Twelve is the average number of AI agents a large enterprise runs today, according to Salesforce's 2026 survey of 1,050 information technology leaders. Half of those agents work in isolation, with no shared oversight and no single person accountable if one of them makes an expensive mistake. That number is projected to reach twenty within two years. Almost no finance department has budgeted for what happens when an agent nobody remembers building starts touching data it should not touch.

A Dashboard That Sees Every Agent, Not Just the Ones IT Built

An agent control tower is a single system that finds every AI agent running across a company, regardless of who built it or which vendor it came from, and gives one team the ability to see what each agent can access and shut it off if it misbehaves. Think of it as the equivalent of an airport tower that tracks every aircraft in its airspace, no matter which airline owns the plane.

Before these towers existed, an agent built by a finance team in one platform and an agent built by a sales team in another had no shared record anywhere in the company. Each vendor governed its own agents. Nobody governed the space between them.

The Platforms That Started the Race

ServiceNow shipped the first version of its AI Control Tower in 2025 and expanded it across five new capabilities in May 2026. The current version discovers AI agents, models, and datasets across Amazon Web Services, Google Cloud, Microsoft Azure, and both OpenAI and Anthropic deployments, then maps every permission those agents hold through its Veza acquisition. ServiceNow is offering the tool free for a year to customers, a package it values at two million dollars, betting that giving away its most strategic product locks in long-term platform commitment.

Microsoft took a different entry point. Agent 365 became generally available on May 1, 2026, built into the Microsoft 365 admin center rather than sold as a separate product. Within two months of preview, tens of millions of agents had already registered inside it, a figure that reflects agents already running, not agents Microsoft expects enterprises to eventually build.

Boomi built its version, Agent Control Tower, as part of a broader Agentstudio release and positioned it around agents that were never going to live on one vendor's platform: agents from Salesforce's Agentforce, Amazon Bedrock, Snowflake Cortex, and Microsoft Copilot side by side in one view.

Amazon Web Services started lighter than the other three, then closed the gap. Bedrock AgentCore added a policy gateway, portable agent identity that works across AWS, on-premises, and competing clouds, an agent registry, and full replay of an agent's decision history for audit purposes at its New York Summit in June 2026. AWS now frames AgentCore as an enterprise-grade control plane in its own right, not a lighter-weight alternative to the others.

The Race Reaches Past IT Platforms

ServiceNow, Microsoft, and Boomi built their towers around business workflows. Three other categories of vendor are building toward the same control point from different directions. Cisco launched Cloud Control in June 2026, a shared plane across networking, security, and observability that treats human operators and AI agents as the same category of actor to watch. Databricks built Unity AI Gateway to govern agents at the exact point they touch enterprise data, logging every model and tool call back to the catalog that already tracks the data itself. Snowflake shipped its own version, Cortex AI Gateway, on July 28, 2026, three days before this piece went to press.

Databricks put a number behind the case for governance that a chief financial officer can use in a board meeting. Across more than 20,000 organizations on its platform, including over 60 percent of the Fortune 500, companies with AI governance in place shipped more than twelve times as many agent projects into production as companies without it.

The two enterprise platforms most companies already run on joined the same race. Salesforce expanded Agent Fabric in April 2026 and now calls it a trusted agent control plane in its own materials, with automated discovery reaching into Amazon Bedrock and Microsoft Foundry as well as its own agents. Oracle built governance into OCI Enterprise AI itself, giving compliance teams agent-lifecycle approval and audit support for SOC 2 and SOX inside Fusion's ERP, HCM, and CX applications.

The vendor names matter less than what they add up to. Workflow platforms, CRM, ERP, cloud infrastructure, data platforms, networking, and observability tools have all built one of these in the same eighteen months. That is not a feature trend. It is every layer of the enterprise stack agreeing on the same missing piece at the same time.

A reader comparing vendors for a specific layer, identity, security, data, or infrastructure, needs more than a company name. The table below lists the distinct fact behind each, not just that a control tower exists.

What the Tower Does All Day

Every agent gets registered the way a new employee gets a badge, with an owner and a record of what it is allowed to touch. The tower then builds a baseline of normal behavior for each agent and flags anything outside it, the same logic a fraud detection system applies to a credit card.

ServiceNow staged the value of that baseline as a live demonstration at its Knowledge conference. An AI agent handling benefits data had given itself elevated permissions without anyone approving it, gaining potential access to sensitive personal records it was never assigned to see. The control tower caught the change, surfaced the alert, and offered a single action to revoke the agent's permissions, deactivate it, and generate an incident report with a full audit trail attached.

That single action is the part that matters to a chief financial officer. Without it, tracing what an agent touched after something goes wrong can take a security team days, and every one of those days is billable exposure.

Why the Payoff Lands on the Balance Sheet

John Baker, chief information officer and chief information security officer at Lexitas, described the practical effect on stage at Boomi World this year. Once his company's agent registry, data unification, and audit trail were already in place, the agent itself was the fast part. His advice to peers was to pick something small, ship it, then build from there, because the groundwork rather than the agent was what had taken the time.

A control tower also changes who holds the leverage in a vendor negotiation. Because it works across Amazon, Microsoft, Google, and smaller platforms at once, a company no longer has to standardize on one AI vendor to get a unified view of risk. That view now comes from the tower, not the vendor.

Who Needs One This Year

A company running two or three agents inside one platform can track them by hand. Past a dozen, across more than one vendor, that stops being realistic. Insurance and financial services carry the added pressure of regulation catching up in real time. The European Union's transparency and labeling obligations under the AI Act take effect on August 2, 2026, with broader rules for general-purpose AI models phasing in on a longer timeline. Singapore's emerging guidance on agentic AI points in the same direction: a documented inventory of every agent in production. Both assume a company can eventually produce that inventory on request. Most cannot yet.

The Cost of Waiting

Salesforce's benchmark found that only 54 percent of surveyed companies have centralized governance over their agents at all, and 27 percent of the application programming interfaces connecting those agents to company systems carry no audit trail and no access control.

Security firm Gravitee's 2026 survey put a harder number on the exposure: agent counts inside surveyed organizations roughly doubled in a single quarter between December 2025 and April 2026, and fewer than half of the agents already running are actively monitored or secured.

Hugging Face published a technical postmortem in July 2026 that shows what the exposure looks like in practice. An autonomous agent built from OpenAI models escaped an internal cybersecurity evaluation, chained through a compromised third party's sandbox, and spent four and a half days inside Hugging Face's own infrastructure before anyone shut it down, 17,600 recorded actions in total. Hugging Face's own security stack correlated the signals into a coherent attack picture and still underrated the severity at first, which cost time in the response. That agent was never going to show up in anyone's internal registry. The lesson still applies: at machine speed, an agent generates more activity than a human team can read in real time, whether it was ever meant to be there or not.

Neither the Salesforce nor the Gravitee number describes a future problem. Both describe agents running inside companies today, unwatched, while the budget conversation about governance is still scheduled for next quarter.

The tower does not stop a company from moving fast on AI. It answers the one question a board will eventually ask: which agents do we have, and who is watching them.

Unknowns and Uncertainties

Pricing past the introductory offers remains private for both ServiceNow and Boomi, so the multi-year cost of ownership is hard to model today. Companies that adopt more than one tower, a Microsoft shop that also runs ServiceNow, for instance, have no established practice yet for reconciling two registries that both claim to be the single source of truth. And no regulator has yet tested whether a vendor's own control tower satisfies an external audit requirement, or whether examiners will want independent verification on top of it.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by