The problem this solves
Autonomous agents are trusted in proportion to how well they ask before they spend. A list build that would cost more than expected, an enrichment across the whole CRM, a watch that turns out to cover thousands of companies - each is a reasonable mission and each deserves a pause and a question when the projected cost crosses a line. The question has to arrive where the person is, with enough information to answer it in one word.
The rule is simple: above <credits>, the agent estimates before it executes and asks in the channel where the mission was started. Slack gets a thread reply, Telegram a message, email a reply in the same thread. The ask states the estimate, what will be delivered, and what would reduce the cost. A 'yes' runs it; a 'no' or a changed instruction reshapes it; silence leaves it queued rather than running by default. The Prospecting Agent, which runs most of the large builds, applies the rule, and so does every other agent.
The approvals are logged with who said yes and when, which is the record finance asks for and the record that makes larger ceilings possible over time.
How the mission runs
- Estimate before executing. When a mission is planned, the agent estimates the credits it will spend from the number of companies, contacts and enrichment fields involved, and compares the estimate with the threshold <credits> you set for the workspace or for yourself.
- Ask in the originating channel. If the estimate exceeds the threshold, the agent posts the question where the mission began - the Slack thread, the Telegram chat or the email thread - with the estimate, the deliverable in one sentence, and one or two ways to lower the cost, such as a smaller list or fewer fields.
- Wait for a clear answer. A reply of 'yes', 'go' or an equivalent runs the mission as planned; a reply that changes the scope re-estimates and asks again if still above the threshold; a 'no' cancels it. Silence leaves the mission queued, and a reminder is sent once after a period you choose.
- Log the approval. The approval, the approver, the channel, the estimate and the actual spend are recorded in the workspace audit trail, and the mission's final message reports the actual against the estimate.
The prompt
This is the exact objective the agent receives. Swap the obvious placeholders for your own domain, segment or channel and run it as-is from the console, Slack, or the API.
What comes back
A standing rule across your missions: anything projected above <credits> pauses and asks in the channel it came from, with the estimate, the deliverable and cost-reducing options, and runs only on a clear yes in that thread. Every approval is logged with the approver, the channel, the estimate and the actual spend, and queued missions are reminded once and otherwise left waiting.
Make it yours
- Set different thresholds per person or per agent - higher for the RevOps lead, lower for new users.
- Route approvals for anything above a second, higher threshold to a named approver rather than the requester.
- Ask for an estimate on every mission regardless of size, with auto-approval below the threshold.
- Post a weekly summary of approvals and spend to #revops.
Frequently asked questions
How accurate are the estimates?
They are computed from the counts the mission will touch and the metered cost of each capability, and the final message shows actual against estimate so accuracy is visible over time. Enrichment costs vary with how many sources answer, so the estimate is given as a range.
What if I start a mission from the console rather than a chat?
The question appears in the console thread, and if you have a preferred chat channel set, it is mirrored there so you can approve from your phone.
Can someone else approve on my behalf?
Only the people you name as approvers; anyone else replying 'yes' in the thread is told they are not an approver, and the mission stays queued.