When AI ranks, drafts or acts, whose signature is on it? The Influence Ladder sorts AI involvement into four rungs and three questions every signer should answer.
A programme officer asks an assistant to rank forty grant applications against the funder's criteria. She reads the top eight properly, skims the next ten, and never opens the rest. Three months later, an unsuccessful applicant asks why their proposal was not shortlisted.
Who made that decision? The officer will say she did, because she signed the shortlist. The applicant will say a machine did. An auditor will ask a third question, which is the one that matters: what, exactly, did the officer see, and what did she rely on without seeing? Accountability for an AI-shaped decision fails not at the moment of the decision but at the moment someone asks to reconstruct it.
Traditional accountability works because the chain is visible. A report has an author, a reviewer and an approver. AI adds a contributor that does not sign anything, cannot be questioned and may behave differently tomorrow. The person who signs is still fully accountable, but their signature now sits on top of work they only partly observed.
Two failure modes follow. In the first, the signer over-claims: "I reviewed everything", when they reviewed a fraction. In the second, the signer under-claims: "the tool suggested it", which is an abdication and will not satisfy a funder, a court or a board. Neither is dishonest by intent. Both arise because the organisation never defined what a signature on an AI-shaped decision is supposed to mean.
We find it useful to sort AI involvement into four rungs, because the accountability each one demands is different.
The rung is not a property of the tool. The same assistant can sit on Rung 1 in the morning and Rung 3 in the afternoon, depending on what the person asked it to do. That is why the rung has to be declared per decision, not per system.
For every decision above Rung 1, an organisation can require the signer to be able to answer three questions in under two minutes.
If the signer cannot answer, the signature is decorative. That is not a reason to ban the practice; it is a reason to make the answers cheap to capture at the time. A three-line note attached to the decision costs a minute today and saves a week of reconstruction later.
Nobody asks for the audit trail when the decision goes well. They ask for it when it is contested, and by then it can only be written down from memory.
The ladder exposes a tension most organisations would rather not resolve. The efficiency case for AI is largest on Rungs 3 and 4, where humans are removed from the reading. Yet the accountability burden is also highest there. A team that adopts ranking or automation because it is fast, and then keeps no record because that would slow it down, has quietly converted an operational saving into an unpriced liability.
Public and mission-driven organisations feel this more sharply. Decisions about who receives a grant, a service or a place are decisions about people, and the people affected reasonably expect a human answer to the question "why". A tool that cannot explain its filtering does not remove the duty to explain. It relocates that duty to the person whose name is on the letter.
You do not need new software to start. Three steps are enough for a first quarter.
A signature on an AI-shaped decision only means something if the signer can say what they framed, what they checked and what they could not see. The Influence Ladder makes that requirement proportionate: light for research, firm for ranking and automation.
Prepared by M.A.I. Consulting.
If this is the question on your desk, a thirty-minute call tells you whether the service fits, or that you do not need us yet.