Delegation Ladder
A four-level test that sorts each step of a workflow into Assist, Recommend, Execute or Never delegate. The level depends on volume, consequence, reversibility, how clear the rules are and whether the decision needs legitimacy a model cannot supply.
What the Delegation Ladder is
The delegation ladder is a four-level test that sorts each step of a workflow into Assist, Recommend, Execute or Never delegate. It answers a question that most AI projects skip: not whether a tool can do a piece of work, but how much authority it should be given over that work.
The task and decision article sets it out as a response to a common mistake. A manager approves “AI for the intake process” and believes one thing has been approved. In fact a chain of steps has been approved, and the steps do not carry the same risk. The ladder breaks the chain apart so that each step can be given its own answer.
The site’s definition says the level depends on volume, consequence, reversibility, how clear the rules are and whether the decision needs legitimacy a model cannot supply. This page walks through those factors and the four levels they produce.
Why a Ladder and not a switch
Many people picture automation as a single dial, with “more automated” at one end and “less automated” at the other. The article argues that this picture hides the thing that matters. AI is not equally reliable across every kind of work, so one dial cannot describe it.
The article reports that AI can automate up to 45% of work activities, but only 5% of jobs are fully automatable. It explains the distance between those two figures as the difference between tasks and decisions. A task can be checked against a right answer and handed over completely once the rules are set. A decision carries a consequence, and a person somewhere has to explain the outcome.
A ladder gives more than two settings. A step can be left entirely to a person, supported by AI, proposed by AI and approved by a person, or done by AI inside clear rules. Each of those is a sensible place for some step in some business, and the ladder names them so that the choice is made on purpose. The article describes the ladder as adapted from a decision-delegation framework.
The four levels
Assist. AI drafts, gathers and compares options, and the person decides. The article says this level fits ambiguous, strategic calls, where there is no fixed answer to check against and the value of AI is a better-informed person.
Recommend. AI ranks or suggests, and a person stays accountable. The article says this fits structured, data-rich calls such as fraud alerts or candidate shortlisting. The AI does the sorting. The person owns the result.
Execute. AI acts inside clear rules, without supervision. The article says this fits high-volume, low-consequence, reversible steps. Nobody should be watching a step that never needed watching, and the article notes that correctly sorted Execute steps are where the fast, measurable return comes from.
Never delegate. Hiring, dismissal and anything with legal, safety or rights implications. The article is direct that these stay human regardless of capability. They are not on a path that a better model eventually unlocks.
The five questions that sort a step
The rungs are not set by what the technology can attempt. The article says they are set by five questions:
- How often does the work happen?
- What does a wrong outcome cost?
- Can it be undone?
- How clear are the governing rules already?
- Does the outcome need a kind of legitimacy a model cannot hold?
The same five questions reappear in the small business article, applied to the smallest possible piece of work. There they decide whether a task is a safe first project. The question about what a wrong result costs works there as a gate: if the answer is an upset customer, a wrong figure in a document that leaves the building, or an unfair outcome for a person, the task is not a first task.
Because the questions are the same at every scale, the ladder works for a single email routine and for a whole department.
Two requests that land on Different rungs
The article’s clearest example puts two similar-looking pieces of work side by side. A password reset and a candidate shortlist are both, in a narrow sense, deciding something about a request.
The password reset is reversible in seconds, happens thousands of times a day, and is governed by a rule with no ambiguity in it. The request either matches the account or it does not. Every question the ladder asks about it comes back low-risk, so it climbs all the way to Execute.
Candidate shortlisting happens far less often, is much harder to undo once a candidate has been told no, and is judged against criteria that are never perfectly fixed in advance. It stops at Recommend, one rung lower, because a person has to stay accountable for the final call however good the ranking underneath it is.
The difference has nothing to do with which task AI is technically capable of doing. It is a difference in consequence, reversibility and clarity, and the ladder makes that visible.
