For administrators

Watch Usage and Control Spending

Read this period's total spend and the per-Agent breakdown in「用量與費用」(Usage & Cost), set a monthly cap on each Agent, and decide whether to block or allow once it is passed.

What you get
You will know how much you spent this month and which assistant spent it, and every assistant will have a monthly ceiling you set yourself.

AI costs are charged by usage, so without a cap there is no floor. This chapter teaches you to read the ledger and fit each Agent with a brake.

Before you start

  • You are signed in as an administrator. 「用量與費用」(Usage & Cost) is administrator-only; an employee who opens the address directly sees only「需要管理員權限」(Administrator access required).
  • Some Agent has already been used. In a brand-new environment every number is 0, which is normal.
  • This chapter covers the numbers and settings you can see on screen. For actual billing terms and payment methods, see your service agreement.

Step 1 · Read this period’s overview

Click「用量與費用」(Usage & Cost) on the left; the title reads「用量與費用 Usage & Cost」.

Two markers in the top-right corner tell you the scope of this ledger:「📅 本期 YYYY-MM」(this period) and「💵 USD」. Statistics are counted by month and start over each month.

Four cards below:

Card What it means How to use it
「總費用 Total Cost」 The total cost of every Agent this period The number to watch most. Compare it to your monthly budget.
「總 Token 用量」(Total token usage) The total used this period, split into「輸入」(input) and「輸出」(output) below Output is usually more expensive than input. Unusually high output normally means some Agent’s answers are too long.
「對話次數 Requests」 How many times it was used this period Use this to read adoption — a low count means staff have not picked it up.
「平均每次成本」(Average cost per use) Total cost ÷ requests Use this to read efficiency. A sudden rise usually means some Agent has too much data attached or a persona that is too long.

Read「平均每次成本」(Average cost per use) before「總費用」(Total Cost). A rising total has two possible causes: more people using it (good) or each use costing more (needs handling). The average tells the two apart at a glance.

Step 2 · Find out who spends the most

Below that is「各 Agent 明細 Per-Agent Breakdown」, one row per Agent, with columns「Agent」,「Token」,「對話次數」(Requests), and「費用」(Cost).

Click any row to expand the detail:

  • 「佔總費用」(Share of total cost) — this Agent’s percentage of the period’s total spend, with a bar chart.
  • 「輸入 Token」(Input tokens) and「輸出 Token」(Output tokens)
  • 「對話次數」(Requests) and「平均每次成本」(Average cost per use)
  • 「使用模型」(Model used) — the model this Agent actually runs on.
  • 「最近活動」(Last activity) — the last time it was used.

Three practical readings:

  • One Agent accounts for more than half: not necessarily a bad thing, but worth confirming it is not being treated as a general-purpose assistant.
  • Few requests but high cost: open that Agent’s「基礎」(Basics) tab and check whether the System Prompt is too long, and its「資料來源」(Data Sources) tab for too many knowledge bases attached.
  • 「最近活動」(Last activity) stuck a month ago: nobody is using it. Consider unpublishing it and saving the maintenance.

Step 3 · Set a monthly cap

Caps live in each Agent’s own settings, not in one platform-wide switch.

Go to「Agent 管理」(Agent Management) → open that Agent →「用量 / 稽核」(Usage / Audit) tab →「用量上限 Usage Policy」block, subtitled「每月重置」(resets monthly). Three fields:

Field Notes
「每月呼叫上限(空=無限制)」(Monthly call cap; empty = unlimited) How many times this Agent may be used in a month
「每月 Token 上限(空=無限制)」(Monthly token cap; empty = unlimited) Another way of counting usage
「超過上限時」(When the cap is exceeded) 「block — 直接擋下」(block outright) or「warn — 放行但記錄」(allow but record)

Setting one of the two caps is enough; fill both and whichever is hit first applies. For a first attempt, set only「每月呼叫上限」(Monthly call cap) — it is more intuitive.

How to pick a number: run one month without a cap (but check it daily), then take the actual request count times 1.5. Or read that Agent’s current monthly usage from「用量與費用」(Usage & Cost) and add fifty percent.

Click「儲存設定」(Save settings).

