This is the final post in a three part series on scaling agentic AI with confidence:
- Part 1: Key Questions Boards Must Ask Before Scaling AI Agents: The six questions every board must answer
- Part 2: Seven Principles for Governing Agentic AI Successfully: The practitioner’s playbook
- Part 3 (this post): Mastering Governance and Architecture for Agentic AI Success. What happens when it’s not one agent, it’s 150,000
Part 1 asked the board level questions and part 2 provided practitioners with seven principles for governing individual agents. Both these posts assumed something that most organisations haven’t confronted yet: what happens when the number of agents outgrows your ability to govern them one at a time?
That’s the problem this post addresses. Not how to govern an agent, but how to govern an ecosystem of agents. Many of them built by people who’ve never read your governance policy, running on permissions nobody explicitly granted and connected to data sources nobody mapped.
If you have read an earlier post I wrote on Defining Enterprise Risk: Shadow AI, you may recognise the pattern;
- 1st Wave: Shadow AI. Employees adopting AI tools outside IT’s visibility (humans using ungoverned tools).
- 2nd Wave: Agent sprawl (ungoverned tools acting on their own).
The Scale Problem That Principles Alone Can’t Solve
The seven principles in Part 2 work when you’re deploying agents deliberately, one at a time, with architecture reviews, identity provisioning and graduated autonomy. This can be considered as the right approach, but It is also not what is actually happening in most enterprises.
Here is some research into what is happening:
- “13% of organisations believe that they have the right AI agent governance in place”, Source: IBM
- “By 2028, the average Fortune 500 enterprise will have over 150,000 AI agents in use, up from fewer than 15 in 2025¹. That’s a 10,000× increase in three years”, Source: Gartner
- “Surveying 2,600 business leaders across 13 countries, found that fewer than half of companies have a clear inventory of their AI agents. Not fewer than half have governance. Fewer than half can even count them”, Source: SAP
- “44% are already using agents, another 40% are experimenting, yet only 17% have visibility into agent performance or conformance, and 48% lack clear roles and responsibilities for managing agents”, Source: Inclusioncloud/SAP
- “76% of organisations are piloting or rolling out autonomous agents, and 42% have already had a confirmed or suspected AI related incident, Source: TechRadar
- “Agentic AI has reached technical viability in 2026, but most enterprises remain stuck between promise and payoff”, Source: Forrester
Part 2’s principles are necessary, however principles applied agent by agent don’t scale to 150,000. You need infrastructure that enforces them automatically. That’s the shift from governance as a practice to governance as a platform.
From Shadow AI to Shadow Autonomy
In my post Defining Enterprise Risk: Shadow AI, I described how shadow AI emerges when employees adopt tools faster than governance can follow. The pattern was predictable: a useful tool appears, someone starts using it, it spreads through word of mouth and by the time IT discovers it, it’s embedded in workflows nobody documented.
Agent sprawl follows the same trajectory with a critical escalation.
Shadow AI tools waited for a human to act. An employee used ChatGPT to draft a contract or fed customer data into an ungoverned model. The risk was real, but it was bounded by human speed and human attention. One person, one action at a time.
Agents don’t wait. They act on their own, with whatever permissions they’ve accumulated, across whatever data sources they’ve been connected to. The governance gap compounds autonomously.
This pattern can be seen through this article: “In a Fortune 50 pharmaceutical company, over 2,000 agent instances shared across the organisation, with 82% built by people who aren’t professional developers”, Source: Zentiy. These were business users solving real problems with the tools their platforms gave them. Nobody tracked what those agents could access, what decisions they were making or who was accountable when something went wrong.
This is not shadow IT, but shadow autonomy and it is the natural consequence of giving people agent building tools without giving them agent governing infrastructure.
The Enterprise AI Control Plane
The architectural response to agent sprawl is reffered to as the Enterprise AI Control Plane.
The seven principles describe how each agent should behave and the control plane is the infrastructure that ensures they do. This can be across thousands of agents, built by hundreds of teams, without requiring each team to implement governance from scratch.
BCG defines it as a unified layer that provides identity, policy enforcement, visibility and governance across the enterprise without slowing innovation. This is similar to what Kubernetes did for container orchestration. This time its for AI agents.
