How to control agents on regulated systems
Give regulated agents a control plane before they touch production data. Keep an inventory record for each agent, test it in a simulated environment so it does not perform unwanted actions, set and monitor its policies,
Give regulated agents a control plane before they touch production data. Keep an inventory record for each agent, test it in a simulated environment so it does not perform unwanted actions, set and monitor its policies, write an audit record of what it did, and keep a human in a supervisory role. The policy check has to inspect combinations: two or three pieces that do not meet a privacy test alone can become PII when they come together, and a compliance pass at launch will miss that later join.
Kris Lovejoy, Global Head of Strategy at Kyndryl, made this case on Chain of Thought in June 2026.
Why are autonomous agents still uncommon in banking and healthcare?
Pilots run ahead of the estates they have to operate on. Lovejoy splits AI so it is not treated as one thing. Individual productivity comes first. People already use generative and agentic tools for ordinary tasks, such as booking a hotel, and those agents inherit the user's authorizations. Enterprise productivity tools are in use at scale.
Mission-critical transformation is the second category. Customers that run banks and healthcare systems are not at full production scale with those programs.
What stops them depends on the situation. Her short answer was all of the above. Critical infrastructure is mostly not cloud first. The estate is hybrid: mainframe and mid-range, machines she called X-series, more than one cloud, vintage systems beside new ones, and data spread across them. Getting to the data, normalizing it, and keeping its integrity so a model can use it is difficult. A pilot can succeed and still leave out security, reliability, sovereignty, economics, and scalability. Guardrails and resilience capabilities production needs often do not exist yet, so teams invent them while they deploy. Her range for agentic AI at scale in most organizations was a few years, maybe up to a decade.
Within an enterprise, productivity tools are being used and they are being used at scale.
Kris Lovejoy, Global Head of Strategy at Kyndryl, on Chain of Thought ep 62
The First Fully Autonomous AI β¦ - Chain of Thought | AI Agents, Infrastructure & Engineering - Apple Podcasts
Podcast Episode Β· Chain of Thought | AI Agents, Infrastructure & Engineering Β· June 11 Β· 46m
What belongs in a control plane for a privileged agent?
If you can't see it, you can't protect it. Lovejoy applied that rule to agentic AI acting as a privileged user.
She listed the operating steps in this order:
- Inventory the agent.
- Evaluate it in simulated environments so it does not perform actions you do not want.
- Generate the policies, establish them, and monitor them.
- Create audit reports and compliance reports that show what happened.
- Keep humans in supervisory roles.
Policy as code sits inside that list, and her relationship with it is fraught. She remembers the early cloud pitch: implement policy as code and security problems go away. That narrative failed, and teams added compliance and secondary controls. Applying deterministic rules to autonomous systems leads her to ask whether the result is RPA or agentic AI. Policy as code remains one element of the reference architecture.
She pushed back on human in the loop. A linear, human-shaped checkpoint gets forced onto work that runs as parallel tasks. Teams then struggle to write constraints where the person is not a roadblock to finishing the process. The model she argued for is human on top: supervisory roles, with secondary evaluations and approvals before the agent can finish the action.
if you can't see it, you can't protect it. And so when you're talking about mission-critical systems that are going to be using agentic AI, and when I think about the role of agentic AI as a privileged user interacting with those systems, then we have to apply certain rules to how we manage that.
Kris Lovejoy, Global Head of Strategy at Kyndryl, on Chain of Thought ep 62
How do you flag a join of data that each field would allow?
Check the set of categories in the query, not each category on its own. Lovejoy's privacy example is two or three pieces of data that, alone, have not met a jurisdictional privacy test and, together, count as personal data.
Privacy, as she defined it, is an individual right. When someone gives data to a third party, that party inherits obligations of access, integrity, and not passing the data to others. Security is one mechanism for meeting those obligations. The rules are a patchwork, so an obligation we have not seen can attach to one element or to a combination.
Compliance, in her account, still works as a point-in-time habit: test the code, mark it compliant, and stop. That habit does not hold once data is joined, connected, provided onward, or processed in a new place. Continuous compliance replaces it. Ongoing observation is what notices when a privacy constraint has appeared.
The category names below are stand-ins for that check.
A minimal sketch to adapt, not a drop-in library.
inventory = {
"agent_id": "claims-review-07",
"owner": "claims-ops",
"nonhuman_identity": "nhi-claims-review-07",
"simulated_ok": True,
"allowed_alone": ["member_id", "claim_code", "device_locale", "visit_date"],
}
forbidden_combinations = [
{"member_id", "device_locale"},
{"claim_code", "device_locale", "visit_date"},
]
def query_is_blocked(categories, inventory, forbidden_combinations):
present = set(categories)
if not present <= set(inventory["allowed_alone"]):
return True # a category this agent was never cleared for
return any(combo <= present for combo in forbidden_combinations)
Run query_is_blocked on the categories a query will touch before the agent calls a tool. Repeat the check when the query, the recipient, or the processing location changes. A pass at registration is the point-in-time review she described as something that does not work anymore.
When two or three pieces of data come together, all of a sudden it's PII.
Kris Lovejoy, Global Head of Strategy at Kyndryl, on Chain of Thought ep 62
FAQ
Is policy as code enough to govern an agent?
Lovejoy called policy as code one element of the reference architecture. Inventory, simulated evaluation, monitoring, audit reports, and human supervision sit with it. She asked whether that result is RPA or agentic AI.
What does human on top require in practice?
A supervisor who can evaluate and approve before the agent can complete the action. Lovejoy set that against a person placed in the middle of the workflow. AI does not think in a linear way, and a midpoint human check becomes a roadblock to finishing the process.
Why is a launch-time privacy review not enough?
Joining data can change the obligation. Pieces of data that are not covered on their own can become personal data in a set of two or three. A test-and-done compliance pass does not track later joins, connections, disclosures, or processing locations. Monitor those changes and block or escalate when a forbidden combination appears.
Takeaway
- Record every agent: owner, non-human identity, allowed categories, and whether simulation passed.
- Block a query when its categories include a forbidden combination of two or three, even if each category is allowed alone.
- Use policy as code as one control, beside inventory, simulation, monitoring, and audit.
- Require supervisory approval before the agent completes a consequential action.
- Re-check combinations when the query, the recipient, or the processing location changes.
The full conversation, with the transcript, is on Chain of Thought.
Subscribe to the Chain of Thought newsletter for new episodes and write-ups like this one.
Drafted with AI assistance from the episode transcripts.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.