There is a specific kind of fatigue that comes from good fieldwork followed by bad admin. You spend three days facilitating FPIC consultations or running a training, capture pages of rich notes, and then face the second job: turning all of that into a deck, a report, a proposal, a brief before anyone can actually use it.

For years, that second job ate the time I should have spent on the first one. It is not the thinking that drains you; it is the reformatting.

This is where AI assistants have earned a permanent place in my workflow, not as a replacement for judgment, but as the colleague who handles the repetitive scaffolding so I can spend my energy where it matters: interpretation, nuance, and voice.

I use “AI assistants” here to mean systems that can take a defined body of source material, perform a sequence of tasks such as organizing, summarizing, restructuring or translating it, and return a usable first pass while I remain responsible for review, correction and judgment.

The tasks that used to cost me the most

A few examples of the repetitive work that shows up on most every project, regardless of sector:

Meeting notes → presentation. Raw notes from a stakeholder meeting or client debrief rarely have the structure a presentation needs. Someone has to find the narrative arc, group related points, and build slides that don’t just dump text on a screen. That someone used to be me, at 11 p.m.

Field notes → daily memos. During active fieldwork, there is no gap between one day’s data collection and the next. Notes taken across FGDs, IDIs/KIIs, and observations have to be turned into a clean, structured memo before the next day’s fieldwork starts, or the backlog compounds.

Training documents → onboarding decks. A full training manual is written to teach depth. An onboarding deck needs to teach direction; the essentials, sequenced, without losing the substance underneath. Compressing one into the other by hand is slow and easy to get wrong.

Scattered planning notes → strategy briefs. Strategy conversations happen in fragments; a point here, a risk flagged there, a decision half-made in a margin note. A usable brief needs those fragments pulled into one coherent argument with a clear recommendation at the top.

These tasks do involve judgment, but not all of that judgment needs to be performed from scratch each time. The distinction I have become more interested in is between interpretation and transformation.

I do not want AI deciding what a participant meant, what a finding means in context, or what deserves to influence a research decision. I am much more comfortable asking it to organize material I have already generated into a different form.

In other words: I do not delegate judgment. I delegate transformation.

That is precisely where AI becomes useful: when the source material is mine, the boundaries are clear, and I stay in the loop as the editor.

What actually changes when you delegate this

The honest case for using AI assistants on repetitive tasks isn’t “it does the work for you.” It is narrower and more useful than that:

• You get a structured first draft in minutes, not hours. The assistant reads your raw notes, identifies the natural sections, and produces an outline or draft you can react to, a much faster starting point than a blank page.
• You stay the author. A good draft from an assistant is not a finished product; it is raw material you shape, cut, and correct until it sounds like you and says exactly what you mean. The judgment, the voice, and the final call stay yours.
• Repetition stops being a cost. Once you have done this once (notes to memo, notes to deck, notes to brief) doing it again for the next project takes a fraction of the time, because the pattern is established and the assistant can follow it.
• You reclaim time for the work that can’t be automated: synthesis, interpretation, relationship-building, and the judgment calls that come from actually being in the room.

I give the system a defined task, a defined body of material, and a defined purpose. It can produce the first pass, but it does not get to decide whether the output is accurate, appropriate, ethically sound, or meaningful. That remains my responsibility.

What makes this more than a productivity trick is that the same pattern holds up across very different kinds of work. Three examples from my own practice show what I mean.

Case one: repackaging a training into a standalone program

The first is a three-day qualitative research training I had delivered and documented mostly as working notes; session outlines, exercises, talking points, timing cues, scattered across drafts written at different points as the program evolved.

What I needed wasn’t new content. I needed the material repackaged: a coherent, sequenced, professionally structured program that someone could pick up and immediately understand (session by session, objective by objective) without me sitting beside them explaining the logic.

Working with an AI assistant, that repackaging looked like this: feed in the raw material as-is; ask for structure, not content invention (group related sessions, surface a logical Day 1-2-3 arc, draft clear objectives per module); review every section against my own standard, correcting emphasis and removing anything generic; iterate until it read as mine, reinserting the specific examples and Kiswahili phrasing that make the training recognizably mine, not a template.

The result was “Qualitative Research: A 3-Day Practical Training Program” a packaged, shareable version of a training I had already been running, now structured clearly enough to hand to a client, a co-facilitator, or a partner organization as a standalone offering.

The AI didn’t teach me anything about qualitative methods. It took the organizing off my plate and gave me back the hours to make sure the content was rigorous and true to how I actually train.

That distinction is important because repackaging can easily become distortion. A system can produce a beautifully structured training program while quietly flattening the particular logic, examples or emphasis that made the original training useful.

That is why I treat the AI output as a draft, not as an authority. The structure can be delegated. The standard cannot.

Case two: What happens when the fieldwork is moving fast

The more interesting test came during live fieldwork.

Over two weeks, I was working through a high volume of focus group discussions (FGDs), key informant interviews (KIIs), and structured observations. At the end of each field day, AI helped me turn my debrief notes into structured daily memos; organizing what happened, recurring observations, emerging questions, and issues requiring follow-up. I still reviewed and corrected everything myself.