What the Accuracy figures say
The article gives two accuracy comparisons, and they explain why the rungs sit where they do.
In ambiguous, judgement-heavy situations, it reports AI decision accuracy of around 65%, against roughly 85% for experienced humans working the same call. That is exactly the territory where a decision, as opposed to a task, tends to live.
On structured, data-rich predictions such as candidate shortlisting, AI reaches about 75 to 80% against 85 to 90% for experienced humans. The article calls this close enough to support the call and not close enough to replace it.
The article is careful about what these numbers do not say. Neither is a verdict on the technology overall. They describe where AI’s error rate sits relative to a person’s at each end of the ambiguity scale, and that is the information a business needs before deciding how much authority to hand over. A system that is further behind on the hardest calls than on the easiest ones is two different tools doing two different jobs.
Using the Ladder to scope a Workflow
The practical use of the ladder is a scoping exercise. Every real AI workflow is a chain of steps, and each step belongs on its own rung. The article’s example is an intake process where the intake might be Execute, the triage might be Recommend, and the final call might be Assist at best, or Never delegate outright.
Treating the whole chain as one automation decision is how a task-shaped process quietly starts making decision-shaped calls nobody scoped it to make. The article says this is usually discovered only after something has gone out the door, when somebody asks who signed off on it.
That is the idea behind the site’s meaning of a well-scoped workflow: every step has already been sorted onto the ladder, and the whole is not approved as one undifferentiated block. The well-scoped article adds that the return concentrates where a workflow is narrow and well scoped, and that broad, loosely defined rollouts are where the return goes missing.
The task and decision article also explains why sorting changes how long the return takes. Steps correctly placed at Execute deliver a fast, measurable return. Steps left at Assist or Recommend take longer to show one, because the return there is a better-informed person making the call faster, and not a person removed from the loop. Expecting an Assist step to deliver Execute speed is a common reason a workflow gets judged a disappointment when the fault was in how it was scoped.
Where a person sits at each level
Every rung except Execute keeps a person in the decision, and the form of the involvement differs. At Assist the person decides. At Recommend the person approves and is accountable. At Never delegate the person does the work.
This is where the ladder meets the human in the loop page. That page says the review has to name who checks, at what stage and against what standard. A rule that says “a person should check” is a hope, not a control. The ladder tells a business which steps need a reviewer. The human in the loop page says what a reviewer has to be.
The small business article makes the same point from the other side. It says the checking step is not an afterthought: decide before you start who reads the output, what they compare it against and what they do when it is wrong. A task with a defined checking step is one you can improve. A task without one only gets quietly worse, or quietly trusted.
How a step moves up without anyone deciding
The article warns about one specific failure. Projects that stalled after a promising pilot were disproportionately ones where a Recommend step had quietly been treated as Execute once the demo looked good. Nobody decided to give the tool more authority. It happened because the tool proved reliable on the easy cases and nobody revisited the original design.
The drift page describes this as a second kind of drift, one that happens inside a workflow. A step that was scoped as Recommend, with a named reviewer, slides up a rung. The article says catching this at the scoping stage costs an afternoon and catching it after launch costs a rebuild, and sometimes an apology to whoever the wrong call affected.
The remedy is to keep the sorting alive. Know which rung each step sits on, record it, and revisit it when the tool or the process changes.
Governance, Audit and the Law
The article ties the ladder to three other things the site covers.
Governance. A written AI governance policy is, among other things, a record of which rung each automated process sits on and who signed off on putting it there. Without that record, a business finds out a process crossed a line only after something goes wrong.
Audit. The article says an audit is largely this same ladder applied to processes that are already running, rather than ones still on a whiteboard. It is often the first time anyone has looked at the whole chain and not only the one step that prompted the review. What an AI audit actually checks covers that in more detail.
The EU AI Act. The article says the EU AI Act draws its own line around what counts as a high-risk AI system, and that the categories it names, employment decisions among them, map closely onto the Never delegate rung. The article treats the overlap as no coincidence, since the legal framework and the practical test are answering the same question: which decisions carry enough consequence that no amount of model accuracy changes the answer.
This page does not interpret the law. For what the Act requires, read the EU AI Act page.
What the Ladder is not
It is not an argument against automation. The article says so plainly. It is an argument for measuring the right thing when deciding how far to take it.
It is not a one-off exercise either. The people running the workflow day to day need to know which rung each step sits on, and the article points to how to brief a team on AI as the preparation that makes the ladder familiar before a step is being sorted.
And it is not a promise about results. The ladder does not say a step will work. It says where a person has to stay, so that when a step does not work, someone is already accountable for noticing.
How we use it
Before automating a process, list its individual steps and sort each one onto the ladder instead of approving the chain as a whole. Anything high-volume, reversible and governed by clear rules can move to Execute. Anything ambiguous or judgement-heavy stays at Assist or Recommend, with a named person accountable for the call. Anything touching hiring, dismissal or a legal or safety obligation stays off the ladder entirely.
If that sorting has not been done for the processes already running in your business, a discovery call is where it starts, using the same five questions on what is running today. See workflow design for where each step is drawn, and the difference between automating a task and automating a decision for the full argument.
Sources
- The difference between automating a task and automating a decision: the four levels, the five questions, the accuracy figures, the password reset and shortlisting examples, and the links to governance, audit and the EU AI Act.
- Which AI tasks should a small business start with first?: the five questions applied to a first task and the lower rungs.
- Well-scoped AI workflow: why the return concentrates in narrow, well-scoped workflows.
- Human in the loop and drift: what a reviewer has to be, and how a step slides up a rung.
FAQ
Questions about Delegation Ladder
What is the Delegation Ladder?
A four-level test that sorts each step of a workflow into Assist, Recommend, Execute or Never delegate. The task and decision article says the level depends on volume, consequence, reversibility, how clear the rules are and whether the decision needs legitimacy a model cannot supply.
What are the four levels?
At Assist, AI drafts, gathers and compares, and the person decides. At Recommend, AI ranks or suggests and a person stays accountable. At Execute, AI acts inside clear rules without supervision. At Never delegate, the work stays human whatever the technology can do.
Why not treat a Whole Workflow as one Decision?
Because every real workflow is a chain of steps, and each belongs on its own rung. The article says approving the chain as one block is how a task-shaped process starts making decision-shaped calls nobody scoped it to make.
What always stays at Never delegate?
Hiring, dismissal and anything with legal, safety or rights implications. The article says these sit permanently at Never delegate, whatever the technology in front of them can do.
Why is a password reset at Execute but candidate shortlisting at Recommend?
The article says a password reset is reversible in seconds, happens thousands of times a day and follows a rule with no ambiguity. Shortlisting is rarer, harder to undo once a candidate has been told no, and judged against criteria never fully fixed in advance, so a person stays accountable.
What does AI Accuracy have to do with the Ladder?
The article reports accuracy of around 65% in ambiguous, judgement-heavy situations against roughly 85% for experienced humans, and 75 to 80% against 85 to 90% on structured calls such as shortlisting. It reads those gaps as the reason to treat the two kinds of work differently.
Where should a Small Business start on the Ladder?
On the lower rungs. The small business article says the tasks that pass its five-question test sit where AI assists or proposes and a person decides, because the aim of a first project is to keep a person reading every result, not to hand anything over.
How does a step end up on the wrong rung?
The article says projects that stalled after a promising pilot were disproportionately ones where a Recommend step was quietly treated as Execute once the demo looked good. Sorting the process before anything is built, and revisiting it, is what prevents that.
What it means for a Business
A well-scoped AI workflow is one where every step has already been sorted onto the ladder, not approved as one undifferentiated block.
Ready to Put This to Work?
Tell us where your team is with AI and we will tell you honestly what would make the biggest difference.