Eero Nevaluoto
Senior Engineer, Intelligent Group
Eero is a certified cybersecurity professional (Azure Security Operations Analyst, CompTIA CySA+, Security+) who specializes in threat detection and compliance automation.
Agentic AI — AI that does not just answer but acts, calling tools, moving data, and executing multi-step tasks with limited human intervention — is moving into production faster than most organizations can govern it. Gartner projects that 40% of enterprise applications will embed AI agents by the end of 2026, up from less than 5%. The capability is arriving. The controls are not.
We build and operate agents in our own stack, so this is not a spectator take. The single most under-managed risk in enterprise agentic AI is not the model — it is the identity the agent runs as.
What makes agentic AI different from a chatbot?
An agent is different because it takes actions in your systems on its own initiative. A chatbot returns text a human reads and decides what to do with. An agent reads a ticket, queries a database, updates a record, sends an email, and calls an API — a chain of privileged actions with no human in the loop for each step. That autonomy is the value, and it is also the exposure.
Why is non-human identity the core risk?
Non-human identity is the core risk because every agent needs credentials, and those credentials are routinely over-permissioned. To ship fast, teams grant an agent broad service-account access, standing API keys, and OAuth scopes far wider than any single task requires. The result is a machine identity with more privilege than the humans it works alongside, no MFA, and often no expiry — a high-value target that never sleeps and rarely gets reviewed.
Non-human identities already outnumber human ones in most enterprises, and agentic AI multiplies them. Each agent, sub-agent, and tool integration is another credential to inventory, scope, rotate, and monitor. Treat them like the privileged accounts they are, or they become the softest path into your environment.
What is the confidence-versus-readiness gap?
The confidence-versus-readiness gap is the distance between how many organizations are deploying agents and how few are actually operating them safely. Forrester has found that only 14% of companies are successfully scaling agentic AI. Gartner goes further, projecting that more than 40% of agentic AI projects will be cancelled by 2027 — driven by escalating costs, unclear value, and inadequate risk controls. Enthusiasm is running well ahead of operational maturity.
That gap is exactly where incidents live: agents pushed to production on a proof-of-concept’s permissions, with no logging of what they did, no kill switch, and no owner accountable for their behavior.
How do you govern agentic AI in production?
You govern agentic AI by treating each agent as a scoped, monitored, revocable identity — the same discipline you already apply to privileged human access, extended to machines:
- Least privilege by default. Scope every agent’s credentials to the minimum tools and data its task requires. No standing admin, no wildcard scopes.
- Short-lived, rotated credentials. Use ephemeral tokens and workload identity over long-lived API keys. An agent’s key should expire on its own.
- Full action logging. Every tool call, data access, and external write the agent makes must be logged and attributable, so you can reconstruct what happened.
- Human-in-the-loop gates. Require approval for high-impact actions — financial transactions, data deletion, external communication.
- A kill switch and an owner. Every production agent needs a named human owner and a one-click way to revoke its access.
Where does this fit in the broader AI-risk picture?
Agentic identity is one pillar of a larger program. It sits directly on top of your AI governance framework, it is a first cousin of the shadow AI problem (an unsanctioned agent is shadow AI with hands), and it depends on the data readiness that determines whether the agent’s actions are trustworthy in the first place.
Govern your AI agents before they scale
We inventory your non-human identities, right-size their permissions, put logging and kill switches in place, and wire agent governance into your broader AI program — so you scale agents on controls, not on hope.
The bottom line
Agents are coming to production whether or not the controls are ready — Gartner’s 40%-by-2026 number all but guarantees it. The organizations that avoid the cancellation pile are the ones that governed the non-human identity first and scaled the autonomy second.