Finance
“Record the payment” cannot be a robot's click
Looking up the order is one thing. Recording that it was paid is another: it changes the system, has consequences and someone has to answer for it.
Confirmation first
Platform · Approvals
The AI stops before acting. A person decides.
An approval is the pause that happens when an agent is about to perform an action classified as sensitive in a connected system: instead of acting, it shows what it is about to do and waits for Approve or Deny. The decision belongs to a person, is recorded and applies to that action, in that company. Not to be confused with document review: here an action is approved, not a text.
See how it worksLOW · MEDIUM · HIGH · CRITICAL
every function has a risk level; high and critical ask for confirmation by default
ALWAYS · INHERIT · NEVER
the company, the agent and the conversation itself can tighten or loosen the rule
APPROVALS INBOX
in the bell and on its own page, only what awaits your decision
SLACK · TELEGRAM
whoever uses Skyller through those channels decides there, with buttons
Where this shows up in your day
Giving AI autonomy is scary until a reliable brake exists.
The fear is not the AI answering wrongly — it is the AI doing something that cannot be undone: sending, deleting, removing, archiving. Without a brake people recognize, the company freezes the use of agents out of caution.
Customer service
The email that goes out under the company's name
A draft helps; an automatic send makes a commitment. People want to read it first, without copying and pasting somewhere else.
Read, approve, send
Management
“Who authorized this?”
When something goes wrong, the question is always the same. A “confirm?” in plain text in the middle of a chat is not an answer that survives an audit.
Record of who and when
How it works
From request to action, with a person in the middle.
The mechanism is the same in chat, in an unattended automation or on Slack. What changes is where the person decides; what gets recorded does not.
01
Every function has a risk
Each function of the connected systems gets a level: low, medium, high or critical. Looking up is low; deleting, removing, archiving and sending are high or critical and ask for confirmation by default. Editing and registering land on medium and go straight through — unless your company tightens the rule.
02
The company adjusts the rule
An administrator can require approval for a specific function or waive it — and lock the decision so nobody loosens it. The agent and the conversation have their own layer too: inherit, always or never.
03
The agent stops and shows
On reaching the action, the agent does not execute: it opens a card with the action and its parameters and stops there. It does not ask “confirm?” in text.
04
A person decides
Approve or Deny, on the conversation card, in the notification bell, on the Approvals page, or through Slack and Telegram. Whoever only reads a shared conversation does not decide — and the administrator does not decide for anyone either.
05
It gets recorded
The decision applies to that call, in that company, and is kept: who asked, what it was, who decided, when. Without a decision, the pending item expires and the action does not happen.
Live proof
Nobody at the screen. The automation stopped anyway.
The most honest test of a brake is when nobody is steering: an early-morning automation reaches a sensitive action and what happens is a pending item, not an action.
Starting point
The “Weekly collections” automation runs at 7am with the finance agent. “Record payment” is marked as always ask for approval.
What happens
- 01The run checks overdue orders and builds the summary; on reaching “record payment for order 4471”, it stops. The history marks the run as paused and the notice says “awaiting approval”.
- 02The pending item appears in the manager's Approvals inbox, with the action and its parameters, and the notice arrives in the bell — and on Slack, if the company uses Skyller there.
- 03The manager approves. The run resumes where it stopped, records the payment and the pending item moves to Approved, with who decided and when.
Result
The action happened because a person said yes — and the difference between “the AI did it” and “the manager authorized it at 8:12” is on record.
What this scene does not promise
- Only actions classified as sensitive stop. Read-only look-ups and text drafting do not open a card — and the company can change a function's classification.
- The Approvals inbox gathers pending agent actions. Sharing invitations and document review are separate mechanisms, with their own screens.
- A pending item without a decision expires: the run is closed without acting and the record shows that nobody decided.
To evaluate calmly
Where the rule is set, and where the decision happens.
Select a topic to see its capabilities, conditions, and limits.
- Risk per function
- Each function of a connected system is classified as low, medium, high or critical. High and critical ask for confirmation by default; the company can review the classification.
- Company policy, with a lock
- The administrator requires or waives approval per function and can lock that choice. Locked, it overrides anyone's preference — in both directions.
- Policy per agent
- In each agent's tools, every item has “inherit”, “always ask” or “never ask”. That is how a read-only agent becomes an agent that acts with a brake.
- Adjustment per conversation and per person
- A conversation can tighten the rule for the session. Each person can mark a function as “always ask” or “always accept” — only where the company did not lock it.
- Card with action and parameters
- The card shows what will be done and with which data, before it happens. Approve resumes the run where it stopped; Deny ends that action, and the agent carries on with what it has.
- Approvals inbox
- Pending items appear in the notification bell and on their own page, with Pending, Approved, Rejected and All. Each person sees and decides their own. You can review the detail, decide a simple confirmation right there or open the conversation; pauses with several actions are decided in the conversation.
- Off screen: Slack and Telegram
- Whoever uses Skyller through a connected channel gets the card there, with Approve and Deny buttons. One click decides; the card closes after the decision.
- Automations stop too
- In a scheduled run, the sensitive action leaves the run paused, awaiting approval: the owner and collaborators are notified and the pending item appears in the inbox, identified by the automation.
- Record and validity
- The decision is tied to that call and that company, with who decided and when. A pending item without a decision expires without executing. Every tool call is recorded too — which one, in which system, by whom and which rule decided whether it asked for approval — in the company's activity trail.
Real application
Two functions enabled. One of them waits for the manager.
The gain is being able to give the agent a function that acts — and sleep well: the action exists, but it does not happen without someone saying yes.
- 01Administrator
In the connected order system, leaves “look up order” enabled without approval and marks “record payment” as always ask for approval, locked.
- 02Finance analyst
In the Finance hub chat, asks the agent: “Check order 4471 and record the payment”.
- 03Finance analyst
Reads the card — order, amount, date — and denies: the amount does not match the receipt. The agent replies with what it looked up, recording nothing.
- 04Finance analyst
Fixes the amount with the customer, asks again and approves. The payment is recorded and the decision sits in the Approved tab.
- 05Manager
In the company's activity trail, sees the sequence: request, denial, new request, approval — who, what, when.
Comes ready
- Risk classification per function and confirmation by default on high and critical
- Card with action and parameters in chat; Approve resumes, Deny ends the action
- Approvals inbox with tabs, detail, decision and pending counter
- Cards with buttons on Slack and Telegram for connected channels
The company decides
- Which functions ask for approval in the company, and which are locked
- The policy per agent: inherit, always ask or never ask, item by item
- What each person may adjust for themselves, where the company did not lock
- Which channels are connected to decide off screen
Questions from whoever decides
What people ask before enabling a real action.
Next step
Enable the function you do not let anyone automate today.
Mark it as always ask for approval, request the action in chat and watch the card appear before anything happens. Or schedule a demo and see the pause working with an example from your routine.