Every AI agent needs an identity, an owner and a limit

I’m pleased to share that I’ve been cited as a contributing voice in a new industry report from Change Makers Club: The Security Shift: Your agents have the keys. But who secures them?

It draws on a roundtable of senior technology, security and product leaders held in London in September 2026, and the full report is available to download below.

The conversation in the room

Most organisations have moved past asking whether to use AI. The harder question now is what happens when AI stops simply answering and starts acting: booking, buying, approving and changing records on someone’s behalf. As the report puts it, AI agents are being handed access faster than anyone can say who they are, what they may do, or who answers for them.

“Identity and Access is now the control plane for every decision AI makes.”
The Security Shift, Change Makers Club

My contribution

My contribution, on page 4, is a simple test any organisation can apply today:

“Every AI agent needs an identity, an owner and a limit. Many organisations are deploying the first element before they’ve decided on the other two.”

Identity, owner, limit: why the order matters

• Identity. An agent isn’t the user, and shouldn’t borrow the user’s credentials. It needs its own identity, so you can see what it did, on whose behalf and when. This is the part most organisations get to first, because the technology makes it easy.

• Owner. Someone has to be accountable for each agent, in the same way a product leader owns a roadmap. Without a named owner, risk falls between teams, and nobody notices when an agent’s behaviour drifts.

• Limit. Every agent needs a defined boundary: what it may access, what it may decide on its own, and when a human has to step in. Limits should be refined as an agent proves itself, not set once and forgotten.

Deploying identity without the other two is like issuing someone a building pass without saying who they report to or which floors they’re allowed on. The pass works perfectly, and that is the problem.

In my own experience of running identity at scale, every entity acting in a system needs its own identity, whether that is a human user, an authenticated service, another non-human entity or, now, an AI agent. Only then can the system strike the right balance between access, productivity, security and risk.

Building trust into products at scale is one of the threads running through my book, Product Truths, and it was good to see the same theme come through so strongly from such a varied room.

Read the full report

My thanks to George Whittington and the Change Makers Club team for bringing the conversation together and turning it into something practical.

[Download The Security Shift report (PDF)]

If you’re working through how identity and trust need to evolve as AI agents take on more, I’d welcome the conversation.

No one photographs the planning. Only the peak.

Setting out on a day's hike last week in Cumbria with my great friend @DavidPull – the preparation was extensive and wide-ranging. Before we even set foot on the trail, there was more planning than most people would ever realise. New walking boots, spare laces, a pre-chilled bottle of water for the forecast heat, a fully charged phone and battery pack.

Summiting the peak is just the final, visible part. The part that makes the memorable photo, the soundbite or the outcome summary.

What no one sees is the journey before the ascent: the compounding experience of previous hikes, checking local weather conditions in advance, the specific kit choices for the day – even who you choose to climb with.

That same invisible preparation is what separates robust, resilient systems from fragile ones. The great systems – the ones that scale seamlessly and bounce back quickly when tested – are the product of stress-testing every angle in advance: informed trade-offs, fall-backs ready to deploy, contingent processes that kick in at exactly the right moment, and the right people on the team. That's what lets a team react gracefully under pressure – an incident, an unexpected outage, a system taking an unexpected turn – rather than scrambling in the moment.

The best, and safest, hill climbs don't just 'happen'. The planning, preparation and forethought account for most of the outcome long before the summit comes into view. Systems at scale are no different – the day-to-day micro-actions of optimising frameworks, honing documentation, and keeping run books slick and ready for the worst. That's no accident either; it's by design.

The next time you see a clean outcome — a summit photo, a smooth incident response, a system that just works — remember you're only seeing the final 10%.

+ Plus Perspectives

Whether it's climbing a mountain or operating a complex system at scale, the steps to success are similar:

> Plan: map out blind spots, the risk areas, the unhappy paths and run through the key 'WhatIf?' scenarios, to stress test your system in advance.

> Prepare: for the worst, whilst always hoping for the best (but the worst will often happen anyway, so best to be ready). Your disaster management and recovery systems, your run books, incident and escalation paths that need to play out in real time (and ensure they're documented, of course).

> Perform: the 'final 10%' of how you act in the moments of crisis or when the system is tested — even these can be 'dry run' or rehearsed, to build the crisis management muscles. The 90% readiness comes from the first two Ps; the remaining 10% comes from how your team 'performs' in the moment of truth.

Photo: me on top of Irton Pike, near Stanton Bridge, Cumbria, July 2026

Photo: Irton Pike, Cumbria - panoramic view, July 2026