What are we really delegating to AI?

We ask for a questionnaire, a report or a briefing. Along the way, the assistant makes choices about what matters. How many of those choices did we knowingly hand over?

When you ask AI to draft an assessment tool, you may also be letting it decide what is worth finding out. When you ask it to write the report, you may be letting it decide which findings deserve attention. This is a question we keep returning to in our AidGPT training, particularly with our ongoing September cohort: what judgement comes with the work we delegate? A request for something “rapid” or “clear” leaves room for the assistant to choose what to leave out, simplify or emphasise. We knowingly hand over the task without necessarily recognising the decisions we have handed over with it.

Consider two hypothetical examples from an emergency response. In the first, a team asks AI to design a rapid market assessment tool. To keep it manageable, the assistant produces a short questionnaire covering prices, stock and traders’ ability to replenish supplies. Questions about whether different groups can safely reach and use the market are left out. The team receives something practical that looks ready to use. But a judgement about what deserves the team’s limited time has been made within the drafting task. If the team accepts the tool, that judgement shapes what it will find out. An issue missing from the questionnaire may remain missing from the evidence used to design the response.

In the second example, the team designs the assessment itself and collects information about access alongside prices and supply. It then asks AI to turn the findings into a report. The assistant has to choose a structure. It leads with market conditions and puts the access findings in a small box near the end. Nothing has been invented, and the information is still there. Yet the report gives prices and supply most of the reader’s attention, while making access easier to treat as a secondary concern. A choice that looks like presentation can influence which problems the team addresses first.

In both cases, the person knows they have asked AI to produce a document. They may not recognise that they have also left it to judge what information is necessary or which findings matter most. The assistant can deliver a polished, apparently helpful result without making those choices obvious. Checking the document for factual mistakes will not, on its own, reveal everything that has been delegated.

This is where I think we need to look more closely. Judgement enters the work well before the final recommendation or approval. It enters through the questions we ask and the way we organise the answers. If those choices help shape the response, we need to recognise them as decisions and establish who has the knowledge to challenge them before the work is used.

What did we actually delegate?

The difficulty is that we usually describe the result we want, rather than every judgement needed to produce it. “Make this quick to use” leaves open what can be shortened or removed. “Write a clear summary” leaves open which findings deserve the most space. An assistant can fill those gaps and deliver something that meets our request, while making choices we would have wanted to discuss if they had been put to us directly.

There are good reasons to let AI make some of these choices. We would gain little from an assistant that needed permission for every heading or sentence. But the boundary matters. Choosing a heading and deciding that safe access needs no separate questions can both happen during the same drafting task. One helps organise the work. The other changes what the team will know. A request to produce a document does not tell the assistant where we intended that boundary to sit.

This also changes what we need to look for when reviewing the result. A short questionnaire may look sensible until we ask what was removed to make it short. A clear report may look convincing until we compare the attention it gives each finding with the evidence behind it. Asking the assistant to identify its assumptions can help us find questions to investigate, but its explanation still needs checking against the actual tool, source material and task. A confident account of what it did is another output to examine.

For me, this is what managing AI means in practice. We need to identify the choices that could change the work, decide which we are comfortable delegating and inspect how they were handled. Approval is more meaningful when the reviewer understands those choices. Even then, the team’s view has limits. We may all agree that a questionnaire is practical or a report is balanced, while missing something that matters greatly to the people whose situation we are describing.

Knowing enough to see what is missing

It is easy to ask an AI assistant to act as an experienced market analyst, protection adviser or public health specialist. It can produce work that uses the right language and looks professionally structured. Assessing that work requires enough knowledge of the subject to recognise what should be there. You might check every statement against its source and still miss an important question the assistant never asked, or a finding it gave too little weight.

An experienced market analyst might look at our short questionnaire and immediately notice that it says little about who can safely use the market. A colleague without that background may see a sensible set of questions about prices, stock and supply. Both could review the tool carefully and reach different conclusions about whether it is ready. The same applies to the report: someone with relevant experience may recognise that a finding placed in a small box should change the whole response.

We cannot assume the assistant will recognise that gap in the person using it. The instruction “act as a market assessment expert” tells it what role to perform. It says little about whether the user has years of experience or is preparing their first assessment. Asking it to identify assumptions and limitations can help, but the user still needs to assess whether that explanation covers the things that matter.

Research from Pakistan gives this concern a practical context. In Local Leadership in the Age of AI, GLOW Consultants, working through the ALLIES partnership convened by Humanitarian Advisory Group, describe staff using AI under resource pressure and tight deadlines. Respondents reported help with proposals, analysis and reporting, while formal training and organisational support remained limited. The report also highlights benefits for staff with strong field knowledge who need support expressing it in English or donor formats. Limited writing support should never be mistaken for limited humanitarian expertise.

The study describes AI mainly assisting human work. It does not establish that staff are replacing missing specialists with AI, or measure their ability to detect subtle omissions. My concern is what happens when an organisation starts relying on AI to cover a technical role it can no longer afford. A capable, committed member of staff may be doing their best with the support available. They can still be left responsible for checking judgements outside their experience.

That distinction matters when we talk about AI saving time. An expert using AI to work faster and a colleague using it to cover a missing specialist may produce equally polished documents. Those documents tell us little about the checking behind them. Before relying on the work, the organisation needs to establish who can recognise what has been omitted, question the priorities and judge whether the result is fit to use. Telling staff to “check the output” is only useful if they have the knowledge, time and support to do it.

Make the delegation visible

Before using your next AI-generated assessment tool, report or briefing, look beyond whether it reads well and contains accurate facts. Ask what the assistant decided in producing it. What did it leave out to keep the work short? Which findings did it give most attention? Then consider whether you have the knowledge to judge those choices, or need someone with different expertise to review them. If you manage a team, make sure staff have the time and support to do this.

This is the practical discipline we teach through AidGPT. Participants learn to decide what to delegate, give the assistant the context it needs and challenge the work it returns. We also look at the system around the assistant: the information it can access, the tools it can use and the points where work needs to return to a person. Across the course, we practise Think, Draft, Review: decide what good work requires, use AI deliberately, then check whether the result is fit to use.

AI training will not turn someone into a market analyst or protection specialist. It can help them recognise where they are relying on the system’s judgement and where further expertise is needed. If you use AI in humanitarian or development work, or supervise colleagues who do, that is what AidGPT is designed to help you practise. You can read participants’ accounts of the course here.

The next time you ask AI to help, ask yourself what judgement comes with the task. Make that delegation a choice you understand.

Tom

AI-use disclosure

I used OpenAI Codex to help develop, draft and revise this article, review the cited research and check references. Claude reviewed earlier drafts and suggested structural changes. I shaped the argument, supplied the assessment examples and directed revisions, including changes to the language and emphasis. I have read the draft and stand behind the argument. Source and link checks were AI-assisted. The market assessment examples are hypothetical and illustrate possible failures, rather than documented incidents.