The Credit balance on Settings → Billing is lower this morning than it was last night, and nobody on the team chatted with Insulin in between. The meter is working as designed. Fours’ metering documentation explains why: “Billed AI is not limited to your chats and agents. These also run for your organization, and each is billed under its own name on the Agent tab of the usage breakdown.”
This post starts from that symptom. It lists what spends AI credit without anyone chatting, where each shows in the Usage panel, what to change, and who can change it. The last column matters: an organization admin can see all of the spend but cannot reach all of its causes.
Three things about the panel itself, before you read it. The balance moves within about a minute of a request, but the breakdown is re-aggregated every couple of hours, so last night’s spend may not be in the table yet. Every grouping totals the same amount, and a filtered total is only a subset. And there is no CSV export or per-request drill-down. Reading Insulin’s AI usage by person, app and model covers each of these in full.
What spends AI credit when nobody is chatting?
Six kinds of work, and the Agent tab names most of them. Open the Usage panel on Settings → Billing, choose the month, and work down this table in order: it starts with the tab that names the most causes.
| What runs without anyone chatting | Where it shows | What to change | Who can change it |
|---|---|---|---|
| Knowledge-base connectors with Auto-sync, and every search of a knowledge base | Agent tab: Knowledge base | Narrow the connector with Edit scope, or Remove it | The knowledge base’s owner, or someone shared on it as Editor or above. Org admins cannot open one they were not shared on |
| Inbox rules that fire on new mail, a deal changing stage, or a Fours event | Agent tab: Inbox | Switch the rule off, or Disable Inbox | The person whose Inbox it is |
| AI features in the rest of the Fours console | Agent tab: each feature by its own name | Take the row to the team that uses the feature | That team. Co-sell AI is switched on by Fours |
| Cron and push jobs, and goals | Agent tab: the agent the job runs as. User tab: No user — scheduled and system runs, for a scheduled job | Switch the job to Disabled, or Cancel goal | Only the job’s owner. Org admins cannot see another member’s jobs |
| An Insulin app’s own model calls | App tab: the app’s name | Ask whether the app needs that model call | The app’s owner, or anyone with Editor or above on it |
| A stopped sandbox, and the month’s storage peak | Resource tab: Sandbox storage and Object storage | Delete the sandbox rather than stop it, and delete files nobody needs | The member who owns them. The organization’s sandbox pool is for org admins |
Two helpers on the Agent tab follow chat rather than quiet days. Thread titles names a conversation and Skill drafting writes a skill with Insulin, so both rise and fall with chat and are not the cause here.
Knowledge base: synced connectors and every search
A knowledge base spends credit each time it embeds a document and each time it is searched, and the Knowledge base row on the Agent tab carries both. The knowledge-base documentation is direct: each embedding request a knowledge base makes, “indexing a document, and running a search”, is charged to your organization’s AI credit.
Three kinds of knowledge-base work happen with nobody in a conversation:
- Connectors with Auto-sync on. Insulin sweeps due connectors about once an hour. It runs an incremental sync when the last sync was more than an hour ago, and a full reconcile when the last full sync was more than 24 hours ago. Documents a sync adds are split into chunks and embedded, and that is billed.
- Searches from automated work. An agent in a job can search the knowledge bases attached to it, and each Inbox draft written for review searches up to three of the knowledge bases attached to Inbox. Every one of those searches is an embedding request.
- An embedding-model change. Switching a knowledge base’s model re-embeds every document, one at a time, and the confirmation dialog says so: it “takes provider time and spend.”
Websites and individual files never sync on their own. A website re-crawls only when someone clicks Re-sync, and an uploaded file is indexed when it is added. If a knowledge base holds only files and websites, it does not spend on a schedule.
What to change. On the connector’s card, use Edit scope to cover fewer repositories, spaces, folders or files, or Remove a connector nobody needs current.
Who can change it. This is where an admin runs out of reach:
- GitHub and ClickUp connectors exist only on personal knowledge bases, and those are private to their creator: “even organization admins cannot access them unless the knowledge base is shared with them.”
- An organization knowledge base supports only the Google Drive connector, and even an org admin cannot open one they were not shared on.
