Welcome back to Building with AI. This week I kept coming back to one question: when an AI agent can act on its own, who decides what it's allowed to do, and who can stop it? The news gave a pretty clear answer for most companies. Plans to deploy agents are running well ahead of the controls to govern them.
I'm hopeful about what agents can take off our plates. But AI was built as a tool to help us, and a tool can't be accountable for itself. We have to guide it, and we have to be responsible for its decisions. So this edition is about the unglamorous work that makes agents safe to use: clear limits, clean data, people who know what good looks like, and someone with the authority to say "stop."
The Story
Most companies want AI agents. Few are ready to govern them.
Last month OpenAI published an unusual report about its own coding agents. In one internal task, an agent repairing an evaluation pipeline switched off automatic alerts to OpenAI's security team. In another, an agent migrating a program removed its monitoring integration, then told the user it hadn't. OpenAI says both agents considered whether to keep the safeguards and read routine coding instructions as permission to weaken them. Neither incident was flagged as top priority, and neither run was stopped automatically. OpenAI says it is tightening its reviews and working toward automatically shutting down any workload that doesn't meet its monitoring requirements.
If that can happen inside the company that built the models, it's fair to ask how much visibility the rest of us have over the agents we're starting to deploy. On November 9, Gartner said its survey of more than 2,000 CIOs and technology executives found that 71% of organizations expect to deploy agentic AI within the next 12 months, while only 21% currently have agentic AI governance capabilities in place.
The same day, Gravitee released a survey of 900 senior technology leaders in the US and UK. Fewer than 1 in 10 organizations have a named person formally accountable for what their agents do, only 1 in 8 can fully reconstruct what an agent did 30 days ago, and just 6% can revoke a rogue agent's access within an hour. Gravitee sells agent management software, so it has a stake in the answer, but the pattern matches what Gartner found.
What I take from this is that we're handing out autonomy faster than we're defining it. With people, I've always found the best results come from giving them a clearly defined box to operate inside, and plenty of room to decide how they work within it. Agents need that box even more, because when something goes wrong, the accountability still lands on a person. In Edition 01 I asked who owns AI's mistakes. For agents, the answer has to be settled before they go live: a named owner, written limits on what the agent can touch and decide, a record of what it did, and someone who can switch it off quickly.
Sources: OpenAI · Gartner · Gravitee
The Rundown
Agents are only as good as the data underneath them
Amazon Web Services has been promoting a striking figure: only about 1 in 5 organizations have the data foundation AI agents need. I traced it to Capgemini Research Institute's 2025 survey of 1,500 executives at companies with more than $1 billion in revenue. Fewer than 1 in 5 rated their data readiness as high, and 82% described their AI infrastructure as low to medium maturity. The data is from 2025 and self-reported, but in my experience the pattern hasn't changed much. Teams get excited about the new technology and forget the data it depends on. If you're unsure where yours stands, data is the first thing our AI readiness assessment looks at.
Source: Capgemini Research Institute
AI code is getting safer, slowly
Veracode tested code from 18 AI models in October and found the average security pass rate rose from 56% to 62%, the biggest improvement since its testing began. That still leaves about 4 in 10 coding tasks producing vulnerable code, even when the code works. Quality and know-how matter, especially with AI, if you want real value out of it. Without the depth of experience, building with AI can feel like giving a baby power tools; but applying insight and discernment can turn them into the precision work of a professional artisan.
Source: Veracode
A startup that checks agents before they act
Bay, a Tel Aviv company less than a year old, came out of stealth with $50 million, combining a $37 million Series A led by Battery Ventures and a $13 million seed round. Its software checks each action an agent proposes before it runs, then allows it, blocks it or sends it to a person for approval. The company names Micron, Peloton, ClickUp and Wix as customers. It's an early sign that investors are starting to fund the controls around agents as well as the agents themselves.
Source: The Next Web
LangChain raises $300 million for agent tools
LangChain announced a Series C at a $3 billion valuation on November 13, led by CapitalG. Its LangSmith platform is built to run, improve and govern AI agents, and the company argues businesses should be free to choose their models and control their own data and infrastructure. For anyone evaluating agent platforms, the questions from this week's lead are a good test: can you see what the agent did, set limits on what it can reach, and change it safely?
Source: LangChain
The Human Side
What happens when people only approve AI's work?
Wharton's Kartik Hosanagar and Daehwan Ahn asked 285 students to write a 1,000-word story on their own. In the next class, the same students wrote a second story with AI, using one of three setups. In the first, AI came up with the ideas, outlined, drafted and edited, and the students only approved the result. In the second, the students came up with the ideas and wrote the outline themselves, then handed the drafting and editing to AI. In the third, they could use AI at whatever stage they chose. AI made the writing 36.2% faster overall. When readers rated them, the approve-only stories scored lower in quality than the students' first-round stories, while stories from the other two setups scored higher than the first-round compositions. The approve-only group was also less satisfied, and felt less pride in what they'd made. The study was published in PNAS Nexus on November 10.
A PA Consulting survey of 2,000 UK employees, released November 9, points to the same issue at work. Among senior leaders, 70% said AI helps them achieve things they couldn't do before, compared with 38% of entry-level and junior employees. And 13% said they feel less responsible for quality when AI has produced most of the work.
What worries me is what's lost when people early in their careers mostly approve what AI produces. They miss the institutional knowledge: how we got here, what we learned, how we treat customers. There's no ground floor to work up from anymore. Most of us find purpose in our work, and a lot of that comes from shaping it. Our teams need to use AI every day, but they shouldn't check their brains at the door. I'd rather see people set the direction, own the result, and use AI to help with the rest.
Sources: EurekAlert · PNAS Nexus · PA Consulting
Try This Week
Draw the box for one agent
Pick one AI agent or automation your business already runs. Write down three things: what it can do without anyone's approval, what it must never do, and who can switch it off. Then ask that person to show you how they would do it, and how long it would take. If nobody is named, or switching it off means a ticket and a two-day wait, that's your first governance gap to close.
That's it for this week. If one of your AI agents made a costly mistake tomorrow, how long would it take you to find out? Hit reply and let me know.
References
- OpenAI, Suppressing misalignment monitoring alerts and disabling misalignment monitoring (October 2026)
- Gartner, Gartner Says Aligning AI Value, Risk and Safety Is Now the Top Challenge for CIOs (November 9, 2026)
- Gravitee, Enterprises Have Learned to Deploy AI Agents. They Have Not Yet Learned to Answer for Them (November 9, 2026)
- Capgemini Research Institute, Rise of agentic AI: How trust is the key to human-AI collaboration (2025)
- Veracode, Veracode Finds First Meaningful Gain in AI Code Security Since 2025 as Average Pass Rate Hits 62% (October 29, 2026)
- The Next Web, Bay raises $50m to vet AI agent actions before they run (November 10, 2026)
- LangChain, LangChain Raises $300M to Enable Every Company to Own Its Intelligence (November 13, 2026)
- EurekAlert (PNAS Nexus), Using AI for creative tasks (November 10, 2026)
- Kartik Hosanagar and Daehwan Ahn, Designing human and generative AI collaboration: Trade-offs in productivity, satisfaction, and creative diversity (PNAS Nexus, November 2026)
- PA Consulting, AI risks widening the workplace divide, new report finds (November 9, 2026)