A six step framework published by Gartner for managing agent sprawl maps directly onto this:
- Establish agent governance and policies: Who can build agents, who can share them, what connectors are permitted
- Centralise agent inventories: A single registry of every agent, its owner, its permissions, its status
- Define agent identities and lifecycles: Provisioning, rotation, and deprovisioning, just like human users
- Govern information access: What data each agent can reach, under what conditions, with what audit trail
- Monitor agent behaviour: Runtime observability, anomaly detection, drift monitoring
- Foster a responsible AI culture: Governance isn’t just infrastructure; it’s organisational muscle
This framework takes principles from Part 2 (identity, layer separation, output assurance, graduated autonomy) and implements them as platform capabilities rather than per agent decisions.
- The team building an agent doesn’t need to design its own audit trail as the control plane provides one.
- They do not need to implement their own kill switch as the control plane enforces circuit breakers.
- They do not need to manage their own credentials as the control plane handles token injection at the action layer.
This also helps to address the shadow AI problem. Shadow AI thrives in the absence of a governed alternative. When building an agent through the control plane is as easy as building one outside it, but with identity, logging and policy enforcement built in, the incentive to go rogue disappears.
The best way to eliminate shadow autonomy is not to ban it, but to make the governed path faster and easier than the ungoverned one.
“CIOs who stop owning delivery and start owning the foundations (including data architecture and the agent governance model) can create an environment where citizen development is safe, consistent, and an engine for progress, not costs”. Source: BCG
The CIO’s job is not to build every agent, but should be to build the platform that makes every agent governable.
The Identity Crisis at Scale
Part 2 argued that every agent should be treated like a new employee, with its own identity, scoped permissions and a probation period. This principle is correct, but scaling to 150,000 agents, it exposes a gap that most enterprise IAM (Identity and Access Management) architectures were not designed to handle.
“Traditional IAM was designed for a human user who authenticates once and operates within a session”, Source: TechRadar
Agents don’t work that way as they authenticate, then make choices, invoke tools and change systems continuously. Legacy IAM validates access at a point in time. Agents need access evaluated in context as actions occur.
In Defining Enterprise Risk: Shadow AI, I argued that one of the core risks of shadow AI is that it operates under borrowed identity:
- An employee’s credentials
- A shared service account
- A team API key
Agent sprawl inherits these and amplifies. When 2,000 agents share a handful of service accounts, you don’t have an identity architecture, you have an attribution crisis.
When something goes wrong, you can’t trace it, you can’t scope it and you can’t stop it without stopping everything.
The control plane solves this by making identity a platform service:
- Agent registration: Every agent gets a unique, non human identity at creation, automatically
- Scoped, time bound permissions: Least privilege with automatic expiry, enforced by the platform, not by the developer
- Short lived credentials: No long lived secrets; tokens issued at the moment of action and revoked immediately after
- Continuous behavioural evaluation: Not just at authentication, but at every action, with anomaly detection built into the platform
“In sectors such as health care and banking, companies must maintain interoperability, security, monitoring, and governance across teams and use cases. Without a common platform, AI efforts become fragmented, harder to govern, more expensive to scale, and slower to deliver impact”, Source: BCG
The identity layer is the foundation everything else depends on. This is the difference between Part 2 and Part 3.
- Part 2 says “treat every agent like a new employee.”
- Part 3 says “build the HR system that onboards 150,000 of them.”
The Economics of the Control Plane
This is not just a risk conversation but one of value.
- “Enterprises leading in agentic AI are cutting costs by 15% to 20% across functions. Their approach prioritizes speed first, growth second, and cost third”, Source: BCG
- “42% of companies expect their agents to act autonomously by 2030, deciding without human approval, and just 5% have the relevant agent controls in place”, Source: BCG
This indicates that many organisations are scaling towards autonomous agents without the infrastructure to govern them.
To be clear about what this measures: 5% with relevant controls, here BCG is counting organisations with the complete set of critical controls (identity, observability, policy enforcement and lifecycle management) all operational. This is distinct from the Deloitte figure cited in Part 2, which found 5% of organisations with business processes highly prepared for AI agents. Same number, different lens. Deloitte measures process readiness. BCG measures control infrastructure. Both of these sources arrive at the same conclusion that almost nobody is ready.
“By 2028, 70% of enterprises will abandon agentic AI built by vendor forward deployed engineering (FDE), trapped by soaring costs and unable to evolve it on their own”. Source: Gartner
The lesson here is:
- If you do not own the architecture, you do not own the governance.
- If you do not own the governance, you do not own the outcome.
The control plane is the investment that separates the 5% from the 95%. Rather than this being a cost centre it is the infrastructure that can make a cost reduction possible providing safety at scale and without indicents that erode board confidence.