The same principle applies to qualitative data analysis. As interviews, FGDs, and field notes accumulate, AI can assist with categorizing and organizing qualitative data; grouping similar responses/themes, surfacing recurring concepts, clustering excerpts, and helping structure preliminary themes across a large body of text.

But I draw a clear line between organizing evidence and interpreting it. AI can help me sort, compare, and structure the data; it does not decide what a pattern means, whether it is significant, or how context changes its meaning. That remains the researcher’s work.

A polished memo can create a false sense of certainty. AI can make two observations appear more closely related than they were, smooth over contradictions, or turn a tentative participant statement into something that reads like a firm finding. Good prose is not the same thing as good interpretation.

So the first-pass structure is useful precisely because I do not confuse it with analysis. I still have to ask: What did I actually hear? What did I actually observe? What is missing? What surprised me? What needs to be followed up tomorrow?

Those questions cannot be delegated simply because the notes can.

There is also an important boundary around what enters an AI system in the first place. I do not treat participant information as ordinary working material simply because a tool can process it. Where AI is used around research material, confidentiality, consent, client requirements, institutional policies, data protection and de-identification still apply. The convenience of a faster draft is never a reason to relax those responsibilities.

For me, the principle is simple: delegate the formatting of knowledge without casually outsourcing the responsibility that comes with possessing it.

The value, for me, is simple: less time spent manually organizing data and more time available for the interpretation, questioning, and contextual understanding that qualitative research actually requires.

Case three: closing the language gap without doing all the translation myself

The third example runs in a different direction from the first two: not compressing volume, but closing a language gap that would otherwise have defaulted entirely to me.

On a recent project, several colleagues on the team don't speak Kiswahili, but a set of documents needed a Kiswahili version before they were usable. Historically, that translation work would have landed on me by default, simply because I am the one on the team who moves comfortably between the two languages, and it is slow, careful work precisely because getting the register right (not just the vocabulary) is most of the job.

Instead, those colleagues ran the documents through their institute’s own AI translation tool first. What came back was sufficiently strong that it did not require me to rebuild the translation from scratch. That meant the translation load never landed on me in the first place, while still leaving room for the kind of human review that matters when meaning, context and register are more important than word-for-word equivalence.

That is, I think, where this is heading for transcription and translation more broadly: eventually the first pass on both will likely sit with the AI assistant more than with me, and my role will shift toward review, checking that the substance holds and that the translation actually carries the meaning and register of what was said, not just the words on the page.

That is not a smaller role. Reviewing against what was actually meant is exactly the kind of check that catches errors a first-pass translation, human or AI, can introduce; it minimizes errors rather than just moving them downstream.

Three very different situations, three different kinds of pressure: a backlog to clear, a deadline that resets daily, and a language gap that would otherwise default to me by habit rather than necessity.

The same underlying pattern holds across all three: the thinking, and the judgment over what a memo, a program, or a translation actually needs to say, stays mine. The reshaping doesn't have to be done entirely by hand.

The infrastructure behind the delegation

AI assistants are not a shortcut around expertise. Used well, they are a way of protecting the conditions in which expertise can actually be used.

For me, that means taking the repetitive reshaping of notes, memos, decks, briefs and other working material off my desk so that my attention remains available for the parts of research that require interpretation, context, relationships and judgment. The goal isn’t less researcher involvement, it’s involvement where it actually matters.

That has changed how I think about AI in research. The question I ask now is not simply, “Can AI do this?” It is, “Is this a part of the work that I need to do myself?”

If the answer is no, and the task can be clearly bounded, checked and safely delegated, I would rather let the machine carry it.

The pattern holds whether the pressure is a slow-burn backlog, a daily fieldwork deadline, or a language gap that would otherwise default to you; which is exactly what tells you it is a pattern and not a one-off convenience.

My scarce resource is not information. It is attention.

The takeaway

AI agents are not a shortcut around expertise. Used well, they are a way of protecting it by taking the repetitive reshaping of notes, memos, decks, and briefs off your desk, so the hours you do spend are spent on the judgment only you can provide.

The pattern holds whether the pressure is a slow-burn backlog, a daily fieldwork deadline, or a language gap that would otherwise default to you; which is exactly what tells you it is a pattern and not a one-off convenience.

And the tools underneath that pattern deserve the same scrutiny you would apply to any other professional investment: not “everyone else is paying for this,” but “what does this actually give back to me, in hours and in output, and does that number hold up.” If your notes are good, your thinking is sound, and your time is the constraint, that is exactly the gap the right AI assistant (and the right premium infrastructure behind it) are built to close.

Disclosure, Confidentiality & AI Use

I have no sponsorship or affiliation with Microsoft, ChatGPT or Anthropic; the above reflects my own paid subscriptions and personal experience. Project references in this piece are described only at the level already part of my public professional record; specific field findings are not disclosed ahead of my clients’ own reporting and publication processes. I also do not treat identifiable participant information as material to be casually uploaded into AI systems: confidentiality, consent, client requirements, institutional policies and appropriate de-identification remain part of the workflow whenever AI is used around research material.

A note on how this piece uses AI: the drafting and restructuring below was done with an AI assistant, from my own notes and edits, which is, appropriately, the whole subject of the article.