What gets lost when AI writes the query?
I think there's a version of the near future where nobody writes their own queries. AI drafts them, platforms refine them, and the analyst's job starts further downstream - closer to the data and further away from the decisions that created it. A lot of people are excited about that future, and I understand why, I’m excited too. Query writing can be painstaking, fiddly, time-consuming work. If a machine can do it faster, why wouldn't you let it?
I've spent fifteen years writing queries. I run a query writing competition that's now in its third year. And over the past twelvemonths I've invested serious time trying to teach AI to do what I do - building a query writing skill in Claude, feeding it proven examples, training materials, first principles. I didn't approach it as a sceptic. I approached it as someone who'd genuinely like to work faster and better.
So what I'm about to say comes from experience, not ideology: I think we're in danger of letting go of something we don't fully understand the value of. Not because AI can't help (it can, and it will), but because the conversation about automation is moving faster than our understanding of what the skill of writing is, actually, all about.
Here's what I mean. Most people think of query writing as a technical task. You define your terms, you build your Boolean logic, you run it, you refine. And at that level, AI is already useful. It can generate keyword lists, expand synonyms, handle syntax, produce a workable first pass.That's definitely valuable, and we don’t need to romanticise the manual version of all that work.
But the technical layer sits on top of something else. It relies profoundly on a foundation of judgement that's harder to see and harder to replicate. When you write a query, you're making a series of compounding analytical decisions. You're deciding where the boundaries of a conversation are.You're working out how people really talk about a category, which is rarely how the brand thinks they talk about it. You're building a taxonomy that needs to be genuinely exhaustive, not just tidy-looking. You're sensing when an exclusion is doing too much (e.g. cutting signal to reduce noise in a way that subtly distorts what you'll find).
Those decisions draw on experience, on your feel for language, and on your understanding of what the data needs to do once you’ve got it. These decisions aren’t procedural, rather they're about interpretation. And they build - each one shapes everything that follows, all the way through to the insight that lands on a client's desk.
When I built my AI query writing skill, this is exactly where it struggled. Not with the mechanics – that was often fine. But it couldn't maintain logical consistency across a complex build. It would lose the thread of a MECE taxonomy halfway through. It would skip steps, ignore constraints, and check its own work inconsistently. I'd set clear principles and find them discarded three moves later. These problems don’t feel like software bugs. They felt like something more fundamental, like a reflection of what the task actually demands, which is sustained, disciplined analytical thinking across dozens of interdependent decisions.
My honest belief is that AI will get better at this. Probably much better. But that's not quite the point. The point is what happens to us in the meantime, and whether we'll still have the skills to know the difference between a good query and one that merely looks good.
Because running the query writing competition for three years has shown me something I didn't expect. Skills drop off quickly. People who write brand monitoring queries every day can do that well. But the moment you step beyond that - into more complex category work, cultural conversations, multi-layered taxonomies – we see confidence outpacing ability. And, I’m sorry, but this gap often seems invisible to the person doing the work. They don't know what they're missing because they've never had to confront it.
This is the risk that worries me most. Not that AI will write bad queries - it'll write increasingly decent ones. But that we'll lose the ability to evaluate them. If you've never worked through the logic of a complex query yourself, you won't spot the weakness in one an AI builds for you. You'll accept tidy where you needed rigorous. You'll miss the exclusion that removed an entire strand of conversation you didn't know existed.
I see an irony here that's worth reflecting on. AI-assisted query writing might actually demand more expertise from the human, not less.When you build a query from scratch, the thinking is embedded in the doing, you discover problems as you go. When you're reviewing an AI-generated query, you need all of that knowledge but without the discovery process (and the skills you developed getting there). You're auditing an AI’s analytical decisions, and that requires a depth of understanding that only comes from having made those decisions yourself, many times, under pressure.
The best version of this future isn't one where AI replaces query writing. It's one where AI handles the procedural layers (e.g. the expansion, the formatting, the first go at logic) and a skilled human owns the judgement. This is delegation, not abdication. But delegation only works when the person delegating understands the task deeply enough to hold it accountable.
So my argument isn't that we should resist the tools. It's that we should invest in the discipline underneath them. Teach people to write queries properly, not because they'll always write them by hand, but because the practice builds something that can't be shortcut. This is the instinct for how language works in the wild, a feel for where the data is and isn't, a sense of what you're really looking for before you start looking.
Query writing was never just a task. It's a form of thinking. And the moment we treat it as something to be automated rather than something to be practiced, we don't just risk worse queries. We risk sloppier work, all the way through.