- The Knowledge base row is one figure for the whole organization. The panel has no per-knowledge-base breakdown, so it cannot tell you which knowledge base spent it or whose it is.
So the fix is a message to the people who connected GitHub, ClickUp or Google Drive, asking them to check the scope of what they sync.
Inbox: rules that fire on mail and events
An Inbox rule that drafts while its owner is away spends credit on each draft, and Inbox’s AI usage has its own Inbox row on the Agent tab. A rule can fire on a new email, on a deal changing stage in Salesforce, HubSpot or Dynamics 365, on a Fours marketplace event, or on a partner program event. None of those needs anyone to be online.
Inbox ships no rules of its own, so this spend always traces back to a rule someone wrote. A rule with pinned canned text involves no model at all; a rule that writes a reply does. A draft waiting unread in Approvals ▸ Pending has already been written, so its tokens are already spent. If Inbox has knowledge bases attached, each draft’s searches land on the Knowledge base row as well.
What to change. In Inbox’s Settings → Rules, every rule row carries an enable switch, so a rule can be turned off without opening it. Disable on the mailbox card pauses Inbox entirely: it stops receiving and processing new email, and keeps what it has already processed.
Who can change it. The person whose Inbox it is. The mailbox and its rules are theirs, and the Inbox row does not say whose rules fired, so ask the people who use Inbox.
AI in the rest of the Fours console
Co-sell, partner programs, workflows and document extraction in the Fours console are not Insulin chat, but the AI behind them spends the same credit. Each feature has its own row on the Agent tab, under the name the metering page gives it:
- Co-sell quality score, Co-sell referral AI fix, Co-sell business problem, Co-sell engagement score and Co-sell contact score
- Marketplace insights, AI field mapping, SCA extraction and AWS invoice extraction
- Workflow AI email, Workflow AI Slack message, Workflow AI Agent node and Workflow AI test event
- PRM and Deal registration
One of these runs on a timer. The co-sell engagement score is refreshed every 6 hours. Fours takes it from the cloud partner where the partner reports one, then from the customer’s marketplace activity, and only when neither is available does it use an AI estimate, which “uses the AI credit of the organization that asks for it.”
What to change, and who. Whether a feature’s AI is worth its credit is for the team that uses it to judge, so take its row to them. Co-sell AI is switched on by Fours, not by your team, so a change there goes through Fours support.
Jobs and goals: work on a schedule or an event
A cron job fires on its schedule and a push job on every event that passes its filter, whether anyone is online or not, and a goal keeps working in rounds until its objective is met or someone stops it. A job runs as the agent that created it, so its spend shows on the Agent tab under that agent. A scheduled job has no user, so on the User tab its spend sits in No user — scheduled and system runs, not under the person who set it up.
A push job with no filter runs its agent on every event that arrives. Adding a filter means events that don’t match are recorded as Filtered, and the agent doesn’t run.
What to change. The owner opens the job and switches it to Disabled: a disabled job never fires on its own, and nothing is deleted. To change a schedule or add a filter, the owner asks the assistant, because the Jobs app has no edit form. A goal stops with Cancel goal in its details drawer, or by writing “Stop.” in its conversation.
Who can change it. Only the job’s owner. The jobs documentation says a job is private to whoever created it: “only its owner can see it, run it, edit it, or delete it. That includes organization administrators.” The exception is work Planner schedules in an organization channel, which belongs to the channel. On the Agent tab, a colleague’s personal agent shows its ID in place of a name, with a dash for Type and Level, and its spend still counts.
What an admin can do from the panel is narrow it down. Type that ID into the Agents filter: the table then shows only that agent’s spend, and the daily spend chart follows the filter, so you can see which days it spent on. The panel does not say who owns an agent shown by its ID, so the last step is to ask your team.
Apps: the model calls an app makes
An Insulin app that calls a model at runtime spends credit each time it does, and the App tab shows that spend under the app’s name. The metering documentation says those tokens are “metered and priced the same way a chat turn’s are.” Whoever uses the app is not chatting with Insulin, but the app’s code is calling a model on your organization’s credit.
What to change, and who. Ask the app’s owner whether the app needs a model for that work. If you can use the app, its card in the console’s Insulin Apps list has an info button that names the owner. Changing the app takes Editor or higher on it.
