The Agentic Delegation Boundary

Problem: Your customers now hand parts of each purchase to AI, and you cannot see where the machine stops.

Outcome: Knowing where that handover sits for each buyer and intent tells you what to fix first, and how fast.

Why it matters: The further the machine goes, the more your sale depends on what it can verify.

Why trust this: Parasuraman, Sheridan, and Wickens graded automation in 2000; the Agentic Delegation Boundary uses that established problem to locate the handover inside one workflow.

Automation research graded the machine, and the Delegation Boundary finds the handover

Back in 2000, Raja Parasuraman, Thomas Sheridan and Christopher Wickens published a model for types and levels of human interaction with automation that engineers still use to decide how much of a task a machine should perform, from gathering information to acting on a decision, graded from no help at all to full autonomy. Their question is about the machine, and the answer is a level. Take that same question inside a single piece of work, turn round to face it from where you're sitting, and it becomes a question about a point: at which step does the machine's work stop and yours resume?

The Delegation Boundary marks where delegated machine work ends and the human takes over

You hand work to an AI assistant or an agent every day, whether that's research, writing, coding, booking a table or buying a replacement printer cartridge. Somewhere inside each of those workflows, the machine stops and you pick it up, and that handover is the Delegation Boundary.

The definition rests on two things you can observe: who does the work, and who holds control. Before the boundary, the machine does the work and controls each step. After it, you do. Nothing else is needed to locate it, which is why the same idea applies to a literature search, a first draft, a pull request or a weekly grocery order.

In e-commerce, the workflow is usually the purchase or the conversion, so that's where most businesses meet the Delegation Boundary: the moment their customer either takes over from the machine or lets the machine carry on to the checkout.

Search, assistants and agents put the Delegation Boundary in three different places

After ranking, after recommending, after acting

Search finds and ranks, then you take over: you read the options, you compare, you choose and you act. An assistant goes further, because it also assesses the options and recommends one, so you take over at the choice. An agent can go further still, because it can also choose and act, which means the boundary can sit after the transaction itself, and you only step in when it's done.

That last crossing is why the term is agentic. Machines have been finding and ranking for you for a quarter of a century, and recommending for a few years now, but when choosing and acting move to the machine side of the boundary, the machine is committing you to something. Acting means committing the person, and that need not include paying: an agent that books your table, accepts a meeting or signs you up to a newsletter has acted on your behalf just as surely as one that spends your money.

The Delegation Boundary, responsibility and confidence each answer a different question

The Delegation Boundary sits alongside two other ideas, and keeping them apart is what keeps the definition clean:

Work and control can be observed and measured. Responsibility explains why the boundary matters. Confidence explains whether the machine can reach it.

The actual boundary can stop short of the intended one, and that is the machine returning control

You set an intended Delegation Boundary every time you delegate: it's the point you authorised the machine to work up to. The actual Delegation Boundary is where the machine really stops. Most of the time they coincide, because the machine has what it needs and carries on to the point you allowed.

Say you tell your agent "buy me dinner". You've authorised it to work all the way through to the purchase, so the intended boundary sits after the transaction. The agent finds restaurants, assesses them and narrows the list, then hits something important it can't verify, and instead of guessing, it stops before choosing and asks you. The actual boundary has moved back up the workflow, to before the choice.

find -> assess -> recommend -> choose -> act
                                             ^ intended (after the purchase)
                    ^ actual (agent returns control before choosing)

Look carefully at what changed hands, because it's control, and only control. You haven't taken back responsibility: you were accountable for your dinner from the start, and you still are. The machine has returned control to you, because it couldn't reach the boundary you set.

For me, this is the gap that decides who gets the sale. Whether the agent reaches the intended boundary depends on its confidence: in the business it's considering, in what it knows about you, and in whether the choice will meet your need. Each of those rests on information that has to be sufficient, current and corroborated, and when any of it falls short, the agent hands the decision back to you, and you might well choose someone else.

The Delegation Boundary is neither the mandate nor the Agentic Spectrum

The mandate is your authorisation. It sets the intended Delegation Boundary and carries the limits inside each step: a budget, a delivery date, the brands you'll never buy. The mandate tells the machine how far it may go and on what terms, and the Delegation Boundary is where the work actually changes hands.

The Agentic Spectrum is a business measure: the share of a company's customers who let an agent buy on their behalf. You read it across a whole customer base, whereas the Delegation Boundary sits inside one workflow, for one person, on one occasion.

Seven factors decide where the Delegation Boundary sits for any purchase

Emotional weight, domain expertise, price relative to income, frequency, reversibility, regulatory context, cultural context

The same person sets the boundary in very different places from one purchase to the next. You'll happily let an agent handle the routine reorder of coffee pods, choosing and paying without a second thought, because the emotional weight is low, the price is small against your income, you buy them often, a mistake is easy to reverse, and nothing regulates the choice. The wedding ring sits at the other end, because it carries every ounce of emotional weight, costs a great deal, happens once and can't really be undone, and you'll want the search results and the advice, then make the choice and the purchase yourself.

Domain expertise pulls both ways, since people who know they don't know a field often delegate more, while people who believe they do delegate less. Regulatory context sets hard limits in finance, health and law, and cultural context shifts how far people trust an agent from one market to the next.

Where your customers' boundary sits tells you what to rebuild, and how fast

You read the Delegation Boundary for each ideal customer profile and intent, as a standard for that combination, with individual customers scattered around it. That standard tells you what work you need to do.

Where your customers take over after the machine finds and ranks, nothing is new: the search optimisation you already do still wins the click. Where the machine also assesses and recommends, you need to educate large language models, search engines and knowledge graphs at once, and be among the top answers for very specific requests, because the assistant names a shortlist and your customer takes over from there. Where the machine acts, you need to rebuild how you sell, transact and serve, so an agent can complete the purchase with you without handing control back to your customer.

The boundary isn't standing still, either. It moves towards the act as people grow used to delegating and as machines earn the confidence to carry on, and how fast it's moving for your customers sets how fast you have to change.

The definition fits in one sentence

The Agentic Delegation Boundary is where the machine's work ends and yours begins, and every business now has to know where that point sits for its customers.