Skip to main content Skip to footer

Interested?

Get in touch

Article AI agents series: part two

Speed with a person on every action that matters

You’ll already have a view on how much autonomy an AI agent should have. Here’s ours, from running agents on live work: give them plenty of speed, and keep a person on every action that matters.

That isn’t paranoia dressed up as strategy. It’s what let us cut threat containment in our own security operations from four hours to under 45 minutes. The approval step wasn’t the brake on that result. It was the reason we could take the handbrake off at all.

What we actually did

Before

4 hours

→

Now

Under 45 mins

Time to contain a threat in our security operations, with an experienced analyst approving every containment action.

In our security operations, agents now help our analysts contain threats in under 45 minutes. It used to take four hours.

The agents do the heavy lifting: pulling the evidence together, working out what’s affected and lining up the response. Then an experienced analyst approves every containment action before it happens. No exceptions, and no quiet carve-outs for “low-risk” steps.

This isn’t a lab demo. It’s live work for real customers, running every day on the same team, controls and Microsoft stack we’d bring to your organisation.

Why approval didn’t slow us down

The usual worry is that a human in the loop turns a fast agent back into a slow process. In our experience, that’s the wrong place to look for the delay.

Most of those four hours weren’t spent deciding. They were spent gathering, checking and piecing things together. That’s the work the agents took on. By the time an action reaches our analyst, the case is already made and the decision takes moments, not hours.

And there’s a second effect that’s easy to miss. Because a person signs off every action that matters, we’re comfortable letting the agents run hard on everything else. Without that approval step, we’d have kept them on a much shorter lead. Governance is what bought us the speed.

“Governance is what bought us the speed.”

Where to draw the line

“Every action that matters” does a lot of work in that sentence, so here’s how we think about it.

Let the agent run

Where it’s reading, gathering, summarising or recommending. If it gets something wrong, a person catches it before anything changes.

Put a person on it

Where the agent would change something in the real world: blocking an account, isolating a device, touching customer data, or anything that’s hard to undo.

  • Make approval quick. The agent should hand over a clear case, not a pile of logs. If approving takes longer than the old manual process, the design is wrong, not the principle.
  • Revisit the line. As an agent proves itself on a type of action, you can move it. But move it on evidence, and write down that you did.

The line will sit in a different place for your use cases than for ours. What matters is that you’ve drawn it on purpose, before the agent goes live.

The model behind it

Approval is the visible part. Underneath it sits a simple set of controls that makes it hold up:

01

An agent register

So you know every agent you’re running, what it can see, what it can do and who’s accountable for it. Every agent is a new identity with access to your data, so it needs managing like one.

02

Approval gates

So the actions that matter always reach a person, and the record shows who approved what, and when.

03

A risk log

So you’ve thought through what could go wrong before it does, and can show your thinking to anyone who asks.

04

Audit evidence

So when your board, a regulator or a customer wants to know how a decision was made, you can show them.

None of this is heavy. It’s what lets you say yes to the next agent with confidence, rather than starting the argument from scratch each time.

Want speed you can trust?

We’ll help you decide where the line sits for your agents, and put the controls in place to hold it.

Talk to our team Our Security Operations Centre

Our vendors, suppliers, and partners