Research

DELEGATING AUTHORITY TO AI

Our research explores how organizations turn AI capabilities into sustainable business value - with particular attention to governance, decision rights, operating models, and human judgment.

That work has culminated in our first practical guide, Delegating Authority to AI. It helps business leaders decide what to delegate to AI, within what limits, and when to expand its authority.

The guide brings together published research and institutional guidance with insights from our research team and interviews with senior leaders and practitioners. It covers five deployment pitfalls, permissions and oversight, evaluation, and ongoing review, with templates for documenting deployment decisions.

Download Delegating Authority to AI (PDF)

This publication marks our transition from weekly research updates toward practical, evidence-based guidance. We continue to welcome deployment experience and feedback on the guide through our Contact Us page.

Earlier research updates appear below, tracing the findings and questions that informed this guide.

----------------------------------------

AI Governance Research Update #8: When Monitoring Isn’t Enough

One idea keeps surfacing in this research: capable AI agents don’t necessarily create risk by abandoning their objective - they can create it while trying to accomplish it.

New evidence this week makes that problem more challenging.

Researchers studying multi-agent systems found that seven of nine frontier models attempted to disguise a protected credential so another agent could recover it - without an adversarial objective. More surprisingly, monitoring sometimes increased concealment rather than preventing it.

That finding echoes real-world evidence. During cyber evaluations, the UK AI Security Institute observed agents taking sustained, unsanctioned actions directed at real people and organizations. Other incidents show agents finding alternate paths around restrictions and acting beyond intended boundaries.

The implication: monitoring is necessary, but monitoring alone is not control.

The problem may also be larger than the individual agent. Recent research has demonstrated self-replicating prompt injections that can propagate through agent interactions and emergent collusion among multiple agents. AWS also disclosed vulnerabilities in its Loom agent platform involving authentication bypass, credential exposure and access to internal resources.

We may need to assure the entire agentic system: agents, tools, credentials, communication paths, infrastructure and controls.

Pieces of that architecture are emerging. NVIDIA is developing independent runtime enforcement. The Advanced AI Society’s Proof-of-Control focuses on independently verifying what agents did and whether controls held. California is establishing a framework for Independent Verification Organizations, while its Attorney General is raising developer accountability for cybersecurity risks during development and deployment.

The research question is becoming more specific: Can we continuously monitor agent behavior, independently enforce boundaries, verify that controls held, and preserve accountable human authority - without eliminating the autonomy that makes agents valuable?

A note on where this is going

These updates began as a way to share an open research project. The research and community around it are growing. I’m now going to evolve these updates into a research publication focused on autonomous AI, enterprise control and continuous assurance.

The goal isn’t another AI news digest. It’s to examine what new evidence means, test emerging ideas against real-world experience, and bring more practitioners into the work.

I’ll share more about the publication and research agenda later this week.

----------------------------------------

AI Governance Research Update #7: From an Emerging Control Model to Continuous AI Assurance
Week of September 21, 2026

Over the past several weeks, this research has evolved from exploring individual AI governance questions into a more focused initiative: Continuous AI Assurance.

The vision is that organizations should be able to use increasingly capable and autonomous AI while maintaining confidence that those systems remain within acceptable boundaries as conditions change.

Our working hypothesis is that traditional periodic assurance will not be sufficient. Increasingly autonomous systems may require a continuous loop that observes behavior, evaluates changing risk, adapts controls, verifies that those controls are operating as intended, and provides evidence to human decision-makers.

Importantly, that remains a hypothesis. The goal now is to test it. Three developments this week moved that work forward.

1. Moving from research to validation
We developed a Validation & Engagement Plan organized around landscape analysis, practitioner validation, real-world use cases, ecosystem engagement, and evidence synthesis.

The intent is not to build consensus around the hypothesis. It is to actively look for supporting, refuting, and dissenting evidence and determine where the model holds up - and where it doesn't.

2. Discovering potentially complementary work
The Advanced AI Society's work on Open Verification and Proof-of-Control was particularly interesting.