Sandboxes and storage: what stopping does not end
Storage keeps billing with nobody working: a stopped sandbox bills Sandbox storage for as long as it exists, and Object storage bills the month’s highest daily reading. Both are rows on the Resource tab.
- A stopped sandbox still costs storage. Stopping ends processor and memory, but deleting the sandbox is what stops the storage charge. The organization’s sandbox pool shows an Auto-stop time for each sandbox, which shuts an idle one down. That ends its processor and memory charges, not its storage.
- Object storage is the month’s peak. It counts the organization’s shared storage and each member’s own together. Deleting files lowers next month’s peak, not this month’s.
Who can change it. Each member manages their own sandboxes under Settings → Personal → AI Sandbox, and an org admin can reach them on that member’s behalf. The organization’s shared pool, which organization agents work in, is under Settings → Organization → AI Sandbox, for org admins only, where Destroy removes a sandbox that is not running.
What does the monthly spending limit stop?
All AI, the background work you want to keep included, and it is the only spend cap your organization sets. Reaching the monthly spending limit pauses AI requests for the rest of the month, on every path that calls a model for your organization:
- a job run that is refused ends Failed, with the reason as its error
- a goal that is refused stops, and stays stopped until someone starts it again
- a document whose indexing is refused fails, and is not retried on its own
- the co-sell engagement score’s AI estimate is left for a later refresh
Three more things to know before you set one:
- It does not stop storage. Sandbox storage, a running sandbox and object storage keep accruing after AI requests stop.
- It stops automatic top-up for the rest of the month as well.
- There is no limit that only warns. The billing FAQ asks “Can we set a limit that warns without blocking?” and answers “No.”
The control that keeps background work running is automatic top-up. It charges a saved payment method once the balance is at or below a trigger you choose, it is checked every five minutes, and the top-up amount must be larger than the trigger. If AI has already stopped, the banner on Settings → Billing names which of the four reasons Insulin pauses AI applies, and each has a different fix.
Your own provider keys do not make this spend free. Your provider bills you for the tokens, Fours charges a platform fee of 10% of the supplier’s price for them, and “a zero balance still pauses BYOK requests.”
How far does the free credit go?
Every organization is granted free credit on the 1st of each month: $100 for each month through November 2026, and $50 from the December 2026 grant on. Some organizations have a different recurring amount arranged with Fours and keep it; the Credit table at the bottom of Settings → Billing always shows the real figure for yours.
Credit is spent soonest-expiry first, so background work draws on the monthly free credit before any credit you paid for, and unused free credit expires at the end of its month. A month with steady background work uses up more of a $50 grant than it did of a $100 one. How Insulin’s usage-based pricing works sets out what is metered and the controls around it.
What the Usage panel can’t tell you
Which document, rule or request spent the credit, and who owns the work an admin cannot see. Besides the missing per-request detail, three limits shape every investigation:
- No per-knowledge-base or per-mailbox split. The Knowledge base and Inbox rows are one figure each.
- No owner for an agent shown by its ID. The Agents filter isolates its spend; it does not name the person.
- No admin view of jobs or personal knowledge bases. Both are private to their creators, org admins included.
So the panel tells you which kind of work is spending the credit, and the last step is to ask the people who own it: whoever synced a knowledge base, wrote an Inbox rule, set up a job or maintains an app.
Sources
Primary sources for the platform rules cited above. Last verified November 18, 2026. Providers change their products, prices and terms without notice — check the source before relying on a detail.
