How I Actually Use AI in Consulting Work

The question I get asked most

Whenever I mention that I use AI heavily in my consulting work, the next question is almost always the same: does it actually work, or am I just saying that because everyone says that. It is a fair question. A lot of what gets published about AI and productivity reads like it was written by someone who has never had to defend a slide deck in front of a client at nine in the morning after three hours of sleep. So this is not a theory piece. This is what I actually do, tool by tool, task by task, including the parts where AI fails me and I go back to doing it by hand.

I work as a management consultant at a Japanese firm in Bangkok, moving between English, Thai and Japanese most days. That combination, more than anything else, is why AI became a load bearing part of my workflow rather than a novelty. Translation and synthesis used to eat hours. Now they eat minutes, most of the time.

Research synthesis, not research

The first place AI earns its keep is turning a pile of source material into something I can reason about. Before a client meeting I might have a stack of industry reports, a regulatory document and notes from three internal calls. I do not ask AI to tell me what to think about that material. I ask it to organise it: pull out the numbers, flag where two sources disagree, list the assumptions each report is making that are not stated outright. That last one matters more than people expect, because consulting work often comes down to noticing an unstated assumption before your client does.

What AI is bad at here is judgment about which disagreement actually matters. It will happily tell you that two sources report different growth rates without telling you that one of them is measuring a narrower segment, unless you specifically ask it to check. So I never skip reading the sources myself first. AI compresses my reading time. It does not replace it.

First drafts, never final drafts

I draft emails, slide narratives and short memos with AI constantly. The trick that took me a while to learn is that a first draft from AI is useful precisely because it is mediocre in a specific, predictable way: too generic, too hedged, too long. Once I can see that generic version in front of me, editing it down to something sharp is much faster than starting from a blank page. This is the opposite of how AI gets marketed. It is not magic. It behaves like a very fast, very patient junior colleague who writes a reasonable first pass and needs every sentence checked.

For anything client facing, I never send an AI draft without rewriting the opening and the actual ask myself. Those two parts carry the relationship, and a client can usually tell when they were not thought through by a person.

Data cleaning is where the real time comes back

This is the least glamorous use, and the one that saves me the most hours. Renaming inconsistent column headers, standardising date formats across exports from three different systems, writing a first pass of a script to reshape a dataset before it goes into Power BI: these used to be the kind of task that quietly ate an afternoon. Now I describe what the data looks like and what I need it to look like, get a script back, and spend my remaining attention checking the output against a few rows where I already know the right answer. I check against rows I know, never just rows that look reasonable, because a script can run without errors and still be wrong.

Translation between Thai, English and Japanese

This is the use case closest to my daily life, and the one I am most careful about. AI translation between these three languages is good enough now that I use it as a first pass for almost everything: an email from a Japanese colleague, a Thai regulatory clause, a client update that needs to go out in two languages. What it is not good at is register. Japanese business writing carries a level of formality and indirectness that a literal translation misses, and a message that sounds perfectly fine in English can land as blunt, even rude, once translated straight into Japanese. So the AI draft gives me the content. Getting the tone right, especially for anything going to a Japanese client, is still something I do by hand, sentence by sentence.

Where I still do it myself

There are three things I do not delegate. Any specific number that will appear in a client deliverable, I calculate or verify myself, because an AI tool will produce a plausible number with total confidence whether or not it is correct. Anything that requires reading the room, deciding how to phrase feedback to a client who is already frustrated, I write myself. And the final structure of an argument, the actual logic of why this recommendation and not that one, has to come from me, because that is the part of the job someone is actually paying for.

What this adds up to

None of this makes me faster at being a consultant. It makes me faster at the parts of the job that are not really consulting: formatting, first drafts, translation, cleanup. That time goes back into the parts that are: thinking about what the client actually needs, checking the numbers, and writing the one paragraph that has to be exactly right. That is the whole pitch, and it is smaller than the hype suggests. In my experience, it is also true.