Smart for Work
Building an AI Team That Works With You · How Far Can AI Take a PowerPoint?
This edition of Field Notes has two parts. The first is about forming a small team of AI assistants to work with you: when it's worth doing, and the principles that keep it from getting messy. The second is a practical one from my day job: what AI can and can't do for a real presentation deck.
When one assistant stops scaling
Most people start with one general AI assistant, and that's the right place to start. Everything goes into one chat, and for a while it works well.
Then it stops scaling. The context gets crowded, instructions from different jobs start to blur together, and one helper ends up holding every tool and every risk at once. That's usually the sign to form a small team. But a team isn't just more assistants. It's a decision about which roles are actually worth splitting out.
I'm not going to hand out a template or describe a setup to copy. Everyone's split will differ with the kind of work they do, the tools they already have, and how much control they want to keep. What travels well are the principles.
Which roles are worth splitting out
Three lenses help:
- By domain: work that needs different knowledge or context.
- By risk: work where a mistake costs more, so it deserves tighter rules.
- By tool access: work that needs powers the others shouldn't have.
Tool access matters more than people expect. A role that only reads and drafts shouldn't hold the same keys as one that can send, pay or delete.
Give every role a brief
Each role needs a clear brief: what it's for, its scope (what's in and what's out), and, just as clearly, what it must never do.
That last line is the one people skip. Never publish. Never contact anyone. Never spend money. Writing it down is cheap. Forgetting it isn't.
What a good hand-off contains
Things tend to go wrong between tasks more often than inside them. A good hand-off carries:
- the goal;
- the deadline;
- what "done" looks like;
- who to report to;
- and a rule to report back even on failure.
Silence is not a status. A small example of why: a hand-off gets interrupted halfway, nobody reports it, and it only turns up the next morning. Nothing breaks, but it shows that nobody had been assigned to check.
Someone owns the check
So someone, or something, owns checking the results: not just "did it finish", but "is it right". And for anything with a real deadline, the human verifies it too. A quick look before the deadline beats a surprise the morning after.
Two shapes, and when to switch
Talk to each assistant directly. Clear lanes and full control. The cost is that you become the router, carrying context from one assistant to the next and remembering what's still pending.
One coordinator that delegates and integrates. One conversation, one combined answer, much less switching. The cost is less obvious: the coordinator has to track everything that goes out, and anything that quietly doesn't come back.
Direct works well while the team is small. Once you're spending more time passing messages than deciding, a coordinator starts to pay off.
The rules that don't move
Whichever shape you pick, the human stays the decision-maker for anything public, anything involving money, and anything that can't be undone. The team prepares; the human makes the call.
And start small. Two or three roles is plenty. Add another only when a real, repeated pain shows up, not because a new role sounds clever.
If I had to put it in one line: a good AI team isn't more assistants. It's clear briefs, hand-offs that report back, and a human who still makes the big calls.
How far can AI take a PowerPoint?
Now the day-job question. Before AI, one proper deck meant understanding the concept first, doing the research behind it, then drawing the flow charts, the visuals and the layout. For me that was half a day to a day, and that's at best.
The "ideal" AI flow looks simple. The company I work at now uses Gemini, so that's where it starts: Gemini gathers the research and writes a first draft, then NotebookLM turns that draft into slides. It's fast, and as a first pass it's genuinely good.
But there's a limit. It condenses. Put in twenty pages and you get back ten-something. Pages merge, sections shrink, and detail gets folded into bullet points. That's sensible, honestly; a shorter deck is easier to follow. But real decks have fixed pages. A company introduction, project pages and reference pages each need their own space. They can't be squeezed into one slide.
What I do now
My current approach has a clear order, content first and layout second:
- Agree the content with an AI assistant first: what goes on each page, and in what order.
- Give it a template, and let Genspark build the deck from that template.
- Tidy the details by hand: spacing, wording, the small things a reader notices.
Why bother with a template rather than letting the tool design everything? Because the template keeps the original layout and colours, so the deck still looks like the company's own. It also makes rebranding fast. The company has two brands, and different clients and vendors need different looks. Same content, the other brand, on short notice. Or switch the language of the whole deck for a different audience. No rebuilding from scratch, just a new look on the same story.
Where AI stops, for now
So how far can AI take a PowerPoint? For content that needs no proofreading, such as marketing research, AI can take it end to end. For formal decks that go outside the company, it still needs human work. AI gets you most of the way. The last stretch, the one a client actually reads, is still mine.
If you've found a different way to handle either of these, I'd like to hear it. More notes are at kpoleung.com, the podcast is on Spotify (https://open.spotify.com/episode/2fymjI4t3YIMIoZOqCRr0g), and comments or topic requests are welcome at m@kpoleung.com.