Five Things to Do on Monday Morning
This series has covered board level questions, practitioner principles and the platform that makes both enforceable. Here’s how to start:
1. Find your agents. If you have not addressed the shadow AI baseline from Defining Enterprise Risk: Shadow AI, start there.
- You cannot govern agents you haven’t found.
- Build a simple registry: what agents exist, who owns them, what can they access, when were they last reviewed?
- If the answer to any of those is “we don’t know,” that’s your first priority.
2. Stand up the control plane. Start with identity and observability. You don’t need the full platform on day one. Start with two things:
- Every agent gets a unique identity
- Every action gets logged.
Those two capabilities alone will transform your ability to investigate incidents and demonstrate compliance.
3. Audit for single loop agents. Apply Part 2’s layer separation principle.
- If any agent both decides and executes with no independent control layer, flag it.
That’s your highest risk deployment. The control plane should enforce this separation as a platform constraint, not a team by team decision.
4. Classify agents by consequence. Not every agent needs the same governance overhead. Use Part 2’s graduated autonomy model as your classification framework:
- Low consequence agents move faster through the dial
- High consequence agents stay at HITL (Human in the Loop) until trust is empirically established.
The control plane should enforce the classification, not just document it.
5. Measure platform outcomes, not agent counts. Stop counting how many agents you’ve deployed. Start measuring what the control plane delivers:
- Mean time to detect an agent anomaly
- Percentage of agents with unique identities
- Percentage of high consequence actions with human approval
- Time from agent creation to full governance compliance.
These are the metrics that tell your board whether the infrastructure is working.
Closing the Series
Across these three posts, I’ve argued a single thesis: Agentic AI is the most consequential shift in enterprise technology since cloud computing and the organisations that will lead are not the ones moving fastest. They are the ones whose governance and architecture are inseparable.
- Part 1 asked the six questions every board should answer before scaling agents, authority, accountability, auditability, reversibility, measurement and architectural readiness
- Part 2 delivered the seven principles that translate board level confidence into production grade governance for individual agents
- Part 3 confronted the scale problem: agent sprawl is already here, principles applied one at a time don’t scale and the Enterprise AI Control Plane is the infrastructure that makes everything else work
The journey from shadow AI to agent sprawl was predictable.
- Shadow AI was the warning shot, employees adopting tools faster than governance could follow.
- Agent sprawl is what happens when those tools start acting on their own.
- The organisations that took shadow AI seriously built the governance muscle.
- The ones that didn’t are now facing the same problem at 10,000× the scale, with autonomous actors instead of human ones.
The gap between the leaders and the rest isn’t technology, it is:
- Architecture
- Governance
- The decision to treat AI agents as actors in your operating model
Not tools in your technology stack.
Close that gap. Your agents won’t wait for you.
The Full Series
- Part 1: Key Questions Boards Must Ask Before Scaling AI Agents: The six questions every board must answer
- Part 2: Seven Principles for Governing Agentic AI Successfully: The practitioner’s playbook
- Part 3 (this post): Mastering Governance and Architecture for Agentic AI Success: Why governance without architecture is just policy on paper
References
- “AI agent sprawl: What it is and how to control it”, IBM
- “Gartner Identifies Six Steps to Manage AI Agent Sprawl.” Gartner
- “AI Agent Sprawl: Why AI Governance Is Now a Board Level Issue”. SAP
- “SAP LeanIX Takes on the Growing AI Sprawl Problem”. Inclusioncloud
- “Why agentic AI demands a new approach to enterprise security”. TechRadar
- “The State Of Agentic AI In 2026: Companies Are Chasing, Few Are Catching”. Forrester
- “Your AI Agent Inventory Is Incomplete. Here’s What That Means for Risk”. Zenity
- “Enterprise AI Control Plane: The CIO’s Guide to Governing and Accelerating AI Agents”. BCG
- “Gartner Identifies Six Steps to Manage AI Agent Sprawl”. Gartner
- “When Your CRM Starts Building Itself: The CIO’s Role in the Agentic SaaS Era”. BCG
- “When identity isn’t human: securing the agentic enterprise”. TechRadar
- “Building Enterprise AI Agents in Regulated Industries”. BCG
- “The Agentic Leadership Playbook: A Scaling Strategy for CTOs and CIOs”. BCG
- “What the Applied AI Index Reveals About AI Maturity and AI Value”. BCG
- “Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward Deployed Engineering by 2028”. Gartner
Leave a Reply