- Insulin Billing: What Is Metered — Fours Doc — Billed AI not limited to chats and agents: the helpers in Insulin (Knowledge base, Thread titles, AI text generation, Skill drafting, Inbox, PRM, Deal registration) and the AI features elsewhere in the Fours console (the five co-sell rows, Marketplace insights, AI field mapping, SCA extraction, AWS invoice extraction and the four workflow AI steps), each billed under its own name on the Agent tab, quoted and listed as the page words them; only AI Fours runs on its own behalf not billed; an app's runtime model calls metered like a chat turn; sandbox storage accruing while the sandbox exists and deleting it stopping the charge; object storage billed on the month's highest daily reading across shared organization storage and each member's own; BYOK tokens billed by the provider with a platform fee of 10% of the supplier's price, and a zero balance still pausing BYOK requests; the balance moving within about a minute and the breakdown re-aggregated every couple of hours
- Insulin Billing: Reading Your Usage — Fours Doc — The Usage panel on Settings → Billing and its six groupings (Resource, User, App, Model, Provider, Agent), every pivot totalling the same amount; Spend by agent, its rows naming the Inbox, Knowledge base and the console's AI features, and a colleague's personal agent or a deleted one shown by its ID with a dash for Type and Level; a scheduled job having no user, and the row No user — scheduled and system runs; the Agents filter finding an agent by a typed ID, a filtered total being a subset, and the daily spend chart following the filters; no CSV export, no per-request drill-down, and support inside the 90-day retention window
- Insulin Billing: How Billing Works — Fours Doc — Free credit granted on the 1st of each month, $100 for each month through November 2026 and $50 from the December 2026 grant on, a different recurring amount arranged with Fours kept, and the Credit table showing the real figure; credit spent soonest-expiry first and free credit expiring at month end; storage and an already-running sandbox accruing after AI requests stop; the helpers and console AI features billed under their own names on the Agent tab
- Insulin Billing: Payment, Limits, and Top-Ups — Fours Doc — The monthly spending limit pausing AI requests for the rest of the month and stopping automatic top-up, from 0 to $100,000, on both billing models; automatic top-up needing a saved payment method, a top-up larger than its trigger, and being checked every five minutes; the debt limit and suspension set by Fours
- Insulin Billing: When AI Pauses — Fours Doc — The billing check on every path that calls a model for the organization, background jobs and goals, the helpers and the console's AI features included; a refused job run ending Failed, a refused goal stopping and staying stopped, a refused document failing and not retried on its own; the four banners
- Insulin Billing FAQ — Fours Doc — No limit that warns without blocking, and automatic top-up as the way to avoid reaching zero; deleting files lowering next month's storage peak, not this month's; a stopped sandbox still charging storage until it is deleted
- Insulin Knowledge Bases — Fours Doc — Each embedding request (indexing a document, running a search) charged to the organization's AI credit and shown as Knowledge base on the Agent tab; only connectors with Auto-sync kept current, swept about once an hour with an incremental sync and a full reconcile after 24 hours, websites and individual files never syncing on their own; the connected card's Re-sync, Stop sync, Edit scope and Remove; Google Drive the only connector on organization knowledge bases, GitHub and ClickUp on personal ones; an embedding-model change re-embedding every document; roles, personal knowledge bases private to their creator and an org admin unable to open an org knowledge base not shared with them; up to 3 knowledge bases searched per Inbox draft
- Insulin Inbox — Fours Doc — Rules triggered by new email, a CRM stage change, a Fours marketplace event or a partner program event; no built-in rules; an enable switch on every rule row and Disable on the mailbox card; a rule with pinned canned text involving no model; Inbox's AI usage on its own Inbox row of the Agent tab
- Insulin Jobs — Fours Doc — Manual, Cron and Push triggers; a job running as the agent that created it; only its owner able to see, run, edit or delete it, organization administrators included; a push job with no filter running on every event; the Enable / Disable switch on cron and push jobs; changes through the assistant, with no edit form
- Insulin Agents: Long-Running Work — Fours Doc — A goal pursued across rounds, the goal card's details drawer with its Cancel goal button, and Stop or Cancel written in the conversation ending it
- Insulin Custom Apps — Fours Doc — Edit App for Editor or above on the app; the info button on an app card in the console's Insulin Apps list showing the app's owner
- Getting Started with Insulin — Fours Doc — Settings → Organization → AI Sandbox, org-admin only, with Auto-stop and Destroy on a sandbox that is not running; Personal → AI Sandbox for a member's own; sandboxes self-owned, an org admin able to reach others on their behalf but not jobs
- Co-Sell Insights — Fours Doc — The Cosell-AI feature switched on by Fours, not by your team; the engagement score resolved from the partner, then marketplace activity, then an AI estimate as the fallback on the asking organization's AI credit, left for a later refresh while AI is paused, and refreshed every 6 hours
Keep reading
Stay Updated
New posts, product updates and marketplace strategy are shared on LinkedIn as they publish.
Follow Fours on LinkedIn