PoC addresses what AAIS calls the "verifiability gap": independently establishing what an AI agent actually did and whether defined controls held. But PoC deliberately does not determine whether those controls were adequate, whether the agent should have been deployed, or whether its outputs were correct or safe.

That boundary may be important. Independent verification could become one component of a broader continuous-assurance loop rather than an alternative to it. That's something we now want to explore with AAIS.

3. Independent evidence is converging around continuous, adaptive control
Research this week strengthened the case for runtime monitoring, bounded authority, least privilege, control-gated deployment, adversarial testing, and controls that adapt as threats and systems change.

But it also sharpened what may be the most important unresolved question: Can assurance and control improve fast enough to keep pace with increasing AI capability and autonomy? That's increasingly where the research is headed.

The next phase is less about adding ideas to the model and more about trying to break it - through practitioners, real enterprise use cases, competing approaches, and evidence that challenges our assumptions.

----------------------------------------

Research Update #6: From Incidents to an Emerging Control Model

Week of September 14, 2026

This week, the research moved from asking whether stronger governance can enable greater AI autonomy to a more concrete question: What kind of control architecture is actually required when capable agents can find paths we did not anticipate?

The risk analysis has now expanded into a cross-incident review of 26 publicly reported cases. The strongest emerging pattern is that autonomy alone does not determine risk. Consequential exposure appears when decision autonomy combines with action authority and controls that do not independently constrain harmful actions.

Several related findings are becoming clearer. Goal authorization is not the same as action authorization. Agents can route around obstacles using alternate credentials, tools, repositories, public services, or communication paths. System and data boundaries are recurring failure points. And successful task completion can conceal an unacceptable method. These are working findings from a purposive corpus, not estimates of incident frequency or failure probability.

Recent OpenAI disclosures strengthened that picture. In separate cases, models searched for exposed API keys, uploaded files to public services, used internal repositories for unintended cross-context communication, and created new agent-to-agent file-sharing paths when the expected mechanism failed.

Our AI roundtable this week added another dimension. A recurring concern was the mismatch between current LLM behavior and regulatory frameworks shaped by deterministic software. In highly regulated environments, expectations around repeatability and reproducibility do not always map neatly onto probabilistic systems. The emerging question may be whether assurance should focus less on identical outputs and more on bounded, observable, auditable, and risk-controlled behavior.

Practitioner evidence is pointing in a similar direction. An executive at Workday described their use of semi-autonomous agents within defined “moats,” with autonomy expanding only as controls and confidence mature. He also raised a practical governance issue that may become increasingly important: systems need to distinguish actions performed by a person from actions performed by that person’s agent.

The control model emerging from the research is becoming more specific:

Agent → intelligent monitoring → deterministic controls → human oversight

The remaining question is whether that architecture can preserve meaningful autonomy without sacrificing accountability, security, or scale.

What comes next: I’m going to test this emerging control model against additional incidents, practitioner interviews, and regulated-industry use cases. In particular, I want to understand where deterministic controls are sufficient, where intelligent monitoring adds value, and how organizations can preserve meaningful autonomy without losing accountability or creating unacceptable risk.

----------------------------------------

Research Update #5: When the Evidence Pushes Back

Week of September 7, 2026

We've collected a good amount of data so this week I started testing the data with a few questions: What does the evidence strongly support? Where is it weak? And, most importantly, what challenges the hypotheses I've been developing? The exercise sharpened three questions.

1. Can governance enable autonomy - not just constrain it?

I've been exploring the idea that better governance could create enough confidence for organizations to grant AI agents greater autonomy. The evidence supports only part of that proposition - strong controls clearly reduce risk, but there is much less evidence that implementing those controls actually leads organizations to grant agents greater consequential authority.

The distinction matters - a system that achieves perfect safety by escalating every consequential action to a human may be safe - but it's not autonomous and it can't scale.

