AI employee budget and cost control: where the spend goes and how to trim it
A budget stops overspending; good design stops waste. Where an AI employee spends, how to see the cost before it runs, and five changes that cut spend without cutting results.
Part of the guide AI employees
Good AI employee budget and cost control has two halves. The first is a ceiling: each employee gets a monthly budget, a warning at 80% and a pause at 100%, so no single flow can drain what the team needs. The second, which saves more, is design: knowing which steps cost anything, putting the cheap filters before the expensive ones, and not asking for results you will throw away. This article covers both, with no prices, because those live on the pricing page and can change.
It is written for whoever owns the credit balance in a team: an operations lead, a founder, or the person who built the first few AI employees and now has to explain the bill.
Where an AI employee spends
Credits are spent by the steps that call a service or an AI, and only by runs that pass (a failed run pays for the data and AI steps that finished). In most flows the spend comes from three places:
- Data searches: each search a data block sends, plus each place or item it brings back.
- AI steps: xMap LLM, the classifier, the AI agent and Check the work pay for each item they read, so their cost grows with the list that reaches them.
- Sends: emails, app messages and records created, counted per call made.
Triggers, logic blocks such as Filter, Sort and limit and Only new, and lookups in your own tables cost nothing. Sample figures are free too, and said so.
See the cost before it runs
Run once shows what a run will cost before you press it. Beside Deploy, an estimate for the month multiplies the runs its timetable makes in 30 days by the cost of the last real run, or the test estimate when there is none yet. The Ready to go live window shows both before the first deploy. Keep in mind that a test run works on live data, so its searches and AI steps are charged; only the estimates shown while you edit are free.
Five changes that cut spend without cutting results
- Filter before the AI. Put Filter, Sort and limit, or Only new in front of xMap LLM. Scoring 30 new places costs a fraction of scoring 300 that include last week's.
- Ask for what you will read. A Most places of 20 on a search whose email shows the top 10 pays for 10 places nobody sees.
- Run as often as the world changes. New openings in a district do not need checking every hour; a daily or weekly timetable finds the same things. Use When something changes for watch jobs that only matter when something moves.
- Prefer free checks. In Check the work, rules about phones, websites, emails, word limits and personal data are checked by code for free; only rules written in words use AI.
- Cap the agent. An AI agent's Most tool calls limits how many paid tools it may use for one message.
The layers of limits
On top of good design, four limits stack, and the tightest wins:
| Limit | Set by | What happens when it is reached |
|---|---|---|
| Monthly budget per AI employee | Its owner or an admin, in its panel | Warning at 80%, the flow pauses at 100% with a one-click raise |
| Team quota | Organization admins, on the Team page | The team's people cannot spend more this month |
| Organization spending limit | Admins, on the Credits page | Nothing in the organization spends past it |
| Daily limits on outside services | xMap | That service is not asked again today; nothing is charged for it |
Every limit starts again on its own period; budgets and quotas each month.
The employee budget is the one most teams adjust. At 80% its owner and the admins get a notice, and the owner an email. At 100%, or when the next run needs more than is left, a live flow pauses: its schedule and checks stop and nothing more is spent. Raising it in one click starts it again at once. More on choosing that first number is in setting a monthly budget for each AI employee.
A worked example (made up)
A bubble tea franchise in Kuala Lumpur runs six AI employees. In the third week of the month, one of them, a scout for new shopping malls, sends its 80% warning far earlier than usual. The operations lead opens Usage, which shows credits by flow, person and month, and sees that this one flow doubled its spend from the week before.
The flow's version history shows why: a colleague widened its area of interest from two districts to the whole Klang Valley and set Most places to its maximum, while xMap LLM still scored everything. The fix takes five minutes. The area stays wide, but Only new now sits before the AI step, and Sort and limit keeps the 30 best rated. The next run's estimate falls back to its old range, and the lead raises the budget once for the rest of the month rather than leaving the scout paused.
Mistakes to avoid
- One giant budget. An organization limit alone means the first busy flow can spend what the others needed. Give each employee its own.
- Budgets set so tight they pause every month. A paused alert flow is worse than a slightly higher bill. Set it from a month of real runs.
- AI on everything. Use xMap LLM where judgement is needed, not to reformat fields a Calculated field can produce for free.
- Ignoring the warning. The 80% notice is the moment to look at what changed, not just to raise the number.
Templates that already filter before scoring, such as new places every morning, only new and competitor openings, are good models. The AI employees guide covers the rest of running a team of them.
Questions
Do failed runs cost anything?
A failed run pays for the data and AI steps that finished before it failed. Nothing else is charged, and sending blocks that did not send cost nothing.
What happens to a paused employee's work?
Nothing is lost. Its flow, settings and results stay; it simply stops running until the budget is raised, taken off or the month starts again.
Can a team member spend past their team's quota?
No. What a person spends counts toward their team, and the team's quota is checked before anything is taken.
Where do I see which employee spends most?
The Usage page shows credits by flow, by person and by month, and each employee's report card lists the credits it used that week.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates