The Delegation Ladder: how much can AI do before a person signs?

Why is "should we use AI?" the wrong question?

Most organisations have already answered it. The harder question is how much the system is allowed to do before a person is involved, and that question is becoming urgent. Gartner forecasts that at least 15% of day-to-day work decisions will be made autonomously by AI agents by 2028, up from 0% in 2024. It also forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027, for rising costs, unclear value or weak risk controls (Gartner, June 2025; forecasts).

Both forecasts describe the same gap: authority given to systems without a clear decision about how much, and on what evidence. The Delegation Ladder makes that decision explicit.

What are the four levels?

Each level hands the system more authority. Each also needs something in place before it is safe to fund.

Recoverable or binding: what is the difference?

The second test is about consequences, and it applies at every level.

• Recoverable: a mistake stays inside the organisation and can be undone cheaply. A wrongly closed internal ticket can be reopened.
• Binding: a mistake reaches a citizen or customer, money or legal rights, or cannot be undone. A wrong reply to a licence applicant, or a wrong credit decision, is binding.

The same system can sit at a different place depending on what it touches. Drafting internal meeting notes and drafting a reply to a customer are both Draft, but only the second is binding, so it needs a stronger check before release.

How do you place an idea on the ladder?

Three questions place any idea:

1. What kind of system is this? The answer is its level.
2. What does it commit us to? The answer is the list of things that must exist before it is funded.
3. Who carries the risk if it fails? The answer is its consequence, recoverable or binding, and the named person accountable.

From there, leaders reach one of three verdicts: fund now; fund once the must-exist items are in place; or move the idea down a level. That third verdict is more useful than it sounds. Many ideas that are unsafe as Delegate can be funded straight away as Draft, because a person still signs before anything leaves.

What evidence moves an idea up a level?

Evidence from the organisation's own operations, not a vendor's demonstration. The research shows why: in a randomised study of 758 consultants, people using AI were 19 percentage points less likely to reach the right answer on a task the AI handled badly (Dell'Acqua et al., Harvard Business School working paper, 2023). Where AI is wrong, it is often confidently wrong, and people follow it.

So each move up the ladder needs its own proof:

• Draft to Advise: a record of past cases and their outcomes, including mistakes, so the system's recommendations can be checked against how often it would have been right.
• Advise to Delegate: a way to reverse errors cheaply, or a named person for every case, plus monitoring that shows exceptions before they compound.

What mistakes do leadership teams make with AI authority?

• Funding the demo, not the level. A system shown acting autonomously is funded as Delegate before anyone checks whether Draft would deliver most of the value at a fraction of the risk.
• Treating sign-off as a formality. At Draft, the approver signs hundreds of outputs a day without reading them, and the level quietly becomes Delegate.
• Ignoring volume. At Advise, a system that flags too many cases trains reviewers to ignore it.
• Forgetting the source. At Assist, answers drift out of date when no one owns the policy or data the system reads.

Measure where your own team stands: take the free Readiness Snapshot →