Data Report · RedactSure Research
Chris Sowa · Founder, RedactSure
Published
Last updated
What Is the AI Delegation Frontier? Harry Ghuman on Who Gave the AI Its Authority
From HAL: Who Signed Off on This?, with Harry Ghuman, CEO of Alpha Decisions
Episode: The AI Delegation Frontier. Listen on Libsyn. The show is also on Spotify and YouTube.
In brief
-
An enterprise is a bundle of decisions. AI creates value, and risk, where it takes part in them.
-
When an agent acts, it did not give itself authority. Someone did, and governance has to say who, for what, and for how long.
-
Agent authority should be mapped to business processes. A $10 refund and a $100,000 refund need different controls.
-
Human in the loop does not scale to 10,000 decisions an hour. Decide the delegation frontier instead: what you delegate and what you never will.
-
Someone has to own the benefit in the business case, and outcomes pursued without controls invite damage.
-
Cutting entry-level roles removes the pipeline of people who will have judgment five years from now.
About the episode
Harry Ghuman is CEO of Alpha Decisions. He spent more than 16 years at Oracle and served as a senior vice president at Palo Alto Networks, so he has seen enterprise transformation from both the value side and the security side. His view is that a company is a set of decisions, which makes AI governance a question of authority: which decisions an agent may take, and who gave it the right to take them.
On HAL: Who Signed Off on This? he explains why so many AI pilots show little value, why agents are assuming more authority, why governance is broader than cybersecurity, why a human in the loop is not a complete answer, and why cutting junior roles is a risk to the judgment a company will need later.
Highlights from the conversation
Where do the decisions that create advantage live?
Chris Sowa: In some of your writings you argue that strategy turns into performance through decision quality, not through technology or data. So where in a company do the decisions that create advantage actually live? And why does that matter more now that AI can take part of them?
Harry Ghuman: An enterprise is like a bundle of decisions. And every department, whether you are in marketing, you're in finance, you are in accounting, at the end of the day, the manager is supposed to make some decision. Every worker needs to make some decision. Nothing can move forward until you can make a decision. So decisions and capabilities, they sort of come together and then you create a business out of it.
And so the same conversation is actually even becoming more important now because yes, AI can give a lot of things for research and things like that, but then used in business, it is actually capable of making a lot of decisions. And that's where a lot of the benefit is and a lot of the risk is.
Who gave the AI its authority?
Chris Sowa: You've described a CEO with five hundred pilots, forty seven vendors and no movement in operating margin. Based on all your years of experience, why aren't people getting bang for their buck for all this work?
Harry Ghuman: Let me first clarify this example. It is a composite example. Just to keep the client confidentiality, I've actually taken two examples and joined them together. But this is actually a very, very common scenario. Everybody's actually facing this.
AI actually took the action, but it didn't give itself the legitimate authority. Somebody else gave it that authority. So you can actually look at that: why was the agent able to do it? And maybe there were some design criteria where it could actually give itself the authority, or it was just a hidden authority, or you thought that it wouldn't do it. This cycle is actually changing every day. If you keep reading about the examples, the agents are becoming smarter and smarter in terms of assuming more authority. And that's where a lot of the controls really need to come in.
Hidden authority is the risk to look for. An agent that inherits a user's full access has every right the user has, whether or not anyone decided the agent should. The fix is to grant authority deliberately, for a task, and to record who granted it.
Is this a security problem or a governance problem?
Chris Sowa: The people implementing this technology, a lot of them know. What does it take to get them to actually do the right thing and put the controls in place, rather than just take on that risk? How did you go about this at Palo Alto Networks?
Harry Ghuman: Things have progressed quite a bit in the whole agent and risk area. When you're looking at the security, that's one topic. But then there's the authority as well. And so you have to actually not just think about it from a cybersecurity standpoint. A lot of it is the governance standpoint as well: what is the agent authorized to do, and where are the controls? Those things have actually mushroomed quite a bit in the last several weeks in terms of the agent being able to go different places.
So it's part of the enterprise design, as opposed to just thinking about one part. You assume that there are people and agents that both have identities. Some of the identities, you actually give authority to one or the other. And once you actually think about all of the main business processes and try to map it to what might be happening here, make sure that you consider all the possibilities that could happen and you put effective controls around them.
How do you size the risk before it arrives?
Chris Sowa: The threat isn't imminent until it's in their house. You have designed methodologies and explained to people how to quantify risk before it happens.
Harry Ghuman: Risk is everywhere, threats are everywhere. First, you have to understand what is it that you're trying to protect. That's where a lot of the past thinking about cybersecurity has been: these are the types of things that we want to protect, the crown jewels of the company, and then the employee IDs, and then you actually create the mechanism by which you can control that. When you're actually thinking about AI, it's everywhere. First you have to worry about the people. And now agents can be so many more than the people in the company, and it's sort of hard to keep track of what all of these agents are doing.
