Nothing happens because the AI asked nicely

Rules you wrote decide which actions need a person. Those actions wait in one list with everything you need to judge them, and the conversation picks itself up the moment you answer.

Start your trial See the AI agent

Shape checkAmount, currency and order id are the values the tool accepts passed
Order A-1042 belongs to MaraChecked before anything reads it passed
Rule: refunds over $25 need a leadExpires in 24 hours, and the fallback is deny matched
Priya approvedConfirmed with her second factor approved
Refunded, and written to the audit logTool, parameters, the rule, the approver, the result 2:14 pm

Enough on the card to decide without opening anything

An approval you have to go and research is an approval that sits there for a day. Every card carries the whole packet, and answering it wakes the conversation exactly where it paused.

Draft expires in 23h 41m

What is being asked for

payments.refund — $48.00 on order A‑1042, to the card ending 4417

Why it stopped here

Rule mid-refund · refunds over $25 need a lead

The customer, and the AI’s working

Mara Ellis · email verified · two open orders

“It came, but the lamp is cracked.” · confidence 0.91

Second factor

Priya · lead

Approved, refunded, and written to the audit log 2:14 pm

Example data, in the shape the console uses. An approval that expires applies the rule’s own fallback, and unless you said otherwise that is deny.

An action the AI wants to run arrives in the queue, with a clock on it.
The rule that stopped it is on the card — refunds over $25 need a lead.
So is everything you need to judge it: who the customer is, and how the AI got there.
Priya approves with her second factor. A password on its own cannot do this.
The refund runs, the conversation wakes where it paused, and the whole chain is on the log.

Denying is not a dead end

The AI delivers the alternative your procedure wrote for exactly this case, so the customer gets a real answer rather than silence.

And the queue itself is measured

Approve or deny in one tap, with a note if you want one, and several cards of the same kind together. How long approvals sit here is a number in your reports, not a feeling.

The Draft lane is a floor, not a suggestion

A rule can say an action is allowed and the lane will still hold it. That is deliberate — the lane is the stronger statement of the two.

In Draft

Anything that writes to another system, or that cannot be undone, waits for review even when a rule would allow it. Reads run freely, so the AI can still look things up while it drafts.

In Route

The AI proposes no actions at all. It reads the conversation and suggests wording to the person handling it, and that is the whole of its job.

In Auto

Actions your rules allow run on their own, and everything else still comes here first. An action with no rule that could refuse it is never granted a handler.

Approve & send in the inbox grants the message and its actions together, in one act, so a reply promising a refund and the refund itself can never come apart.

An approval is only ever recorded by somebody who proved it was them

A password is not enough to approve money moving. Anyone whose role can approve signs in with a second factor, and the approval carries that fact with it into the audit log.

Which role is needed is part of the rule, not a global setting: a rule can require a lead, a manager, an admin or the owner, and a higher role always satisfies a lower one. An approval can be delegated once, and only once — a chain of hand-offs is not accountability.

Granted — the run wakes and continues.
Denied — the AI delivers the alternative the procedure wrote.
Expired — the rule’s own fallback applies, and unless you said otherwise that is deny.
Delegated — passed to one other person, once.

Approve less over time, on evidence you can check

Nobody wants to approve the same $12 refund forever. Aqvitus watches what you actually decide and offers to loosen a rule only when your own decisions say it is safe.

When one rule and one band of values has collected at least 50 decisions in 60 days and you changed almost nothing — 97% or more approved as proposed — Aqvitus suggests letting that band run on its own, and shows you the counterfactual in plain numbers: “this would have approved 212 of the last 219 by itself.”

It never applies itself

A suggestion is a suggestion. Accepting one creates a new version of your rules, and that version runs the same test gates as any other change before it goes live.

It tightens as readily as it loosens

If a rule’s approval rate falls below 70%, or an action that ran automatically ever had to be reversed or was charged back, the suggestion goes the other way.

It reads your decisions, not the lane’s

Approvals the Draft lane held are marked as the lane’s, not a rule’s, and are left out of the statistics — otherwise the lane would quietly argue for its own removal.

AI First. Human Always.

Put the AI in front of every conversation, and keep your team in charge of everything that matters.

Start your trial See the platform