AI agents are multiplying faster than IT can track them. Here's what to budget for in 2027.

The 2027 budget season is underway, and AI agents have secured a line item at most organizations. What's missing is equally important: a budget for visibility and control over those same agents.

The problem is already visible. Agents are entering enterprises through multiple channels simultaneously—some built by internal engineering groups, others activated within existing SaaS platforms, and still others launched by individual employees using nothing more than a web browser and a credit card. Each agent receives access credentials. Each agent connects to systems. Many operate without human approval for individual actions.

IT leaders recognize this pattern. The critical difference now is that the entities proliferating don't simply read or write data. They make autonomous decisions and execute actions based on that data.

Round three of a familiar pattern

The first wave was SaaS adoption. Business units deployed applications faster than IT could evaluate them, and visibility typically arrived only when contract renewals, breaches or compliance audits forced a reckoning.

The second wave involved APIs. As organizations decomposed monolithic systems and integrated with external partners, the number of connections exploded beyond what any team could manually track. Organizations that achieved control followed a consistent playbook: establish a single control point, maintain a registry of all integrations, assign clear responsibility and enforce rules centrally rather than negotiating them repeatedly across teams.

Both experiences revealed three consistent principles:

1. Governance requires visibility first. Continuous inventory matters more than periodic audits.
2. Enforcement mechanisms must sit in the access path, not in policy documents. Rules embedded in PDFs are routinely bypassed. Rules enforced at the point of access are followed.
3. Every asset requires an owner. Unmanaged tools and orphaned endpoints become repositories for risk.

Agents present all these challenges plus an additional layer of complexity. Unlike SaaS applications or APIs, which execute predetermined instructions, agents determine their own next steps. This means governance must address not just "what systems does this connect to" but also "what actions is it permitted to take, and can we document what it actually did."

Five questions every agent should answer before production

A straightforward evaluation: an agent that cannot address these five questions should not be deployed to production systems.

1. Who is it? Each agent requires a distinct identity independent from the user or service account that created it. Using shared credentials makes it impossible to determine which agent performed which action.
2. What can it reach? Specify which systems, data stores and environments the agent may access, and enforce these restrictions at the access layer. The principle of least privilege applies to agents as much as to human users.
3. What can it use? Agents operate through tools—APIs, connectors and, increasingly, other agents. Each tool must be vetted, rate-limited and capable of being disabled, following the same protocols IT uses for API governance.
4. What is it for? An agent with a specific, documented mission is substantially easier to control than one designed for general purposes. Its function should identify a responsible human sponsor and establish clear escalation points.
5. What did it do? Every action must generate an audit trail that security teams, auditors or regulators can examine. If an action cannot be reconstructed and explained after it occurs, it cannot be defended.

These questions are not novel. CIOs routinely ask similar versions about employees, external contractors and software systems. The novelty lies in applying them to autonomous software operating at scales that no manual review process could handle.

Incorporating agents into the 2027 plan

The encouraging aspect is that this work aligns with budget discussions already occurring. Four concrete steps merit immediate attention:

• Allocate resources for agent discovery before deployment accelerates. Budget for identifying agents already in operation, including those embedded within vendor applications. Most enterprises will uncover more than anticipated.
• Establish the five questions as a prerequisite for launch approval. Incorporate them into the existing approval workflow for new applications and integrations. Agents unable to satisfy these questions should not proceed to production.
• Direct agent communications through existing governance infrastructure. Since agents communicate via APIs, extending current gateways and access controls to encompass agent traffic is more efficient than constructing separate control systems.
• Assign ownership to each agent. Connect every agent to a business sponsor and a technical owner, and review assignments with the same frequency applied to privileged access reviews.

These measures also strengthen the business justification. When leadership or audit teams inquire about agent governance, "we can demonstrate it" carries substantially more weight than "we have documented policies."

Seizing the opportunity in round three

SaaS and API proliferation were largely managed retroactively, after the consequences of inadequate oversight became apparent. Agents present an unusual opportunity to establish governance proactively. CIOs who establish control frameworks in their 2027 plans will navigate agent scaling with assurance throughout the year. Those who postpone will spend 2027 conducting inventory.

Rory Blundell is CEO of Gravitee, which helps enterprises manage, secure and govern AI agents.