Notes from the middle of something: finding the trAIl

At the start of this year I gave myself a loose brief: use AI wherever I plausibly could, and pay attention. Not to prove a point — I didn't have one to prove — just to understand it from the inside rather than the outside. The word was everywhere: headlines, meetings, job descriptions, casual conversation. I wanted to know what it actually felt like to use it, rather than just hear about it.

Here's where I've been using it so far.

Writing coach

This is probably where AI is most visibly useful — and where I'm most watchful. I use it to pressure-test tone, trim structure, and sense-check length. The benefit is real: I come away more confident in what I'm putting out. But confidence and quality aren't the same thing, and there's a version of this where I gradually outsource the friction that actually makes writing personal. I haven't resolved that tension. I'm just aware of it.

The other thing worth noting: when I've asked it to support arguments with references, it has occasionally invented them. Plausible-sounding, specific, wrong. That's not a minor footnote — it's a reason to treat it as a thinking partner rather than a source.

Internal search

An Enterprise LLM is great for search. It surfaces answers faster than conventional search, and finds data I didn't know existed. That's a genuine time saving. What's disconcerting is that I can't get comfortable with what I am not seeing. With a traditional search, I have some sense of the edges — what I searched, what came back, what I might have missed. With AI-assisted search that perimeter is less clear. More answers, but less visibility of the territory.

Excel

I've used it to structure new workbooks, strengthen reporting data and improve formulas. Because you can describe what you want in conversational text, it makes it easier to find the right formula and significantly reduces the time spent on trial and error. The analysis it enables is more consistent and less dependent on whether I happen to remember the right syntax.

Interestingly, when I've tried to have it work directly in a sheet the results have been inconsistent and frustrating at times — including describing capabilities that don't exist.

Weekly reporting

For the last three weeks, I have used an LLM to summarise correspondence and generate a working draft of a weekly report. As a starting point it's solid — the shape of a report, roughly the right content, something to react to rather than start from scratch.

The pattern I've noticed is that it weights frequency over consequence. Threads with high message volume get more space; quieter exchanges that contain something important can get marginalised. A report structured around what was loudest isn't the same as one structured around what mattered. That's still an editorial judgement I have to make myself.

Process design

I've used two different LLMs to assist with process design — one to find a digital solution to a problem I was sure didn't need to be paper-based but didn't know how to solve, another to flesh out a data collection and assessment process that was still finding its shape. In both cases the LLM generated potential solutions and, in the case of the digital solution, coached me through implementation. I came away feeling I'd genuinely expanded my capabilities and my value in that area.

The uncertainty I sit with is compliance. A solution can be workable and coherent and still fall foul of a regulation I didn't know to check. AI doesn't know what it doesn't know about my specific context — which means the output here requires a different kind of scrutiny than elsewhere. Useful, but not sufficient on its own.

Summarising documents and regulations

Quick synthesis is probably AI's most straightforward win. A long document turned into a readable summary, with areas flagged for closer reading, is genuinely useful — especially with dense regulatory material.

The thing to watch is conflation. When I've asked it to synthesise across multiple documents it has occasionally blended them in ways that are subtle enough to miss on a quick read. The summary reads coherently; the problem is it's pulling from slightly different contexts. Not a reason to stop using it for this, but a reason to stay close to the source material.

Covering letters

I've used it to compare a CV against a job description and identify what to emphasise. In principle this should have saved time. In practice, the output usually requires checking and rewriting — to confirm accuracy and make sure the letter still sounds like me. Next time I'll be using a different prompt: to identify strengths and gaps only.

What I've actually learned

Take these for what they are — things I've picked up along the way. Use what's useful, ignore what isn't.

Ask if it's sure. It will often revise, qualify, or row back when pushed. Sometimes this surfaces areas of inference rather than fact, other times it acts as a useful peer review.

Demand references with page numbers. Then check every single one. Not most. Every one. The invented reference problem is real and it looks convincing.

Tell it to flag assumptions and inference. The difference between what it knows, what it's inferred, and what it's assumed is not always obvious in the output. Ask it to make that visible.

Slow it down. The default mode is to answer quickly. Quick isn't always what you need. I prompt it to think before responding — to consider the question rather than just react to it. The output is usually better.

Set it up as a collaborator, not an expert. This one has made the most difference. When it assumes the role of expert it produces confident answers. When I frame it as a thinking partner working through a problem with me, it asks better questions and hedges more honestly. That's the more useful version.

Where this leaves me

Still in the middle of it. Some of these use cases will probably drop away; others will develop into something more reliable. What I didn't expect was how much using it regularly would sharpen my awareness of what I actually need — from a tool, from a process, from my own judgement.

That might be the most useful thing it's given me so far.