I've been using AI as a way to organise and coordinate some (but not all) of my work since about February - I work across a commercial platform, my own consultancy, a research programme, and everything at home and in my personal life. For most of that time I'd have told you it was going well, I'd have meant it, and it is! I am getting more done and my focus is more on the actual work rather than wondering where version 8 of a powerpoint should be stored.

However, much of my work lately has been putting together larger strategies. What I've realised is that I'd been working almost entirely in discrete packets of work; Write this email. Check these records. Draft this section. Fix this thing. Every one of those is self-contained - everything it needs is in the request, you can see immediately whether it worked, and when it's done, it's done. AI is genuinely excellent at that, and if that's most of your work then you may not hit the problem I hit at all.

The limitations of that started when I tried to pull a strategy together.

Why the packets hid it

A packet of work (I'm defining it as a set of tasks smaller than a project, like "draft a document on topic x and send it to y for review") doesn't need continuity. It doesn't need to know what I decided in June, or what I said I'd do and then didn't, or how this piece connects to the other four. You hand over enough context in the request itself, and the fact that nothing carries over between sessions may never come up.

A plan or strategy is the opposite: as it exists across a longer timeframe, it needs last month's decision to be available for context, it needs to know what's stalled and what's moved, and it needs to still be there when I come back to it a fortnight later. So when I actually went to make these strategies and plans, I realised I was managing a lot of that in my head, with pieces of it across docs, files and emails.

To address this, I'd written good context files for each of my companies - who decides what, where things live and the files/emails that shape the decision being made. When I checked and found that in six months many sessions had never loaded them, which I assumed was down to context windows or a limitation of the tool. For packet and task-based work, it made little difference (as I had the context in my head), and there was enough context from whatever I was doing (editing a file etc.) which is meant it didn't matter as much. For anything that needed continuity that I didn't want to keep in my head, it was much more of a problem. It did appear subtly in areas that in hindsight, make more sense.

I took the famous John Wanamaker quote of "Half the money I spend on advertising is wasted; the trouble is I don't know which half" and modified it match my AI experience:

"Half of the work I do on setting up context is wasted; the trouble is I don't know which half"

As an example, a weekly planning job stopped running months ago and while I noticed, it never felt particularly useful as the context was always limited (due to the issue above). None of it hurt while I was working task to task. It was glaringly obvious the moment I tried to plan across projects and companies.

The one job that did survive was a daily review I read every morning. It's had exactly the same context problem as everything else, but because I actually read it, it gave me the thoughts on what tasks were missing as much as what I needed to do.

What I got backwards

I'd been optimising for the individual conversation or short(er) project. Getting a good answer out of a session felt like the win, and it usually was, so I kept doing that and assumed the bigger picture would take care of itself. Spoiler: it doesn't with these tools.

Every session starts from nothing unless something deliberately carries state forward. What I actually had was a large pile of finished packets and no accumulated position - which feels productive right up until someone asks where the plan is - which was in my head more than on paper/disk.

The other thing that makes this hard is that these tools fail politely. They don't crash, they carry on with whatever thin context they've got and hand you something that looks fine. They may not refuse either (I regularly asked did they have enough context with the answer 100% in the affirmative), they just skip the step and say nothing. So there's no moment where you're told it's not working. You find out later, when you try to zoom out for the bigger plan and find a pile rather than an organised system.

What I've changed

The context that doesn't change - the companies, who decides what, where things live, the rules I don't want to explain twice - now loads automatically at the start of every session, rather than sitting in a file I thought I'd pointed at. That one change is the difference between explaining my own project structure for the twentieth time and never explaining it again (so far).

A scheduled job reads the day's transcripts each evening and writes down what actually moved, because I know for a fact I won't, or I'll forget. I said as much when I was asked, and I'd rather design around that than pretend otherwise. In the morning I get it back as two short lists: what moved yesterday, what to do next. I don't write either of them.

And everything that runs on a schedule now records that it ran, even when there was nothing to do. Otherwise a quiet week and a dead job look identical, and I've now learned the hard way how long you can go without telling those apart.

What this doesn't mean

Working in packets (bundles of tasks) isn't wrong. Lots of work can be done in packets, and AI is very good at them. But packets don't add up to a plan or strategy on their own, and I'd used my brain (and other docs) to bridge the gap.

None of this means that I'm now outsourcing my thoughts or about handing it all over and forgetting about it. What I've pushed onto the system is the continuity - what happened, what's outstanding, what I decided back in June - which is precisely the part humans are bad at holding. It doesn't decide what matters, what to drop, or whether the plan is any good. That's still mine, and I don't particularly want it to be anything else. I'm working on how to visualise that in a way that works better for me than streams of text.

If anything, I think work now needs more attention rather than less. A tool that fails politely will hand you something plausible and then wait for you to nod at it, and the better the output looks, the easier that is to do. The point of getting the continuity out of my head isn't that there's less thinking to do, it means that I've got more attention and brain power left over for the part that actually needs me.

If you're finding these tools useful for tasks and vaguely unsatisfying for anything bigger, I don't think that's you doing it wrong. It's that the thing which makes a plan a plan - continuity - is the one thing you have to build deliberately, because it won't spontaneously arise out of your threads.