Overview
If your usage is higher than you expected, or you are approaching your limit, this page is where to start. It explains what actually consumes your AI capacity and the specific changes that reduce it — most of which take a few minutes and cost you nothing in capability. Your workers use AI capacity to do work — every source they check, document they read, search they run, and tool they call uses some of your allowance. So the question that decides your usage isn’t how often you talk to your workers or how long their answers are. It’s how much work each task involves, and how often that task runs. A request pointed at one inbox uses a fraction of what the same request uses when it’s pointed at “everything”. That difference — repeated every day by a recurring task — is where almost all of the available saving lives.AI capacity is measured in Work Credits. That’s the unit you’ll see on your plan, in
Settings > Usage, and in any message from a worker telling you a limit has been reached.
Everything on this page reduces the number of Work Credits your workspace gets through.
The three things that matter most
How much work a task involves
A tightly-scoped task checks a few things and stops. A vague one keeps going. The biggest lever by far.
Your recurring tasks
Scheduled work is the largest single source of AI capacity usage for most customers.
Conversations that never end
A thread kept alive for weeks across many topics costs more with every message.
- Answer length. What your worker writes is a small part of what a task uses.
- Follow-up messages on the same topic. Continuing a focused conversation is efficient — often more so than starting a new one. Far more usage goes into the opening task of a conversation, including every time a scheduled task or an incoming email starts one.
This is why “send fewer, longer messages” isn’t the win people assume. Focus on scope, not wording.
1. Scope the work, not the wording
The difference between an efficient task and an expensive one is almost always how many places the worker has to look before it can stop.- ✅ Tightly scoped
- ❌ Open-ended
“Check the Vendas inbox for unread emails from today. For anything from an existing client,
draft a reply. Skip newsletters and notifications. If there’s nothing, reply ‘nothing today’
and stop.”Bounded: one inbox, one day, a clear skip rule, and an explicit stop condition.
- Name the sources. “Check the Vendas inbox” instead of “check my email”.
- Set a boundary. A time window, an item limit, a named list — “the top 3 competitors”, not “the competitive landscape”.
- Give an explicit stop condition. Tell the worker what “done” looks like, and what to do when there’s nothing to do.
- Say what to skip. Exclusions prevent whole branches of unnecessary work.
This matters most for recurring tasks, because the saving repeats on every single run.
2. Audit your recurring tasks
Recurring tasks are the largest single source of AI capacity usage for most workspaces — and the easiest thing to fix, because every improvement compounds on every run. Open your scheduled tasks and ask four questions in this order:a) Are any of these doing the same job?
A common and easily-missed pattern is a set of near-duplicate reports — a separate daily briefing per region, per inbox, or per client — that each reload similar context. Tasks that fire at the same time of day are worth looking at first. Fix: merge them into one task producing one combined report. You get one email instead of five, and the workspace does a fraction of the work.b) Does this really need to run that often?
A task running every six hours uses roughly four times what the same task uses running daily. Ask what actually changes between runs. If you read the morning report and ignore the rest, you’re paying for three reports you don’t use. Tasks that run more than once a day deserve the closest look. There are usually only a handful of them, and they account for a disproportionate share of what recurring work costs.c) Is each run doing more than it needs to?
This is where the real saving is. How often a task runs sets how many times you pay; how much work it does sets how much you pay each time — and that varies far more between tasks than frequency does. Two customers can run the same hourly email check and use very different amounts of capacity, purely because of how the task is written and how the worker is set up. Open your most involved recurring task and apply the four scoping rules above.d) Is anyone reading the output?
If a report goes to an inbox nobody opens, delete the task. This is more common than you’d expect.3. Keep a conversation to one topic
A conversation carries its history forward. That is what makes follow-ups on the same subject efficient — the worker already has the context and doesn’t have to rebuild it. The cost appears when a single thread is kept alive indefinitely and used for everything. As it accumulates months of unrelated history, every new message carries all of it, and each message costs more than the one before. A small number of these very long threads can account for a surprising share of a workspace’s total usage. What to do:- Start a new conversation when the topic changes. Moving from last quarter’s reporting to a new client is a new conversation, not a new message in the old one.
- Don’t keep one permanent thread as a general-purpose channel to a worker. If you’ve been in the same conversation for weeks, that’s the signal.
- Keep working threads for work in progress. A conversation that runs for a day or two on one task is exactly right — this is about threads that never close.
Nothing is lost by starting fresh. Your worker keeps what it has learned in its memory, so a new
conversation still knows your business, your preferences and your standing instructions.
4. Trim tools down to what the worker actually uses
Every tool connected to a worker adds to what that worker carries into every task it performs. A worker with a long list of connected tools is meaningfully less efficient than a focused one — and that difference applies every time it works. Review each worker: which connected tools has it actually used in the last month? Disconnect the rest — you can always reconnect them later. This is also the strongest practical argument for specialists over generalists. A worker with one clear job needs a handful of tools, not a dozen.5. Watch for the heavy tail
A small number of very large tasks will usually dominate your usage. They’re typically:- Open-ended research (“find everything about X”)
- Whole-document analysis, where only a section was needed
- Multi-step work with no stop condition, where the worker keeps going
6. Pause what you’re not using
Workers you’re not currently using can be put on holidays — they stop processing entirely and resume when you need them. For recurring tasks you want to keep but not run right now, pause rather than delete, and you keep the setup.Ask your worker to audit itself
Your workers can review their own setup. Paste this into a conversation with any worker:Your worker can see how it is configured, but not what it costs — the prompt above is
deliberately written so it reports the former and doesn’t guess at the latter. Pair its answer
with the actual figures from Settings > Usage.
A quicker version
For a fast check on a single task:Quick Reference
Next Steps
Usage Limits
Understand what each limit means and what happens when you reach them
Scheduling Tasks
How to set up and adjust recurring work
Plans and Pricing
Compare plans if you need more capacity
Managing Multiple Workers
Best practices for organizing your AI team