Who Watches the Agents?


The agents are in production. The guardrails are not.

That is the gap this week, and the seed money went straight into it.

Start with the numbers. Zapier surveyed 513 US enterprise leaders in September and found that 88% of enterprises are now using or testing AI agents, up from 72% a year ago. Production use jumped from 40% to 71% in one year, and the share still piloting or testing fell to 17%. Nine out of ten leaders say their organizations are certain or likely to increase investment over the next year, up from 84% in 2025.

Then look at what they actually trust the agents to do. Only 30% of leaders say their agents complete more than 60% of assigned tasks without human intervention. When they are asked what holds them back, the answers are data privacy, security, and regulatory compliance risks (27%), then accuracy, reliability, and errors (24%).

Genpact surveyed more than 2,000 executives across 17 industries and found the same split from the other side. 79% of enterprises have AI agents in production, but just 20% use them broadly and consistently across functions. Only 15% have fully tied AI investment to P&L outcomes. On average, just 49% of AI initiatives funded in the past 24 months delivered the value expected. Only 9% have completed significant operating model transformation.

And Gartner has put a date on the bill. The firm predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because governance gaps are found only after production incidents occur. That is a prediction, not a fact — but it tells you where the smart money is looking.

The seed rounds this week are the answer to that prediction.

Metano raised $12 million on Nov 19, led by Sierra Ventures, to watch what agents do once they are loose in the enterprise. The platform discovers approved and unapproved agents, MCP servers, skills, and software packages; follows agent behavior in context through Hive, a real-time context graph connecting agent tasks, identities, tools, data, and actions; and enforces guardrails where agents operate, including activity that may bypass traditional network gateways. It is generally available now. “Traditional security tools were not designed for AI agents that can make decisions and change their behavior as they operate,” said Neelima Rustagi, co-founder and CEO. “We built Metano to understand agent behavior in real time, connecting intent to action, and stopping threats before autonomy turns into risk.”

Unsiloed AI raised $7.5 million on Nov 18, led by Felicis, to make sure agents read the right numbers in the first place. The software breaks down complex business documents — tables, charts, complicated layouts in PDFs, images, and spreadsheets — extracts the information into structured data that AI agents can use, and shows where each extracted item came from in the original document so businesses can check results rather than take the AI’s output at face value. “The problem with LLMs is they’re inherently probabilistic,” said Aman Mishra, co-founder and CEO. “There is going to be some deviation induced. There is going to be some hallucination involved, and that has a massive negative impact downstream.” The company says its tests reached 99.78% accuracy across roughly 84,000 extracted data points from 542 financial documents — its own tests, not independently verified, so treat the number as the company’s claim.

Two rounds, two directions. Metano guards what agents do. Unsiloed guards what agents read. Both sell the same thing in different forms: proof you can check.

I have seen this before, in a different market. In The Impact Factor, Chapter 16, “Systems,” I wrote about affiliates who had never used the product they were selling: “How can these people possibly believe in what they are selling? How can they understand the product? How can they identify with it? How can they in turn instill trust in their target audience? They cannot. It is as simple as that.”

An agent that has never really read the document cannot be trusted with the answer. An agent nobody watches cannot be trusted with the action. And the founders who win the next round are the ones who build the checking in — not the ones who promise the model will behave.

Here is what that means if you are raising in the next 90 days.

First, show the guardrail, not just the agent. Your investor has read the same surveys. A demo of what your agent can do is table stakes now. A demo of what stops it when it is wrong — that is the part they remember.

Second, name the human in the loop. Who reviews what, and when? If you cannot answer that in one sentence, you are not ready for an enterprise customer’s security review, and your investor knows it.

Third, bring one customer who can say the agent’s work held up under their own checks. Not a logo. A person who looked at the output, verified it, and kept using it. That is the whole story, told by someone who is not you.

The agents are in. The question is who watches them.

Read the full issue in Raise Smarter, and subscribe here: https://www.linkedin.com/newsletters/raise-smarter-7514657628552183808/

For founders: https://connectlab.live/founders — 21-day free trial of Founder Pro. For funds: https://connectlab.live/join/funds

Connection is the currency of the 21st Century.


About Ken McArthur

Ken McArthur is the founder of ConnectLab.live, an AI relationship-intelligence and warm-intro platform that helps founders who are raising find the funds that fit and the people who can introduce them, and helps fund managers connect with the founders they back. He was recently accepted into Harvard Business School Foundry. He also writes Raise Smarter, a daily newsletter on what's really moving in early-stage fundraising. Ken challenges us to realize we ALL have an impact, whether we want to or not, on thousands of people we touch in our day-to-day lives, and he shows that simple things make a HUGE difference. ConnectLab.live grew out of that belief. Connection is the currency of the 21st Century, and the right introduction can change a company's future. Long before ConnectLab, Ken was the popular host of live events that brought together top-level marketers, entrepreneurs, business owners, corporations and non-profit organizations to build multi-million dollar joint venture relationships. He has managed product launches ranked in the top 400 sites on the Internet and is regularly asked to speak at leading marketing events. He created AffiliateShowcase.com, a pioneering affiliate program search engine and directory. He also founded the MBS Internet Research Center, which conducted the world's largest survey ever attempted on creating and launching successful information products. Ken was the official mentor for Sterling Valentine as he took his launch from zero to over $100,000 in less than 8 days, a proof of concept documented in Info Product Blueprint. Today he puts that same skill for connecting people and launching ideas to work for founders and the investors who back them.

Leave a comment