If you cannot say when you would stop:
eleven AI leaders, one record.

On October 15, 2026, eleven speakers took the virtual stage at the AI Innovation Summit, hosted each month by Gerald “Solutionman” Haman and the AI Innovators Network. They came from predictive analytics, enterprise strategy, workforce development, cloud partnerships, manufacturing and public policy. Each had fifteen minutes.

Heard one at a time, they covered very different ground. Heard together, many of them described the same thing from different sides: what an organization should be able to show, in writing, about each AI system it runs.

Helen Teplitskaia, founder of the Global Alliance on Sustainability and AI, put it in one sentence in her slides:

“If you cannot say when you would stop, you are not in control of the system.”

That sentence is a practical test for any AI deployment. Below is what each speaker added to it, arranged as the questions a well governed AI system should be able to answer.

Eleven speakers, one record

Read across, each row is one speaker’s idea and the line it adds to an organization’s record of its AI systems. Read down, the last column is that record.

One thread runs through every row. None of these answers can be assumed. Each has to be written down, owned by a named person, and dated, or nobody can check it later.

Five questions every AI system should answer in writing

1. Where does it come from?

Anita Luthra asked the questions most organizations skip at the start: who built the model, what it was trained on, where it runs, and whose laws govern it. Her practical advice was to keep an exit, so a change outside your control does not stop your business. The answers belong in the record before the system goes live, along with the name of the person who checked them.

2. When may it act on its own?

Eric Siegel showed how a predictive reliability model scores each output an agent produces. The riskiest ones go to a person before they reach a customer. In his words, “we need to predictively and judiciously escalate to that more expensive human in the loop.” At Twilio, he said, certain customer interactions are “quarantined and held until a human can audit and potentially intervene.” The agent gives up a little autonomy and gains a great deal of reliability.

Someone chooses the score where that line sits. The confidence score where an agent is allowed to act on its own is not a tuning setting. It is a permission. Like any other permission it needs to be written down somewhere an auditor can find it, with the name of the person who set it and the date it last changed.

3. Who can stop it, and how fast?

Helen Teplitskaia’s rule was simple: no human stop, no launch. Every high-impact system needs a named owner and a named person with real authority to pause it. Her four steps make that concrete:

1. Put a human stop on the system before it ships.
2. Scale in stages, and write the stop condition before each stage, with the measured time to reach a safe fallback.
3. Give each agent only the access its task needs.
4. Let the people who see a failure first call a stop without a meeting.

Her slide on accountability said it plainly: someone owns the decision to deploy and the decision to stop. In her talk she added that each agent should be traceable, with its isolation designed, monitored and revocable.

Teplitskaia also discussed the OpenAI and Hugging Face agent incident, which SanctumShield has written about in AI Did Not “Colonize” Us. But Humans Did Lose Containment.

4. Why was this decided, and what earns more autonomy?

Langdon Morris argued that as intelligence becomes abundant, competitive advantage comes from good judgment, and that AI should be used as a tool with exceptional rigor. His company Crucible pairs a protagonist AI with an antagonist AI to test each line of reasoning, and keeps persistent reasoning and institutional memory: a durable record of why each decision was made. Rigor matters most, he said, “when we want to create a record of the decisions that we’ve made and why we’ve made them.” Advantage, he said, comes from “allocating authority wisely, learning faster, and then remembering why you made the decisions you made.”

Eric Pilon-Bignell’s advice for leading in an agentic world was to release agency within guardrails, starting small. “Not every agent should have unrestricted access,” he said. “We need appropriate permissions, oversight, clear guardrails, and boundaries.” “You don’t start in a 20-foot wave,” he said, and autonomy “should expand as reliability, capability, and trust are demonstrated.” Together they describe the record that makes growth safe. Each expansion of autonomy is a decision, and each decision deserves its reasoning, its owner and its date.

Rushikesh Kulkarni brought evidence to the same question. He put one escalating ethical dilemma to several AI systems and found that their final answers were often close while their reasoning differed widely. His advice to anyone deploying AI was direct: “I would assess AI solutions more deliberately,” and test their values “before you actually implement these solutions.” The results of those tests belong in the record as well.

5. Are the people ready?

Three speakers turned to the people around the system. Richard Pozniak urged organizations to judge readiness by the floor of the workforce rather than its five best users, and to measure each person against what the role requires. Carol S. Scott described AI fluency as knowing what matters rather than everything: what an agent is delegated, what it costs, and where human judgment adds value. Jerry Abrams reminded the audience that human difference matters most when a situation is new, which is exactly where an agent’s boundary ends and a person’s judgment begins.

Nearly every idea on the program depends on a person in the loop: the reviewer who receives a doubtful output, the person with authority to stop, the judge of a case nobody designed for. That person can only do the job well if they understand what they are looking at. That is why readiness belongs in the record too.

Article 4 requires measures that support AI literacy, suited to each person’s role and context, and has applied since February 2, 2025. AI literacy (EU AI Act Article 4) in the glossary carries the amended text and its sources.

What it looks like on one page

Here is an illustrative entry for a single agent that reviews insurance claims. It is an example, not a real customer, and every value is invented to show the shape of the record.

None of this requires new technology. It requires someone to decide each line, write it down, and keep it current.

Where SanctumShield and the Academy fit

The speakers spoke for themselves and their own work, and the arrangement here is mine. Still, it was striking how closely their principles match the ones SanctumShield and the SanctumShield AI Literacy Academy were built on. Each of the five questions has a practical place to start.

Organizations that are still deciding where to begin do not have to take one person’s word for any of this. Eleven respected leaders, from eleven different directions, described the same responsibilities. The record is how an organization meets them.

Walk the walk, not just the talk. No consulting engagement, network appliance or guardrail product can fill in this page for you. Those all have their place, and they work best when the page already exists. Only the people who own the system can write it.

Closing

Eric Siegel closed his talk with a reminder that applies to every speaker on the program: excitement about the technology is only the first wow. “That’s just the theatrics,” he said. “It doesn’t count until you deploy it.” Deployment is where the record matters, because a deployed system makes real decisions for real people.

Several speakers arrived at the same point from different directions. Siegel said that “the concept of agency is sacred and belongs to people.” Morris said that on strategy and innovation, “people need to be making the decisions.” Brian Fried reminded inventors that under US patent law “the human is the inventor.” Teplitskaia asked everyone to “maintain the agency in our hands.” AI can reason, draft and act, and responsibility stays with named people. A written record is how an organization keeps it there.

The eleven speakers came at AI from very different directions. What they share is a practical confidence that AI can be deployed well, by organizations that can say where each system came from, when it acts on its own, who can stop it, why each decision was made, and whether the people around it are ready. That is what responsible AI governance looks like in practice, and it is within reach of any organization willing to write it down.

Thank you to Gerald Haman and the AI Innovators Network for a thoughtful program, and to each speaker for sharing their work. The full recording, just under four hours, is available on the AI Innovators Network YouTube channel at https://www.youtube.com/watch?v=TDcZoforVYU.

Start with your own record. SanctumShield is AI governance software that helps any organization quickly and effectively identify and assess the governance needs of its own AI systems. It is free to access at https://sanctumshield.com. You can also run the free Shadow AI Risk Calculator at https://sanctumshield.com/calculator and take the free AI governance literacy quiz at https://sanctumshield.com/quiz/ai-governance-literacy. For role-based training, the SanctumShield Academy at https://academy.sanctumshield.com provides AI governance and AI literacy training with verification for board members, CEOs, CISOs, legal counsel and employees.