But we still need to start from the business process: what are the main business processes that the client has, and who in there is authorized to do what. If you start from that side, then you can actually figure out what authority the people have, what authority the agents have, and which type of authority it is. Is it permanent authority, or is it for a few actions? The fundamental thing is looking at which process it is. It's just like issue a $10 refund or create a $100,000 refund. At some point you have to think about what would be the next control.
Permanent authority or authority for a few actions is the practical design choice. Task-scoped authority that ends when the task ends, with a threshold where a person must approve, is how the $10 refund and the $100,000 refund end up with different controls.
Should companies start small or go after core processes?
Chris Sowa: Some leaders check the box. They roll out Copilot so teams can organize their inbox, and at the end of the year it didn't generate much value. On the other hand, it's a little scary to know your most important supply chain process can get the most value from AI. Do we start small and accept that we won't get the value, or go after the core processes that are critical?
Harry Ghuman: Yes, and yes, because what I'd say is that companies should actually continue to experiment, but don't have a lot of expectation that it is every time going to give you a measurable result. Because if you don't experiment, you cannot really progress to the more consequential decisions that you can actually apply AI to, and the more consequential processes. Every company knows what their most consequential processes are, where the money comes in from and where the costs are.
But if you want to apply AI to the more consequential processes, you have to first draw up a plan that this is what we are going to do. Based on that, whatever the goals that you set up, somebody has to own those benefits that you wrote in the plan. The business case says that this will be the benefit, and then that person should be accountable for the benefit.
What happens when an agent gets an outcome but no controls?
Harry Ghuman: This is actually very relevant to agentic AI. You can give an agent: this is the outcome that I'm looking for, and this is the mission. And they just don't have all the controls, and then they get out of their sandbox and do some damage, and then everybody starts looking at how did this happen? But anytime you actually look at creating an outcome, you have a lot of responsibilities as well. You have to do things ethically, you have to do things correctly. AI is a tool.
Is a human in the loop enough?
Chris Sowa: Are there some processes where you say, I want to keep a person in the loop? Even if the machine is right ninety-eight percent of the time, you would feel strongly that a human should still be in the loop for that two percent.
Harry Ghuman: Maybe that is correct. But it's not the one answer for everything. Because the AI can produce these decisions thousands of times faster than a person can actually do anything about them. If you say human in the loop and everything is going to be fine, but the agent is producing 10,000 decisions in an hour, what is the human going to do?
So what is it that you can delegate to the agent? Think about it as a delegation frontier: what would you delegate, and what is it absolutely you would not delegate? The most consequential decisions, that's where you want the human to be there, but it's not the answer for everything, because you would not be able to keep up.
The delegation frontier turns a slogan into a design. Decide in advance which actions an agent may take alone, which need a named approver, and which never leave a person's hands. Approvals then go where they carry weight, instead of being spread so thin that they become a rubber stamp.
Where will judgment come from in five years?
Harry Ghuman: If you're going to commit the company to doing something extremely consequential to their survival, I would not put that in the hands of an AI. You have to think about the judgment, because humans have this judgment that they've developed over taking those decisions many times. And many of those judgments, so far, I've not seen that can actually be built into the AI.
Now that creates a very interesting question. A lot of the business cases, in this current generation of AI and before as well, in technology transformation, people think: I'm just going to let go of the junior people. I'm not going to get entry level people. The point is that if you don't get the new employees, the fresh employees, how are you going to get the people five years from now who have judgment? Senior people were junior people at some point. Partners were interns.
What executives can take from this conversation
-
Map agent authority to business processes, and decide whether each grant is permanent or limited to a task.
-
Look for hidden authority: agents that inherit a user's full access without anyone deciding they should.
-
Set thresholds where a named person must approve, so a $10 refund and a $100,000 refund are controlled differently.
-
Write down your delegation frontier: what agents may do alone, what needs approval, and what stays with people.
-
Name the owner of every benefit in an AI business case.
-
Protect the entry-level roles that build tomorrow's judgment.
Where RedactSure fits
Authority is the design question RedactSure is built around. Under Supervised Delegation, a named person confirms the policy for a task before it runs and approves consequential actions at the moment they happen. The agent works on tokens in place of sensitive values and cannot settle, pay or decide on its own. Everyone else controls where data can go. RedactSure controls what the AI can see.
Frequently asked questions
What is the AI delegation frontier?
Harry Ghuman's term for the line between decisions an organization delegates to AI agents and decisions it keeps for people. Because agents can produce thousands of decisions an hour, a person cannot review them all. The frontier decides in advance where human judgment is required, usually the most consequential decisions.
Why is human in the loop not enough for AI agents?
Agents work far faster than people can review. A human in the loop on every action either slows the work to a halt or becomes a formality. He recommends putting humans where decisions are consequential and controlling the rest through scoped authority.
Who gives an AI agent its authority?
A person or a design decision always does. Agents can also accumulate hidden authority, for example by inheriting a user's access. Governance means granting authority deliberately, mapping it to business processes and recording who granted it.