Zendesk’s claim, and any money figure says “at most”.

Checked 16 October 2026 against About automated resolution tiers.

Zendesk add-on

Zendesk QA

Quality assurance inside Zendesk: AutoQA scores every conversation, your team’s and your AI agents’, against ready-made and custom categories, with one-to-one coaching built on the results.

Where it is stronger
  • A mature QA product: scorecards, categories you write yourself, and coaching across the whole team.
  • Native to Zendesk, and it reviews AI agents alongside people.
Where we differ
  • An answer is checked against the article it should rest on, as it stood that day, rather than scored on categories.

Checked 16 October 2026 against Zendesk QA, Evaluating AI agents in Zendesk QA.

QA tool

Rippit (formerly MaestroQA)

Quality assurance for support teams: every eligible conversation scored against your rubric, with coaching built on the results.

Where it is stronger
  • A full QA program: your own rubric, every eligible conversation scored, and coaching.
  • Lists integrations for both Zendesk and Intercom.
Where we differ
  • An answer is checked against the article it should rest on, as it stood that day, rather than scored on a rubric.

Checked 16 October 2026 against Rippit: QA every support conversation, Rippit integrations.

By hand

QA in a spreadsheet

A person reads a sample of conversations each week and scores them against a rubric the team wrote.

Where it is stronger
  • Free, and the rubric is entirely yours.
  • A person reads every conversation that is scored, with all the context they carry.
Where we differ
  • Every conversation you grant is read, up to your plan’s allowance, not a sample of a few a week.
  • Each answer is checked against your help center as it stood the day it was given, and a verdict shows the passage it rests on, checked against the article.
What this adds, whichever you use
  • Each answer is checked against your help center as it stood the day it was given, and a verdict shows the passage it rests on, checked against the article.
  • Each resolution, or where the helpdesk reports none each conversation, is followed for 7 days: whether the customer came back about the same thing, said it was not fixed, or got an answer we found wrong.
  • Failures are grouped into patterns and counted in people, not conversations.
  • A fix is drafted for each pattern, a person approves it, and it is measured before and after on the people it was for.
  • Nothing scores or ranks people. Coaching speaks about one person at a time, to the manager who may see it.

Questions about Zendesk

Does FunctioningHQ work with Zendesk?
Yes. You connect Zendesk by signing in, and choose how much to grant. At rung 1 we read the tickets Zendesk marks as handled by its AI agent alone, filtered before anything leaves Zendesk. A conversation the AI handed to a person becomes an ordinary ticket, so it is read from rung 2. Nothing is written into Zendesk until you choose to allow it, and then only what a person approves.
Will it change anything in Zendesk?
Not unless you allow it. A connection starts read-only. Writing a fix back is a separate choice, and nothing reaches a customer without a person approving the exact words.
How is this different from Zendesk AI’s own reports?
Zendesk’s reports are its own count of what its AI resolved, and its resolutions are part of what it bills. We keep that count as Zendesk’s claim and check each resolution against what happened next, and each answer against your help center.
Can I try it for free?
Yes. Up to 500 of your recent conversations, read once and checked, ending on what to fix first. No card.

Start with a free audit of your Zendesk.

Up to 500 of your recent conversations, read once and checked, ending on what to fix first. No card.

Request beta access