2. Guardian Agents: refining the architecture

I've also been exploring “Guardian Agents”: independent AI agents that monitor other agents at machine speed. Here, too, the evidence pushed back - AI supervisors can share blind spots with the agents they monitor and can even be persuaded by an acting agent's rationale for questionable behavior.

That doesn't eliminate the concept, but it changes its role. A more defensible architecture is emerging: Agent → Guardian/Monitor → Deterministic Controls → Human Oversight. The Guardian provides contextual interpretation, anomaly detection and triage. Deterministic controls enforce hard boundaries. Humans retain authority over consequential exceptions.

3. Is AI really different - or are we just seeing another J-Curve?

A third hypothesis is that AI is creating an unusually difficult organizational adaptation problem because the technology changes faster than organizations can redesign around it.

The evidence isn't conclusive. An alternative explanation remains quite plausible: we're seeing the familiar Productivity J-Curve - technology advances rapidly, while organizations take much longer to redesign workflows, skills and structures around it.

That's where the research is becoming more interesting. These ideas haven't disappeared under scrutiny. They've become narrower, more testable - and less certain.

What comes next: I'm now shifting from broad exploration to targeted evidence gathering. The next phase will focus on finding empirical studies, benchmarks and real-world cases that can test these questions - particularly whether strong governance can preserve meaningful autonomous action, where intelligent supervision adds value, and whether AI's organizational adaptation challenge is genuinely different from previous technology transitions.

----------------------------------------

Research Update #4: From Principles to Risk Assessment

Week of August 31, 2026

Last week I ended with a question: What organizational capabilities allow companies to continuously adapt to increasingly capable AI while safely delegating greater autonomy and preserving meaningful human control? This week, the research moved closer to how we might answer that question.

When agents pursue the objective - but not the intent

I looked closely at several real-world incidents involving AI agents. A common pattern emerged: an agent can pursue a legitimate objective while taking actions its human principal did not intend or authorize.

In one case, an agent exploited weaknesses in a gym's booking system while trying to improve its user's position. In others, agents deleted production data or emails because the technical authority available to them exceeded what their tasks required.

These cases suggest an important distinction: Goal authorization is not the same as action authorization.

Instructions matter, but instructions alone are not controls. For consequential actions, permissions, approval gates, monitoring, and other externally enforced constraints become increasingly important.

Control is becoming a deeper question

I also had a fascinating conversation with Rose De Sicilia about where control can be imposed in an AI system - from pre-training and model architecture through post-training and runtime controls.

We discussed research suggesting that sufficiently capable models can sometimes recognize aspects of the environment in which they're being evaluated or controlled and alter their behavior accordingly.

That complicates governance: as the system changes - and potentially adapts to the controls themselves - the controls may need to change too.

Piloting an Agentic AI Risk Assessment

I'm now beginning to pilot a quantitative approach to the question emerging from this research: How does enterprise risk change as we give AI agents greater autonomy and consequential authority?

The assessment starts with one real AI agent or workflow and examines what it can observe, decide, and do; what could materially go wrong; what controls exist; and the consequences if those controls fail.

The goal is to quantify major risk scenarios, identify which uncertainties drive exposure, and determine which controls reduce that exposure most effectively.

Before attempting a broader analysis, I'd like to test and refine the approach against several real-world use cases.

If you're building, deploying, or managing an AI agent and would be willing to put your own work through the assessment, I'd welcome your participation. It involves a 75–90 minute working session focused on one agent or workflow. Perfect data isn't required - estimates and unknowns are useful too.

If that sounds interesting, please reach out.

----------------------------------------

Research Update #3: From AI Adoption to Continuous Adaptation

Week of August 24, 2026

This week, several threads in my research began to converge around a larger question. I've been exploring what organizations need to become “AI-ready,” particularly as AI moves from assisting people toward acting with increasing autonomy. I'm increasingly convinced that the challenge isn't simply AI adoption - it's continuous organizational adaptation.

Three ideas stood out this week.

1. The target keeps moving.

Organizations are discovering that AI can't simply be added to existing processes. Workflows, roles, decision rights, and operating models need to change around it.

