Cisco gave 90,000 employees access to a platform for building their own AI agents. Half of them were using it daily within two weeks. Around 700 employee-built agents are now running inside the company, each one making decisions, calling tools, and touching systems that nobody centrally signed off on.
That's not a thought experiment. That's this month, at one company.
The Governance Gap Nobody Budgeted For
A recent Collibra and Harris Poll survey put numbers on what most engineering teams already suspect: 70% of decision-makers say unaligned data foundations are the reason their AI pilots fail. At large organizations, the ones pulling in $100 million or more in revenue, 64% say a human still has to manually review what the agent did before anyone trusts it, and 96% point to bad data as the root cause when something goes wrong. Across the board, 87% of leaders say they have to re-verify an agent's context before acting on its output.
Read that again. The agents are shipped. The review process to actually trust them is not.
PwC's chief AI officer put the legal reality in one line: "the agent made me do it" is not a valid defense, morally or legally. Somebody in the company is accountable for every action an agent takes, whether or not anyone built the paper trail to prove it.
What the New Platforms Are Actually Admitting
Look at what's shipping right now and the gap becomes obvious. Intuit built GenSRF, a security, risk, and fraud layer that tracks every request and response an agent makes. Workday launched what it calls an "agent system of record," treating AI agents as a digital workforce with identities to manage, not just tools to license. ServiceNow says its AI Control Tower is one of the fastest-growing products it has ever shipped.
These aren't features. They're an admission that identity, scope, and audit trail for AI agents is now its own category of enterprise software, because almost nobody building agents six months ago planned for it.
Build This Before You Need It, Not After
If you're a founder or engineering lead shipping agents right now, the lesson isn't "wait for enterprise tooling to mature." It's that you can build the basics yourself, cheaply, before you have 700 agents to retrofit:
Give every agent its own identity and scoped credentials, never a shared service account inherited from a human user. Log every tool call with the reasoning behind it, not just the outcome, so a review six weeks later actually means something. Put a kill switch in front of anything that writes to production or touches customer data, and make it something a human can hit in seconds, not a ticket that takes a day to route. Review the action logs weekly while the agent count is small. Waiting until quarterly review is how you end up needing an "agent system of record" instead of a spreadsheet.
Retrofitting an audit trail onto agents that are already touching production is a much bigger job than building it in from the start. The companies buying GenSRF and Control Tower right now are the ones that skipped this step.
We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!