How Many AI Agents Is Your Company Running? AI Agent Governance Starts With an Inventory

Ask a company how many servers it runs and the IT lead can usually answer on the spot. Ask how many AI agents it is running, and what data each one can touch, and you will mostly get silence. That was the point Dataiku CEO Florian Douetteau made when launching a new product in late September. From assistants that draft email replies, to automations wired into spreadsheets, to AI tools that can read and write shared drives, AI agent governance is no longer a problem only large enterprises face. It is a basic discipline every brand adopting AI needs to put in place.

How many AI agents are running in your company?

An AI agent is an AI system that can call tools, read data and even take actions on its own. Agents tend to arrive quietly: a marketer connects an AI writing tool to the shared drive, a salesperson links the CRM and inbox through an automation platform, an engineer lets an AI assistant read the codebase. Each decision makes sense on its own, but together they form a web nobody has seen in full.

IBM research found that fewer than one in five organizations maintain a complete, current inventory of their AI systems. On 24 September 2026, Dataiku announced a standalone product called Agent Management, built to bring agents scattered across platforms such as AWS, Google, Microsoft and Salesforce into one inventory, recording each agent's owner, purpose, tools and models, cost and quality. When even specialist AI platforms start treating the inventory itself as a business, the problem is too widespread to ignore.

The cost of losing control: one "Allow All" grant

In April 2026, the cloud platform Vercel disclosed a security incident. According to post-incident analyses from several security firms, it began when an employee signed up for a third-party AI office tool with a company account and chose "Allow All" when granting access to Google Workspace. When that AI tool company was itself compromised, the attacker used the grant to get into the employee's account and moved on from there into Vercel's internal systems.

This is not an isolated case. A survey published in April 2026 by the Cloud Security Alliance and Token Security found that 65% of organizations had experienced at least one security incident caused by AI agents in the past year, and 61% of those incidents involved exposure of sensitive data.

The core issue is not AI itself but access: whatever an agent holds, an attacker inherits.

Four basics of AI agent governance

In Notion + make.com: the AI era we looked at the efficiency automation tools bring. To keep that efficiency from turning into risk, these four practices come first:

1. Build an inventory: record each agent's name, owner, purpose, and which accounts and data it connects to. An agent without an owner is an agent nobody will ever switch off.
2. Least privilege: if read-only will do, do not grant write access. If one folder is enough, do not open the whole drive. "Allow All" on a consent screen should be the exception, not the default.
3. Review third-party grants regularly: admin consoles for Google Workspace, Microsoft 365 and similar tools show which outside apps hold access. Clean them up every quarter and revoke anything no longer in use.
4. Keep logs and a kill switch: know what each agent has done, and know how to disable it immediately when needed.

Where smaller businesses can start: one spreadsheet

Not every company needs to buy an enterprise agent-management platform. For most brands and small businesses, a shared spreadsheet is enough for step one, with columns for agent name, owner, purpose, connected accounts, permission scope and last review date. Getting the current picture on the table matters more than choosing a tool.

The division of labor is especially clear here:

• AI handles: running repetitive work, organizing data, completing tasks within the access it has been given.
• People handle: deciding which work can go to AI, how much access to grant, who is accountable for the result, and coming back to review it regularly.

Conclusion

The more useful AI agents become, the more easily they accumulate access and risk without anyone noticing. Mature AI adoption is not about how many tools you use but about knowing exactly what each one is doing. AI delivers perfection; people deliver emotion, and oversight. If you want to design an AI workflow for your team that is both safe and genuinely useful, explore TitApeX services or talk to us.

References

• Dataiku Launches Agent Management for enterprise AI agents (Dataiku)
• Dataiku Agent Management reveals unmonitored AI agents (Help Net Security, September 2026)
• Dataiku debuts cross-platform Agent Management (SiliconANGLE, September 2026)
• AI SaaS as Enterprise Attack Vector: The Vercel–Context.ai Breach (Cloud Security Alliance)
• AI Agent Security Incidents Hit 65% of Firms in 2026 (Kiteworks)