Approval gates
Mark the actions that need a person. The agent suspends and waits, up to a window you set.
Set the goal.
For the work you cannot draw as a flowchart. Give an agent an objective and the tools, and it decides what to do next, shows you its reasoning, and stops for you on anything that matters.
Every step it takes is recorded, including the ones it decided against.
Agent runChurn watch · daily
The goal you set
Every morning, find the accounts that stopped using us last month and tell me which are worth a call.
Pulled usage for 1,284 accountsYour API
Found 37 with a drop over 60%Analysis
Checked open tickets and invoicesData store
4Ranking by contract value and riskReasoning
5Draft the outreach noteWaiting
Three accounts need your call, not an email
Contracts over 50k with an open complaint. Want me to book time with the owners?
Yes, book itNot this time
A goal
Not a script you have to maintain
Its reasoning
Visible on every step it takes
Your tools
Only the ones you hand it
Your sign-off
On anything you mark sensitive
An agent is only as useful as what it can reach
On its own, a model can write. Connected to your systems, an agent can find the record, check the invoice, update the sheet and tell the customer. You decide exactly which doors it has keys to.
- Grant tools per agent, not per account
- Read-only where read-only is enough
- Every call it makes is logged
- Revoke access without redeploying anything
WWeb searchPublic information
DData storesYour own documents
YYour APIsAnything you expose
SSpreadsheetsExcel files, read and written
JJiraIssues and comments
EEmailRead and send
CCode executionSandboxed
OOther agentsDelegate a sub-task
You choose which of these each agent may touch.
Answers from your material, not from the internet's
Point the agent at your policies, contracts, manuals and past tickets. It searches them by meaning, answers from what it finds, and attaches the source so anyone can check.
- Semantic search across your own content
- A citation on every claim
- Says so when your material does not cover it
- Nothing you upload is used to train a model
Asked
Can a reseller in Germany return opened goods after 21 days?
No. Resellers have 14 days for opened goods, and only when the seal is intact. Consumers have 30. The exception is a documented fault, which has no time limit.
reseller-terms-v4.pdf · p.12returns-policy-2026.docx · §3
Answers come from documents you uploaded, with the source attached. If your own material does not cover it, the agent says so rather than inventing a policy.
Autonomy with a handbrake
The question is never whether an agent can act. It is what happens the day it gets something wrong.
Visible reasoning
Read why it chose a step, what it found and what it ruled out.
Scoped tools
It can only touch what you granted. Nothing discovers itself along the way.
Model of your choosing
Pick per step. Whichever model ran is recorded with the run.
Budgets and limits
Cap spend and step counts so an agent cannot run away with a task.
Ask a colleague
When it is genuinely unsure, it asks a question rather than guessing.
One queue for everything waiting on you
Approvals do not scatter across inboxes. They land in one place with the context attached, and you approve, edit or reject. Editing is the one people underrate: correct the draft and the agent carries on with your version.
- See exactly what the agent proposes to do
- Edit before approving, not after the fact
- Reject with a reason the agent takes on board
- Nothing sensitive moves until someone says so
Waiting for you3
R
Refund 148 EUR to order #4821
Refund agent · Over the 100 EUR limit
ApproveEditReject
O
Send 12 follow-up emails
Outreach agent · Outbound to customers
ApproveEditReject
O
Cancel duplicate subscription
Ops agent · Touches billing
ApproveEditReject
Give it the work that needs a brain
If a person has to read something and decide, that is agent work. If they just follow steps, use a workflow.
Market and competitor briefs
Gather from the web and your own files, weigh it up, write the brief a person would have written.
Churn and win-back watch
Spot the accounts going quiet, work out which are worth chasing, draft the outreach.
The hard half of the queue
The tickets a script cannot close: read the history, work the problem, escalate what it should.
Exception handling
The invoices that do not match. It investigates, explains the mismatch and proposes the fix.
Vendor and partner chasing
Follows up until it has an answer, then updates the record and tells you what changed.
Answers from your own material
A copilot over your policies, contracts and wikis, citing the document every time.
Competitor briefsChurn watchInvoice exceptionsTicket investigationVendor chasingPolicy answersLead researchContract reviewData clean-upWeekly summariesCompetitor briefsChurn watchInvoice exceptionsTicket investigationVendor chasingPolicy answersLead researchContract reviewData clean-upWeekly summaries
Watch one work through your problem
Bring a task your team does by hand and we will point an agent at it, on your data, with you watching every step.