The rung model
How much of your helpdesk we read, and why rung 1 reads nothing at all on Help Scout and Gorgias. A boundary in the application, not only in the token.
A rung is how much of your helpdesk you have granted us. You choose it when you connect, you can move it afterwards, and it is a boundary in our application as well as in the access token — because a token's scopes are your vendor's setting, not your decision.
| Rung | What it is | What we read | What we write |
|---|---|---|---|
| 0 | Upload | A conversation export you send us ([importing a file](importing.md)). Nothing is connected. | Nothing. |
| 1 | Your AI's conversations | Only the conversations your AI took part in, filtered before anything leaves the helpdesk. | Nothing. |
| 2 | Every conversation | Every conversation, your team's answers included. | Nothing. |
| 3 | Approve and run | Every conversation. | Articles, saved replies, AI content, internal notes, replies to customers — each one only after a person approves the exact text, immediately before it goes. |
A new connection starts at the rung you choose when you connect, 1 or 2, never above what your plan reads. Nothing is written below rung 3, and rung 3 is always a separate choice made after connecting.
Rung 1 is the vendor's own filter, or nothing
This is the part people are surprised by, so it is stated first.
At rung 1 we ask your helpdesk for AI conversations using the filter your helpdesk itself provides — Intercom's ai_agent_participated, Zendesk's support_type_scope=ai_agent. Help Scout and Gorgias have no such filter, and neither does Freshdesk, which reports nothing about its AI at all, so at rung 1 they read nothing at all. We will not "make it work" by downloading everything and filtering afterwards: widening what a rung reads is a promise to you being broken, whatever the token would allow.
If your helpdesk is Help Scout, Gorgias or Freshdesk, rung 2 is where reading begins.
What each rung makes possible
- Rung 1 is the AI audit: what your vendor billed as resolved against what the evidence supports, the patterns behind the failures, and the help center gaps those failures point at.
- Rung 2 adds your team's answers — quality by topic, answers given from memory, and the line between the AI and your team on each topic. Work this out from rung 1 alone and you are guessing about half the inbox.
- Rung 3 lets an approved fix actually be written back. Nothing is written without a person approving the exact words first; rung 3 is permission for that approval to mean something, not permission to act unattended.
How it is enforced
Three things are intersected before any call is made:
- What the vendor's API allows, which the adapter reports per capability.
- The rung you granted, and the scopes the token exchange actually returned.
- Your workspace's status — anything other than active gets nothing.
A write needs all three, and the engine asks again immediately before each step runs, so lowering the rung stops a write that has not been made yet. The step then fails at its preflight, with the reason recorded. That is correct behavior, not an error to route around.
Your plan caps the rung as well, and never chooses it. Every plan reads at rung 2 if you choose it — the free audit and Starter included — so reading every conversation is never what a plan sells. Writing approved fixes back, rung 3, is on Team and above and the trial. Starter reads a connection granted at rung 3 as rung 2: nothing stored is rewritten, and the connection writes again the moment the plan allows it. A connection at rung 1 stays at rung 1 on every plan.
The free audit reads the newest 500
The free audit reads once, so it takes in no more than that one run reads and grades: the newest 500 conversations your rung reads that changed in the last 30 days, and nothing past them is stored. Where they begin is found first, from a list of ids and change times alone — newest first where your helpdesk lists that way (Intercom, Help Scout, Gorgias, Freshdesk), counted in change order where it cannot (Zendesk's incremental export), and at the same rung, so at rung 1 that list is your vendor's AI filter too. The window ends a day before the read, because a conversation still going is never read or graded. What the audit covered is recorded as that period and a count.
Freshdesk: a key, and a rung that is ours to hold
Freshdesk is connected with an API key you copy from your own profile settings and paste into the connect form, not through a consent screen. Two things follow, and the connect page says both before you paste it:
- A key carries every permission of the agent it belongs to, whatever we ask of it. There is no list of permissions to compare, so the rung you choose is held by FunctioningHQ, in the application, on every read and every write, not by the key. Give it a dedicated agent with a narrow role: one that can see every ticket and does nothing else. The key is stored encrypted for your workspace alone, never shown again, and never placed in an address, a log or an audit event; replacing it is the same form.
- Freshdesk reports nothing about its AI. There is no filter for it, so rung 1 reads nothing here; and no field says whether it answered or resolved a ticket, so there is no vendor number to check and none is invented. Grading and repeat-contact checks apply to every ticket we read, and the line your AI signs with (Settings › How we work) is how its answers are told apart from your team's. Chats your AI answers inside Freshchat never become tickets, so they are not read.
Reading goes by when a ticket last changed. A ticket nobody has touched since it arrived is found through the plain ticket list, which covers the last 30 days: on a first read of a longer history, never-touched tickets older than that may be missed. A ticket deleted or marked as spam after we stored it is not noticed, and stays counted. Freshdesk shares its hourly allowance of requests with everything else on your account, so a read stops when a fifth of it is left and carries on the next night. At rung 3, after you approve the exact words, we can leave a private note, reopen a ticket with one, raise a ticket for your team, reply on a customer's own ticket and write a help center article as a draft. We never write a tag, and saved replies are read and not created.
An AI agent from another company
Some teams put an AI agent from another company (eesel, Pluno, Chatbase and the like) on top of their helpdesk. It writes its replies into your tickets, so the helpdesk's own AI filter and resolution field never mark them, and rung 1, which is the helpdesk's filter or nothing, reads nothing for them. We do not guess at them. They are found the way any AI the helpdesk does not report is found, and only at rung 2, where every conversation is already read:
- The line your AI signs with, pasted exactly, matched literally against customer-facing replies. An internal note is never read as an answer.
- Tags that mean the AI handled it, which set who answered and nothing else.
- On setup, a short list of agents to tick. It adds only a tag the agent's own documentation names (today Chatbase's
chatbase-involved), says where that was read, and for the others says plainly that nothing is documented to match.
Either way it applies from the next read and never rewrites what was stored, and it never widens what a rung reads. For an agent from another company there is no resolution claim from a vendor: its answers are checked for whether they were right and supported, billed against held uses the rate you enter, and only a tag you name as meaning your AI counted it resolved becomes a claim, recorded under that tag's name.
Seeing it from outside
Ask the product what it may actually do here and you get the intersection, not the brochure. Over MCP (what_is_possible), or on Settings › Sources:
{
"rows": [
{ "key": "helpdesk.write_article", "name": "Write a help center article",
"state": "available_if", "todo": "Move the helpdesk to rung 3, approve and run" },
{ "key": "helpdesk.note_handoff", "name": "Leave an internal note",
"state": "not_here", "reason": "Gorgias does not let anything outside it add a note." },
{ "key": "manual.send_links", "name": "Links a person sends",
"state": "available" }
]
}
And a step that runs after the rung was lowered fails at its preflight, with the reason on the step rather than a stack trace:
The helpdesk is connected at rung 1, which reads and never writes.
An admin can move it to rung 3, approve and run, under Connections.
That is the correct outcome, not an error to route around.
Moving a rung
Under Settings › Sources. Moving up asks your helpdesk for consent again; moving down takes effect immediately, including for a play that is part way through. Nothing already read is deleted by lowering a rung — retention is a separate window, under Settings › Sources.
Something missing, or a field that does not behave as this says? Tell us — the contracts here are what the code is held to, so a page that is wrong is a bug in the same sense the code would be.