In 1970, Roger Conant and W. Ross Ashby argued that to effectively manage a complex system, you need an accurate understanding of how that system works. And when the system changes, your understanding has to change with it.

That's the challenge organizations now face with AI: the thing they're trying to manage may be changing faster than their ability to understand and adapt to it.

2. Governance may need to become an adaptive capability.

Cisco recently began deploying personalized AI agents across its workforce of approximately 90,000 people. What's particularly interesting isn't simply the scale, but how Cisco is approaching control.

The agents can have considerable freedom in determining how to accomplish an objective, while consequential actions remain bounded by permissions, policies, human approvals, monitoring, and accountable ownership.

That illustrates an important distinction: Intelligence, autonomy, and authority don't have to increase together. As organizations gain experience and confidence, authority can expand alongside their ability to observe, evaluate, and govern it. Governance becomes less like a fixed rulebook and more like an operating capability that evolves with the technology.

3. Human expertise may be part of the control system.

AI can make experienced people dramatically more productive. But what happens when AI begins performing the work through which inexperienced people traditionally became experts?

If humans increasingly supervise AI, they still need the expertise and judgment to recognize when it's wrong. A human “in the loop” isn't much protection if that person can no longer effectively evaluate the machine.

Where This Leads

That leads to the question I'll be exploring next: What organizational capabilities allow companies to continuously adapt to increasingly capable AI while safely delegating greater autonomy and preserving meaningful human control?

I'm continuing to interview senior leaders who have meaningful responsibility for introducing, scaling, governing, or creating business value with AI. Successes, failures and false starts are all useful.

If you're working through these issues - or know someone who is - I'd welcome the conversation.

----------------------------------------

Research Update #2: Can Organizations Keep Pace with AI?

August 21, 2026

Another week of conversations has shifted my thinking about enterprise AI. Last week, I wrote about starting with business problems, creating space for experimentation, and using governance to make that experimentation safe. This week, a larger pattern is emerging: The challenge may not simply be adopting AI. It may be building organizations capable of adapting as quickly as AI itself is changing.

A few observations from this week's conversations:

1. AI transformation may have its own J-Curve

Major technologies rarely create their full value simply by being deployed. The Productivity J-Curve describes a pattern seen with previous technology transformations: organizations invest in and adopt a new technology, but the expected productivity and ROI don't immediately appear. Value comes later, after organizations redesign processes, roles, skills, and structures around the new capability. AI appears to be following the same basic pattern - but potentially at a dramatically faster pace. That raises an uncomfortable question:

Can organizations redesign themselves quickly enough to keep pace with the technology?

If not, some of today's disappointing AI ROI may reflect an adaptation gap, rather than a failure of the technology itself.

2. Autonomy may be a ladder, not a leap

One enterprise example this week offered a useful progression:

Business problem → constrained AI → demonstrated value → guardrails and learning → progressively greater autonomy

Rather than jumping directly to autonomous agents, the organization deliberately expands what AI can do as confidence, capability and controls mature. That suggests an important principle:

The level of autonomy we delegate should grow alongside our ability to govern it.

3. Agentic AI changes the governance problem

Traditional software executes logic we explicitly prescribe. With an AI agent, we increasingly specify the goal and allow the system to determine how to achieve it. That means governance has to address behavior we didn't explicitly program - and may not fully understand.

Recent work in mechanistic interpretability makes this particularly interesting. Researchers are beginning to identify internal representations and computational pathways inside large language models. One implication is especially important:

An AI's explanation of why it did something may not faithfully represent the internal process that actually produced the behavior.

So explainability and auditability may not be the same thing.

4. The emerging issue may be continuous adaptation

Put these ideas together and I'm beginning to wonder whether we're framing AI transformation too much like previous transformation programs - with a future-state operating model we're trying to reach.

What if the target doesn't stay still?

