A CISO at an aviation business we work with thought his organisation had around 20 AI agents. Then he discovered there were another 60.
The employees building those agents weren’t necessarily doing anything wrong. They were finding ways to improve how they worked, which is exactly what businesses are encouraging their people to do.
But it raises an uncomfortable question. If the person responsible for security doesn’t know how many AI agents are operating across the business, how can they be confident those agents are being managed securely?
I think that example tells us something important about where we are with AI adoption.
Businesses have become very good at asking what AI can do for them. They’re less consistent about asking what needs to be in place to manage it.
And that gap is becoming harder to ignore.
We’re measuring AI progress in the wrong way
There’s no shortage of enthusiasm for AI in boardrooms. I hear it from clients, and our own board regularly asks what more we could be doing with the technology.
It’s a reasonable question. No business wants to overlook an opportunity to improve how it operates.
But I think we need to be more careful about what we consider progress.
Launching another AI pilot or deploying another set of agents is relatively easy to measure. Whether those agents are delivering meaningful value, operating securely and being managed properly is a more demanding question.
And arguably, a more useful one.
An organisation could have hundreds of AI agents and very little to show for them. Another could have a handful that make a genuine difference to how its people work.
The number tells us very little on its own.
We should be measuring AI adoption by the business problems it solves, not simply by how much technology we’ve deployed.
That starts with understanding where AI can make a difference.
When clients come to us wanting to build agents, we encourage them to start with their business rather than the technology. What are they trying to improve? Where are people spending time on repetitive work? What would a better outcome look like?
Those questions need to come before decisions about which agents to build.Security then has an important role in making sure those capabilities can be introduced responsibly, with clear controls and accountability from the outset.
AI is changing what accountability looks like
Managing who and what can access business systems has always been an important part of cybersecurity. But AI agents are adding a new layer of complexity.
Depending on how they’re designed, agents can retrieve information, interact with applications and carry out actions on behalf of users. They can also operate across several systems as part of a single process.
That raises questions about how organisations manage access, permissions and accountability when it’s no longer just people interacting with their systems.
Who decides what an agent should be allowed to do? Who reviews its permissions as its role changes? And who takes responsibility if something goes wrong?
These aren’t decisions security teams should be expected to make alone.
If an AI agent is handling part of a finance, HR or customer service process, the business still owns that process. Automating the work doesn’t remove the need for someone to oversee it, understand the risks and take responsibility for the outcome.
You can delegate a task to an AI agent. You can’t delegate away responsibility for it.
I think that’s an important distinction for business leaders to understand as AI becomes more widely used.
Security teams have a role in establishing the right controls and making sure agents operate within them. But the people introducing AI into business processes also need to take ownership of how those agents are used.
That means agreeing on responsibilities from the outset, rather than trying to work out who owns what once agents are already operating across the organisation.
Because without clear accountability, it’s entirely possible to have AI agents delivering useful results while leaving nobody quite sure who’s responsible for the decisions they’re making.
The right operating governance makes innovation easier
There’s a tendency to treat AI governance as a choice between speed and security. Either businesses move quickly and accept more risk, or they introduce tighter controls and slow everything down.
I don’t think that’s the right way to look at it.
A lack of governance can create just as many obstacles as too much of it.
When organisations don’t have a clear approach to AI, every new initiative can become a separate discussion about permissions, data access, security requirements and responsibility. Teams end up solving the same problems repeatedly, often bringing security into the conversation when the work is already well underway.
That’s not particularly efficient for anyone.
A better approach is to establish clear expectations from the outset. What can teams build? What information can agents access? Who needs to approve their use? And what responsibilities come with deploying them?
Those decisions shouldn’t require a lengthy approval process for every small experiment. But there should be enough consistency that employees understand what’s expected and security teams can support adoption without having to start from scratch each time.
Good governance shouldn’t be measured by how many restrictions it introduces. It should be measured by how confidently people can work within it.
For organisations already using AI, that starts with understanding what’s in place.
Not simply the tools that have been formally approved, but the agents employees are actually using, the systems they’re connected to and the information they can access.
Only then can businesses make informed decisions about where stronger controls are needed and how to support further adoption.
