Control
Approval on the actions that matter, budgets that stop, and a record of everything. On from the first run.
Needs your decisionjust now
Refund €48.00 for order #20417
Creating a refund in Mollie always needs approval before it runs.
Try it: the call only runs when you say so.
Approvals
Choose which actions an AI agent may run alone and which pause for a person. A gated call stops before it happens, with its arguments, the AI agent that asked and the task behind it in front of the reviewer.
In one sentence: An approval workflow for AI agents stops a proposed action before it runs and asks a person to decide. In Infinized it works on the call itself, action by action, not as a standing permission.
Every connector lists its actions, and its page shows which of them wait for a person. Reading runs alone on purpose: gating every lookup would drown a reviewer and teach them to click approve without looking.
Refunds, payouts and cancelled payments, commands on your servers, custom HTTP tools and MCP servers always wait for a person. Your organisation can let the other actions run alone once they have earned it.
Decide who may approve what: finance actions to finance, infrastructure to operations. The reviewer approves or rejects that one call, and a rejection goes back to the AI agent with the reason.
Work through a queue in bulk when the cards are routine, and get a notification on your phone for the ones that should not wait.
An action that would ask a person in chat asks a person in a flow too, and a flow can add checkpoints of its own.
Budgets
Set a monthly ceiling per organisation and per team. It is checked before every run, and at the cap new runs stop, instead of running up a bill you only see next month.
In one sentence: A budget limit for AI agents is a spending ceiling the platform enforces before a run starts, not a report you read afterwards.
!New runs of Iris won’t start until the budget goes up.The line is 80%: that is where the bar changes colour.
The ceiling is checked before a run starts. At the cap, new runs stop until the month turns or you raise it.
One ceiling for the whole organisation, and one per team underneath it, so a busy department cannot spend what another one needs.
The month-end projection turns red before you get to the cap, not after.
Each department has a limit on tasks running at once, so a burst of work queues instead of hitting your model provider all at the same time.
Tokens, cost, latency and outcome per run, rolled up by team, AI agent and model, and exportable as CSV.
Audit trail
Every run, tool call, approval and setting change is recorded: who, when, with which arguments and what came back. Filterable on screen, and exportable when someone needs the file.
In one sentence: An audit trail for AI agents records what each agent did and on whose authority, so that "who authorised this change?" still has an answer months later.
Showing 7 of 7 events
mollie_create_refund · €48.00 on #20417Jorienmollie_create_refund in the refund flowIrisclaude-sonnet-5Mattzendesk_add_comment · reason sent to NovaIlsemistral-small-latest · $0.0004NovaFilter on screen, export when someone needs the file.
Which AI agent ran, on which task, which tools it called with which arguments, and what came back.
Approved, rejected and executed calls are all recorded, with the reviewer who decided and when.
Who changed an AI agent's instructions, a budget, a connection or the approval list, and when.
Tokens, cost, latency and outcome sit next to the record, and failures carry their error messages.
Filter by what you are looking for, and export when an auditor or a customer needs the file rather than the screen.
Where it stops
Questions
Actions that change something in a connected system, such as replying to a customer, creating an issue, booking an entry, changing a DNS record or deploying a site. Each connector page lists its own. Reading does not wait.
Yes, per organisation, once it has earned it. The floor stays: money leaving the account, commands on your servers, and custom or MCP tools always wait.
Approvals are human oversight a deployer can point to, and every decision is recorded. Whether your use case is high-risk is yours to classify.
New runs are stopped until the next month starts or you raise the ceiling.
You, directly to your model provider, on your own API keys. Infinized charges for the platform and adds no markup to tokens.
Yes. Filter it on screen and export it when someone needs the file rather than the screen.
Keeping records is one of the things the Act asks of a deployer of a high-risk system, and the audit trail and per-run records are building blocks for it. Whether your use case is high-risk is yours to classify.
We start by listening to your question. Then we show you a demo that fits your organisation.