The fastest way to lose trust in an AI system is for it to do something you didn't expect. One wrong email to a customer, one deleted record, one meeting booked with the wrong person, and a team stops delegating anything important. Many AI products respond by asking for less permission and hoping for the best. We designed Verkko the other way round.

The rule

Reads run freely. Sending, booking, deleting and deploying always stop and ask. Verkko can search your email, read your documents, analyse your pipeline and prepare everything an action needs, without interruption. But the moment something would leave the building or change a system of record, it waits for a person.

This isn't a setting someone might forget to switch on. Underneath, it's enforced by Baton, our protocol for governed AI: any tool with side effects pauses the agent for approval and doesn't execute until a person says yes. Any tool that hasn't been explicitly classified is treated as having side effects, so the safe default is the only default.

How it works in practice

Anything gated waits in the Approval Center. Each card shows what wants to happen and why: the draft email and who it's going to, the ticket and where it will be filed, the record and exactly what will change, along with the steps that led there.

  • Approve it and it happens immediately.
  • Reject it with a reason, and the agent uses that reason to adapt rather than simply failing.
  • Edit the draft first, so "almost right" doesn't mean starting over.

A workflow waiting for approval is paused, not failed. It might wait minutes or overnight; its state is saved, and it continues from exactly where it stopped the moment you say yes. Approvals can be routed by amount or risk too, for example a second Finance sign-off on purchase requests above £5k, so policy is enforced without slowing everyone down.

A "no" is useful

In most systems, a rejected action is an error. In Verkko, it's information. When you deny something, that denial goes back to the agent as the result of its request, and it can revise its approach or finish the task without that step. Reject an email because the tone's wrong, and you get a better draft, not a broken workflow.

Audition before you trust

New automations don't have to go live on day one. They can run in suggest-only mode: for a week, you read what they would have done, every email and every update, without anything real happening. Once you're confident, switch to live.

Simulate mode does the same for a single run. Every step shows exactly the email or message it would have produced, and nothing real is touched. It's the fastest way to check a change before it affects anyone.

Everything on the record

Approval only works if you can see what you're approving and, later, what happened. Every action Verkko takes is logged and attributed: who asked, what each agent saw, which facts it relied on and who approved the final step. When someone asks "why did the system send that email?", there's a precise answer, not a guess.

The paradox

You might expect all this to slow teams down. In practice the opposite happens. Teams that can see and stop every action end up delegating more, not less, because the cost of a mistake is a rejected card rather than an apology to a customer. They start with drafts, move to routine follow-ups and eventually hand over whole workflows, because they've watched them work.

Control is what makes autonomy possible. Read how the mechanism works in Why agents shouldn't talk to each other directly.