Most enterprises govern their AI agents with a single set of rules. Every agent passes the same review, carries the same approval steps and is monitored in the same way, regardless of what it actually does. The intent is consistency. The result is a regime that fits almost nothing. An assistant that drafts a meeting summary is subjected to controls designed for an agent that can move money, and an agent that can move money is governed by rules calibrated for an assistant. The first is slowed down for no gain in safety. The second is under-governed in the place it matters most.
Guidance shared at Gartner’s IT Symposium on October 20 points in a different direction. Governance, Gartner’s analysts argued, should evolve iteratively rather than be fixed up front. Organisations should collect evidence now, ahead of the external scrutiny that is coming. They should build layered guardrails rather than rely on any single control. And oversight should be tailored to each agent’s level of autonomy, with agent implementations working best where reliability and security keep pace with the autonomy granted. Gartner also predicted that by 2028, 25% of enterprise security breaches will be traced to AI agent abuse — a risk rooted in unmanaged machine privileges.
The common thread is proportionality. Oversight should scale with autonomy. This blog is for governance, risk and compliance leaders, and for the technology owners who work with them, who need a governance model that is strong where it must be and light where it can be.
Why A Single Regime Fails In Both Directions
A uniform regime fails first by being too tight. When every agent faces the same approval process, low-risk, high-value use cases queue behind high-risk ones. Teams experience governance as delay, and some begin to route around it — which produces exactly the ungoverned agents governance exists to prevent.
It fails second by being too loose. A regime calibrated to be workable for the many low-risk agents cannot also be demanding enough for the few agents with real authority: those that issue documents, change records, approve transactions or act across several systems. Those agents need stronger identity controls, more evidence and faster intervention, and a uniform regime does not provide them.
It fails third on evidence. Regulators, auditors and boards increasingly want to see how an agent’s behaviour was controlled, not only that a policy existed. A single regime tends to produce the same thin evidence for every agent, when the agents that carry real consequence need a far richer record.
Gartner’s 2028 prediction sharpens the point. Breaches traced to AI agent abuse are, at root, a privilege problem: agents holding broader, longer-lived access than their task requires. Proportionate governance addresses that directly, because access is scoped to the tier.
A Four-Tier Autonomy Model
The first tier is assistive. The agent reads, summarises, drafts and recommends; a human takes every action. Oversight is light: registration in the agent inventory, a named owner, read-only access scoped to the data it needs, and standard logging. Approval to deploy should be fast, because the risk is contained by design.
The second tier is supervised action. The agent prepares actions — a draft invoice, a record update, a customer reply — and a human approves each one before it executes. Oversight adds approval checkpoints, logging of every approval decision, and measurement of how often humans amend or reject the agent’s proposals. That rejection rate is the evidence that tells the organisation whether the agent is ready for more autonomy.
The third tier is bounded autonomy. The agent acts on its own within defined limits — value thresholds, document types, customer segments — and escalates anything outside them. Oversight shifts from approving every action to sampling actions, monitoring behaviour against expected patterns, and reviewing every escalation. Identity controls tighten: credentials are scoped to the specific actions permitted and are time-limited.
The fourth tier is delegated autonomy. The agent acts across multiple systems, sequences its own steps and handles material decisions within its mandate. Oversight here is continuous: real-time behavioural monitoring, layered guardrails at the model, tool and data levels, an immediate means of suspending the agent, and the most complete evidence record in the estate. Few agents should sit here, and each should have earned its place through evidence gathered at the lower tiers.
Two rules make the model work. Promotion between tiers is earned with evidence, not granted by request. And demotion is normal: an agent whose behaviour drifts, or whose environment changes, moves down a tier until confidence is restored.
The Gulf Compliance View
The tiered model maps naturally onto the obligations Gulf enterprises are preparing for. An agent that summarises supplier correspondence is assistive. An agent that issues tax invoices is not. Under ZATCA’s integration requirements in Saudi Arabia, and under the UAE’s mandatory e-invoicing regime from January 1, 2027, invoices are structured, validated and transmitted records with defined consequences for errors and delays. Any agent involved in creating or transmitting them belongs, at minimum, in the supervised tier, with evidence that would satisfy a tax audit.
Gartner’s advice to collect evidence now is particularly relevant here. The time to establish approval logs, exception records and audit trails for invoicing agents is before go-live, not after the first query from the tax authority. Data residency belongs in the same design: the evidence record for regulated processes should be retained in-region, under access controls proportionate to the tier.
What Inspires Us At Lynt-X Global
What inspires us about this guidance is its pragmatism. For a long time the governance debate has been framed as a trade-off between speed and safety. Proportionate oversight dissolves much of that tension: light controls where risk is contained, rigorous ones where agents carry real authority, and evidence that allows trust to grow step by step. It treats governance as something that enables progress rather than something that holds it back.
We are also encouraged by the idea of autonomy as something earned. It mirrors how organisations develop people — responsibility expands as a track record builds — and it gives teams a clear, fair path for bringing more capable agents into production.
In the Gulf, the regulatory environment makes this especially timely. National e-invoicing programmes are establishing a culture of structured, auditable records, and that culture is a natural foundation for evidence-based agent governance. We expect the region’s enterprises to be well placed to adopt tiered oversight early.
Looking ahead, we are watching how AI agent identity matures as a discipline — Gartner lists it among its strategic technology trends for 2027 — how auditors and regulators begin to ask for agent evidence, and how layered guardrails become standard practice rather than bespoke engineering.
The Compliance Read
A single governance regime for every agent is either too loose or too tight, and often both at once. Gartner’s Symposium guidance — iterate, collect evidence now, layer guardrails, tailor oversight to autonomy — points to proportionality, and its prediction on breaches traced to agent abuse shows why scoped privileges matter.
A four-tier model gives that principle structure: assistive, supervised action, bounded autonomy and delegated autonomy, with oversight, identity controls and evidence rising at each tier, promotion earned through evidence and demotion treated as routine. For Gulf enterprises approaching e-invoicing go-live, the model also clarifies where the strongest controls belong. Oversight that scales with autonomy lets low-risk agents move quickly and keeps high-autonomy agents under the governance they require.
“Governing every agent the same way slows the assistants that pose little risk and under-governs the agents that can act on their own. Oversight should scale with autonomy: light for agents that recommend, approval checkpoints for agents that prepare actions, sampling and scoped credentials for agents that act within limits, and continuous monitoring for the few trusted to act across systems — with every promotion earned through evidence.”