Businesses are experimenting with AI agents. For many, deployments are still small, but it’ll soon grow to hundreds and perhaps thousands. The real challenge won’t be building them – it will be governing, securing and managing them in the long term.
The autonomous AI Workforce is coming
As businesses move beyond experimentation, they’ll have a growing ecosystem of autonomous AI agents – each performing specialist work on behalf of people. That might be procurement, or recruitment, financial forecasting, contract review, producing sales proposals, resolving service desk incidents, or manufacturing planning.
Many agents will work together. Some will trigger others. Some may operate 24×7.
AI will no longer consist of individual copilots, but hundreds of digital workers.
Who manages them, and how?
The Governance challenges you’ll face
You’re likely to face five key governance challenges.
Agent sprawl
As individuals and departments create AI agents to solve business problems, organisations will inevitably accumulate hundreds of perfectly legitimate agents. Without effective lifecycle management, agent sprawl – much like VM sprawl – can take hold, making governance, security and optimisation increasingly difficult.
Permission creep
Over time, agents can accumulate permissions beyond their original purpose. This may be through changing business requirements, cloned configurations, or expanding responsibilities. Unless regularly reviewed, the principle of least privilege is gradually eroded and eventually an HR agent may be able to access Finance data, or vice versa.
It is very similar to the problem organisations experienced with SharePoint permissions.
Change and version control
Superficially, change control sounds simple enough. But agents evolve, prompts change, models are updated and organisational knowledge grows – and this happens dynamically.
How will you keep track of which version is live, who approved changes, and what effect they’ve had on performance?
Understanding how AI work gets done
Increasingly, agents will trigger other agents. The governance challenge isn’t that individual AI agents make decisions; it’s that dozens of interconnected agents can collectively reshape business processes in ways that become increasingly difficult to understand, audit and optimise.
Ownership and lifecycle management
If ownership isn’t clearly established, agents will become the next generation of orphaned digital assets. You’ll recognise this problem from applications, SharePoint sites, Teams, Power Apps and virtual machines – without clear ownership they persist, because nobody wants to switch them off.
Effective governance can prevent these challenges becoming operational problems. But it’s not a one-time fix and will require ongoing activity.
The emergence of AI Operations
Look ahead to when you’ve got hundreds, or even thousands, of agents. You’ll want to know:
- How many agents you have
- Who created them
- Which departments own them
- Which agents are still being used
- Which have access to HR or Finance data
- Which agents communicate with each other
- Which are generating business value
- Which are costing thousands in AI consumption each month.
Microsoft has recognised this problem and created an entirely new management platform to address this challenge. Microsoft Agent 365 provides a central control plane covering agent registry, mapping, analytics, and lifecycle management.
Agent registry
This is perhaps the biggest benefit, quickly helping you address ‘agent sprawl’. Instead of hunting through multiple platforms, administrators get a central registry showing what agents exist, who owns them and where they came from.
Agent map
As AI workflows become increasingly interconnected, understanding those dependencies becomes essential. Usefully, Agent 365 doesn’t just list agents but it visualises the relationships between them. So, if a sales agent is linked to a pricing agent, which is linked to an approval agent, and in turn to a contract agent, that’s easy to see.
Analytics
Organisations will increasingly want to understand which agents are delivering value. See ‘Why AI needs FinOps: Controlling costs without limiting innovation’ for more on this. Agent 365 provides a range of relevant measures, such as usage, performance, quality, business impact, and ROI.
Lifecycle management
AI agents will need to be treated like other IT assets, and Agent 365 makes it easier to manage an agent’s lifecycle. From onboarding and applying governance policies, through ongoing reviews, to retirement.
Agent 365 doesn’t replace or duplicate existing capabilities but orchestrates them, so you’ll have:
- Microsoft 365 Admin Centre providing central administration
- Microsoft Entra giving agents individual identities and controlling what they can access
- Microsoft Purview governing what data agents consume and create
- Microsoft Defender monitoring agent activity for threats or malicious behaviour, and
- Microsoft Intune applying policies across managed environments.
So, Agent 365 is a management layer, while Entra, Purview and Defender provide the underlying identity, compliance and security capabilities.
Foundational capabilities are included within existing Microsoft 365 subscriptions. However, advanced governance features will require an Agent 365 licence which is available standalone or through Microsoft 365 E7. This is in line with Microsoft’s view that every organisation needs basic visibility, but larger estates require more sophisticated governance.
Technology alone isn’t enough
The technology isn’t likely to be the difficult part. In a mature Microsoft environment, with Entra, Purview, Defender, Intune and good identity management, deploying Agent 365 should be relatively straightforward. But getting governance and ongoing operational management right is far more demanding.
Governance
Your governance model must anticipate and answer a plethora of organisational and operational questions. For example:
- What types of agents are permitted?
- Who can create them?
- What approval process should new agents go through?
- What business data may they access?
- How are they tested before going live?
- Who owns each agent?
- When should agents be reviewed?
- What happens when an owner leaves the business?
- When should an agent be retired?
This can be a significant piece of work that demands plenty of thought – although a knowledgeable Microsoft partner can save you a lot of time by outlining good practices.
Operational reality
Agent 365 provides valuable operational insights. But someone will need to interpret what the platform reveals – and most organisations will significantly underestimate the time and effort involved. It’s likely to give rise to a new operational discipline that’s similar to a Security Operations Centre (SOC).
Just as many organisations don’t want to run their own 24×7 SOC, they’ll want assistance governing and managing their AI agents.
It’s not just a case of reviewing new agents, monitoring agent health and usage, or responding to security incidents that involve agents. Someone will need to continually determine if agents are still needed, have drifted from their original purpose, or have appropriate permissions. These aren’t one-off tasks – they’re operational processes.
AI is rapidly maturing to the point where it requires its own operational discipline, in much the same way that cloud computing led to CloudOps and FinOps.
Building AI agents is only the beginning. As organisations scale their AI workforce, success will depend on far more than technology. It will require governance, operational discipline and continuous oversight. Those that establish these capabilities early will be best placed to realise AI’s potential safely, securely and at scale.
When we ask, ‘who manages the AI workforce?’, it’s not Agent 365. People do.
Cloud Direct helps organisations put this into practice, with advisory, and management services. Use the form below to request a call with a subject matter expert and discover how Cloud Direct can help you manage your agent workforce.