This field is required in order to publish. An Agent with no usage cap is blocked when you click「🚀 發布」(Publish), with the message「尚未設定用量上限(usage policy)」(No usage cap set). That is deliberate — no assistant goes live without a budget ceiling.

Step 4 · Decide what happens once it is passed

The difference between the「超過上限時」(When the cap is exceeded) options is a practical one:

「block — 直接擋下」(block outright)

  • The next employee to ask is stopped and sees a message that the usage cap has been reached.
  • 「稽核紀錄」(Audit Log) records a「用量攔截」(Usage blocked) entry.
  • Suits: customer-facing services, departments with hard budgets, and new Agents still under observation.

「warn — 放行但記錄」(allow but record)

  • Employees carry on as usual, uninterrupted.
  • The platform records it all the same, and you can see it in「稽核紀錄」(Audit Log).
  • Suits: critical assistants that are running smoothly, such as an IT help desk — better to overspend than to go dark during working hours.

Caps reset automatically every month, with nothing for you to zero out by hand. A blocked Agent recovers on its own on the first day of the next month.

To let it through right away, go back to the「用量 / 稽核」(Usage / Audit) tab, raise the cap, and save; it takes effect immediately, with no waiting for next month.

Step 5 · Check one Agent’s detail

To see one Agent’s own ledger, you can expand its row in「用量與費用」(Usage & Cost), or open that Agent’s「使用紀錄」(Usage Records) tab.

The two numbers at the top are「本月呼叫」(Calls this month) and「本月 Token」(Tokens this month), with the detail of every single use below — one entry per use.

An Agent that has never been used shows「尚無使用紀錄 — 到「對話」使用此 Agent 後就會出現。」(No usage records yet — they appear once you use this Agent from Chat.)

How you know it worked

  1. 「總費用 Total Cost」in「用量與費用」(Usage & Cost) shows a number (it stops being 0 once anything is used).
  2. Expanding any row shows the「佔總費用」(Share of total cost) percentage and its bar chart.
  3. On some Agent’s「用量 / 稽核」(Usage / Audit) tab,「每月呼叫上限」(Monthly call cap) holds the number you entered and「超過上限時」(When the cap is exceeded) holds the mode you chose.
  4. Deliberately set a test Agent’s cap to 1, use it twice, and the second use is blocked while「稽核紀錄」(Audit Log) gains a「用量攔截」(Usage blocked) entry. (Remember to change it back afterwards.)

Common problems

Symptom Cause / fix
「用量與費用」(Usage & Cost) shows「需要管理員權限」(Administrator access required) This page is administrator-only.
Every number is 0 Nobody has used any Agent this month. Statistics are counted monthly and start over on the first.
Requests are going up but total tokens stay at 0 This environment may currently be on a test configuration that is not metered. Usage counts and audit entries are recorded as normal, but tokens are not measured. Check with your rollout contact.
The cost is far higher than expected Expand「各 Agent 明細」(Per-Agent Breakdown) to find the largest share, then look at its「輸出 Token」(Output tokens) and「平均每次成本」(Average cost per use). Go back and check the length of its System Prompt and how many knowledge bases are attached.
Staff say an Agent suddenly stopped working It may have hit the monthly cap in「block」mode. Filter「稽核紀錄」(Audit Log) by「用量攔截」(Usage blocked) to confirm, and raise the cap and save if needed.
The cap is set but seems to do nothing 「超過上限時」(When the cap is exceeded) is set to「warn — 放行但記錄」(allow but record), which never stops anyone and only keeps a record.
Clicking「🚀 發布」(Publish) is blocked with「尚未設定用量上限」(No usage cap set) Go to the「用量 / 稽核」(Usage / Audit) tab and fill in at least one of「每月呼叫上限」(Monthly call cap) or「每月 Token 上限」(Monthly token cap).
You want to know how much one employee spent This page counts by Agent. For use-by-use records, look at the「Agent 使用」(Agent usage) events in「稽核紀錄」(Audit Log), which name the person who acted.
Where did last month’s numbers go? This page shows the month named by the「📅 本期」(this period) marker in the top-right corner. For the historical trail, see「稽核紀錄」(Audit Log).