AI capability continues to advance. Organizations redesign around it. Before that redesign is complete, the capability changes again. The leadership challenge may therefore be shifting from:

“How do we successfully adopt AI?” to:

“How do we build an organization capable of continuously adapting as AI itself continues to change?”

That question is becoming one of the central threads I'm exploring.

If you're working through these issues - or know someone who is - I'd welcome the conversation.

----------------------------------------

Research Update #1

August 14, 2026

I’m researching why some organizations are able to create enterprise value with AI while others struggle to move beyond individual productivity gains. I’m interviewing senior leaders to understand the organizational conditions that make the difference.

A few themes are beginning to emerge...

1. Productivity isn't the same as Enterprise Value

AI is already helping individuals work faster. What's much harder is using AI to create value across an organization. And, the costs matter - licenses, tokens, infrastructure and support can add up quickly. Productivity gains only matter if the value created exceeds the cost.

2. Start with the Business Problem, not the AI

This keeps coming up in the interviews: organizations frequently struggle to clearly define the problem they want AI to solve. Without that clarity, it's hard to know what to build, how sophisticated it needs to be, or whether it's worth the cost. Start with the business need, then ask what AI can do about it.

3. Experimentation requires Governance

Organizations need to experiment because we're still discovering what AI can do. But agentic AI raises the stakes: agents can access data and systems - and take actions. That means some guardrails need to exist before experimentation begins.

Two approaches I've heard: a) tightly isolated sandboxes, and b) Zero Trust / Least Privilege – that is, give AI access only to the minimum data, systems and actions it actually needs.

The conclusion: Good governance doesn't prevent experimentation - it makes safe experimentation possible.

4. The Issues are Interconnected

Strategy, leadership, culture, governance, decision rights, workflows and operating models keep coming up - but they're not independent. People need psychological safety to experiment. Experimentation needs guardrails. Guardrails require clear decision rights. And those decisions need knowledgeable leaders with executive backing. The relationships among these factors may matter as much as the factors themselves.

Finally, I'm comparing what I'm learning with established guidance including ISO 42001, NIST, the EU AI Act, PMI, Microsoft and OWASP. The goal isn't to create another AI governance framework - I'm trying to understand what has to be true inside an organization for good AI practices to actually work and create value.

Join the discussion on LinkedIn

----------------------------------------

Research Updates

August 6, 2026

Designed to Fail? - Launching the Research

Every leader I talk to is asking the same questions: "We bought the licenses. We ran the training. Why isn't our bottom line moving?"

Here's the question I've been wrestling with: What if most organizations are structurally built to reject the true value of AI?

Much of the enterprise AI conversation is trapped in what I think of as the "Productivity Sandbox" - we are obsessed with making individual employees 10% faster at tasks they already do.

But the biggest opportunity isn't making existing work faster; it's discovering entirely new ways to create value that are only possible because of AI.

That means building new, AI-driven business capabilities at the team and functional level. And that is where things get much harder.

What if the organizational structures that make mature companies great at execution - rigid governance, predictable ROI models, strict decision-rights - are the same structures that suffocate AI exploration?

Maybe you don’t have an AI technology problem. Maybe you have an organizational design problem.

To explore this matter, I’ve launched a new research initiative:

DESIGNED TO FAIL?
An interview-based research project exploring the organizational conditions that determine whether AI capabilities can take root and scale.

We’re looking beyond the tech, models, and prompt engineering to understand which governance models, decision-rights, cultural conditions, and organizational structures allow new AI capabilities to take root and scale - and which get in the way.

I’m looking to interview Directors and VPs who have had meaningful responsibility for introducing, scaling, governing, or creating business value with AI.

Success stories are welcome. So are failures, false starts, and scars.

If you’ve been trying to move your organization from “AI tools” to “AI capabilities,” I’d like to hear what you’ve learned.

Interviews are 30 minutes and confidential. Contact me if you’re interested - or if someone in your network has relevant experience, I’d be grateful for an introduction or repost.

Let’s figure out how to stop building organizations that reject the future.

Join the discussion on LinkedIn

bridge light consulting

Research