AI doesn’t require us to reinvent security
AI introduces new risks and responsibilities, but many of the principles needed to manage them are already familiar.
Identity management, data protection, access controls, and continuous monitoring remain important. The challenge is applying those principles consistently as AI agents become more capable and more closely connected to business systems.
I think there’s a risk that organisations become so focused on what’s new about AI that they overlook the foundations they’ve already invested in.
For businesses using Microsoft technologies, there may be opportunities to build on existing security and compliance capabilities rather than introduce another set of disconnected tools.
That matters because security complexity is already a challenge for many organisations. Adding AI shouldn’t automatically mean adding more products, more processes and more administration.
We should be asking whether our existing controls are being used effectively, where the gaps are and what genuinely needs to change.
But there’s another consideration.
An agent that’s appropriately configured today may need different permissions tomorrow. Its purpose might expand, it might connect to additional systems or the information it handles might become more sensitive.
That’s why AI security can’t be treated as something that ends once an agent has been deployed.
The real challenge isn’t just introducing AI securely. It’s maintaining that security as the technology and the business change.
That requires regular review, clear ownership and the expertise to understand how changes in one part of the environment affect another.
Specialist expertise matters more as technology becomes more complex
One of the lessons I’ve learned from building CWSI is that expertise requires focus.
We started in mobile security and, in 2017, made a deliberate decision to invest in Microsoft security. We believed there was more value in developing deep knowledge of a particular technology environment than trying to cover every area of IT.
That decision has shaped how we work with clients today.
As organisations adopt more technologies and connect more systems, security becomes increasingly difficult to manage in isolation. AI adds to that challenge because agents don’t operate independently of the wider business environment. They rely on existing identities, permissions, applications and data.
Understanding those connections matters.
Cybersecurity has also become a crowded market. Many technology providers now include security within a much broader range of services.
There’s a place for those offerings. But I think businesses need to distinguish between a provider that can implement a technology and a specialist that understands how to secure and manage it over time.
AI makes that distinction particularly relevant.
Deploying an agent is one part of the work. Understanding its permissions, protecting the information it uses, monitoring its activity and adapting the controls as its role changes are equally important.
And those responsibilities don’t disappear when the initial project is complete.
We’re seeing organisations recognise the value of longer-term security partnerships, where specialist expertise remains available as technologies, risks and business requirements evolve.
That reflects a wider change in how businesses need to think about security.
A one-off assessment or implementation can address a particular requirement. But it can’t account for every change that follows.
Security needs to keep pace with the business, not just with the next technology deployment.
For us, that’s why building lasting relationships and maintaining specialist expertise go hand in hand. The better we understand a client’s environment and ambitions, the more effectively we can help them manage change without introducing unnecessary complexity.
Boards need a better measure of AI progress
I expect AI to remain high on the agenda for business leaders. The opportunities are significant, and organisations are already finding practical ways to improve how they work.
But I think the conversation needs to move beyond how quickly businesses are adopting it.
It’s relatively straightforward to report how many AI initiatives have been launched or how many agents have been deployed.
Those numbers don’t tell us whether the technology is delivering value, whether the organisation understands how it’s being used or whether the right controls are in place.
I’d like to see leadership teams asking a different set of questions.
Are our AI agents solving problems that matter to the business? Do we know what they’re doing and what they can access? Is someone clearly accountable for each one? And can we continue expanding their use without creating security problems we’ll have to untangle later?
These questions should be part of the discussion before an organisation commits to its next wave of AI investment.
That doesn’t mean slowing down or waiting until every possible risk has been addressed. It means making decisions with a clearer understanding of both the opportunity and the responsibility that comes with it.
The first phase of AI adoption has understandably focused on exploring what’s possible.
I believe the next phase needs to focus on making those capabilities dependable, manageable and genuinely useful.
The real measure of AI maturity isn’t how many agents a business has deployed. It’s whether those agents deliver value, operate securely and remain under the organisation’s control.
Businesses that can demonstrate those things will be in a much stronger position to build on what they’ve started.
And that’s a more meaningful measure of progress than being first to deploy the next hundred agents.
