AI agent adoption is accelerating but production value depends on multiple factors. You need data quality, solid integration, identity management, governance, and continuous control to ensure your agents work for your business and not just in it. Microsoft recorded a 15-fold year-on-year growth in active agents across its Microsoft 365 ecosystem – 18-fold in large enterprises – and the company’s research has doubled down on the need for entities to be correctly managed with the right controls.
AI agent readiness comes down to the state of your data, identities and access controls. Gartner predicted in 2025 that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025. Many of these agents are arriving inside software you already license, and each one will work with whatever data and permissions it is given.
This article unpacks what you need to put in place before your agents reach your production workflows and how a zero trust AI architecture keeps them in check.
Key takeaways:
· Active agents in the Microsoft 365 ecosystem grew 15-fold year on year, and 18-fold in large enterprises, between March 2025 and March 2026.
· Gartner predicts that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025.
· Nearly 80% of organizations say they can’t share data across teams in ways that make agentic AI work. Only about one in four strongly agree that they have clearly defined owners maintaining their knowledge sources.
· Zero trust applies verify explicitly, least privilege and assume breach to AI agents, and role-based access control for AI limits each agent to the task it was built for.
· Multifactor authentication protects the people who build, approve and use agents. Autonomous agents authenticate through the client credentials flow and can’t complete a second-factor challenge, so they need tightly governed credentials, narrowly scoped permissions and continuous monitoring.
What do enterprises need before deploying AI agents at scale?
Before agents go into production, you need to put three key things in place. The first is data and ensuring it is both classified and current; the second is permissions, ensuring they match what each person and agent requires; and the third is a register of every agent in use. These are the data foundations for AI agents, and they are essential because, without them, an agent will do exactly what it was built to do using information it should never have seen.
Your data also needs to stop being a production constraint. Agents cannot reliably execute cross-functional workflows if your data is siloed, stale or ownerless. Microsoft found that 80% of companies surveyed couldn’t share their data across teams in ways that support agentic AI, and only one in four agreed that clearly designated owners kept their knowledge sources current and reliable.
Agents raise the data stakes because they act on that data. They can access high-level systems and accidentally share data that isn’t theirs to share – think access to financial records and expense reports with no controls over who can speak to the agents and use that access. They can sprawl – Microsoft warns of agent sprawl where business units build agents without IT oversight and temporary agents stay in production with permissions nobody reviews. For a South African business, an agent with that level of reach into HR or customer records is also running the risk of agents processing and sharing personal information without the right levels of security and privacy.
How does zero trust support safe AI adoption?
Zero trust treats every agent as an identity that must prove who it is, justify what it can reach and be monitored for signs of compromise. The three principles that govern human access – verify explicitly, use least privilege and assume breach – apply to AI agents in the same way.
Role-based access control for AI is the first step towards managing least privilege. For example, an AI agent that reconciles supplier invoices needs read access to invoices and the vendor master file, but it doesn’t need to open payroll. When you scope each agent to a role and review that role when the task changes, you can then lock a compromised or misbehaving agent in one small corner of the business in the event of a worst-case scenario.
Multifactor authentication (MFA) is another key part of the process, but it must be applied differently to people and autonomous agents. MFA protects the people who build, administer, approve and use agents, however, when an agent operates on behalf of a signed-in user, tools like Microsoft’s Entra Conditional Access, evaluate the user’s identity and apply relevant user policies, including MFA requirements but the agent itself Is not directly subject to these policies. The agent authenticates using its own agent identity through the client credentials flow and because they don’t authenticate with passwords, passkeys or authenticator apps, they cannot complete an interactive second-factor challenge.
AI identities require tightly governed credentials, narrowly scoped permissions, risk-based access policies and continuous monitoring. Zero trust doesn’t mean forcing the same authentication control onto every identity, it means verifying each identity appropriately and consistently reassessing access.
How can your business benefit from agentic AI?
Start narrow to reduce the risk of escalating costs, unclear business value and inadequate risk controls. And remain skeptical of what the market has on offer. There are thousands of vendors marketing agentic AI solutions, but very few are the real thing. Before you sign, ask what an agent can do without a human in the loop, what data it needs and what permissions it requires. The answers to these questions will immediately cut out agents that don’t align with your security and with what you need your agents to do within the business.
The strongest initial use cases are not necessarily purely rules based, they are bounded, high-volume and measurable but complex enough to benefit from an agent’s ability to interpret context and choose between permitted actions. Purely repetitive and predictable tasks may remain better suited to conventional automation, so for agents, define your baseline before deployment and measure the same indicators afterwards to gain credible evidence for deciding whether to improve, scale or retire an agent.
Agents are multiplying in the business making it essential to prioritize AI agent readiness. Close the gap before an agent reaches production workflow with data that has named owners, a register of every agent in use, roles scoped to each task, MFA for the people who build and approve agents, and governed credentials for the agents themselves. For agentic AI in the enterprise, companies setting a baseline and measuring against it can retire poor performers and prioritize high performers while still scaling adoption securely.