A brutally honest account of using AI to build a real corporate presentation — not a demo, not a toy project.
I just finished a week-long project building a half-year industry analysis presentation for a large company with six subsidiaries across multiple business verticals. Thirty-plus slides, six companies, three years of operational data to cross-reference, and a deadline that didn’t move.
I used Claude — specifically a combination of Claude’s chat interface, its desktop Cowork app, and its Office plugins for Word and Excel — as my primary co-worker throughout. Not as a toy. Not as a “let’s see what AI can do” experiment. As an actual production tool on a real deliverable.
Here’s what I learned. The honest version.
The Setup
The brief: a 2026 H1 analysis report for a large company with six subsidiary businesses across multiple verticals. Leadership wanted forward-looking projections where possible. The data lived across dozens of operational reports — Word documents, each one extremely long — plus spreadsheets. Six companies. Three reporting periods each. That’s 18+ documents just for the core data layer, before touching anything about structure or narrative.
I had no template. I had no clear starting point. I had a pile of source material and a deadline.
This is where AI actually earns its keep — or fails you.
What I Did Wrong First
Let me start with the mistakes, because they’ll save you more time than the wins will.
Mistake #1: I tried to build on top of an existing PPT file.
My first instinct was sensible: don’t start from zero, grab a previous version from a colleague and iterate from there. Hand it to Claude, ask it to update slide by slide.
This is a trap.
Every time Claude touches a task involving an existing PowerPoint file, it re-reads the entire thing from scratch. It tries to understand the whole structure, all the existing slides, the formatting logic — before doing anything you actually asked for. I watched one session burn through 60% of its context window just trying to insert a single new slide. The deck barely moved. My progress stalled completely.
The fix: stop trying to edit an existing file. Generate each slide as a standalone output and paste it in yourself. More on this below.
Mistake #2: I gave Claude too much context, thinking more was always better.
It isn’t. Dumping all 18 source documents into one session and asking Claude to “just figure it out” produces confused, generic output. The model gets pulled in too many directions. What I started calling “context contamination” is real — irrelevant material in the context window quietly degrades the quality of the output you actually wanted.
The fix: treat each slide as an isolated task. Only load the files relevant to that specific slide.
The Workflow That Actually Worked
After the wrong turns, I landed on a four-phase system.
Phase 1 — Plan in chat, not in Cowork.
Before touching the desktop app or generating a single slide, I spent time in a Claude chat Project just reading and thinking. I uploaded source documents, asked Claude to help me understand what data I actually had, and together we worked out a chapter structure and a rough outline of what each section needed to contain.
The critical thing here: you have to do this yourself. Claude can help you brainstorm, but you have to be the one who actually understands your material. If you can’t judge whether a proposed structure makes sense, you can’t course-correct when Claude gets it wrong — and it will get things wrong. Your judgment is not optional.
Phase 2 — Extract data using the Office plugins.
This was the biggest surprise of the project: Claude inside Word and Excel is genuinely powerful for data work, in a way that the standalone chat interface isn’t.
With Claude in Word, I opened three operational reports simultaneously and asked it to cross-reference them — pull the same hotel’s revenue figures across Q1 2025, H1 2025, and Q1 2026 in one pass. It did. That would have taken me an hour manually. It took maybe three minutes.
With Claude in Excel, I could ask open-ended analytical questions directly against the spreadsheet data. Not “apply this formula” — actual questions like “which segment showed the steepest decline in the back half of the year and what’s driving it?” Claude would locate the relevant data, analyze it, and discuss the findings with me interactively. This is not a smart autocomplete. It’s a different kind of tool.
Phase 3 — Build slides one at a time in Cowork, with strict isolation.
Here’s the discipline that made Cowork actually work: one slide (or one cluster of closely related slides) per conversation. One task. Clean context.
My process for each slide:
- Copy the relevant source files into a fresh working folder
- Open a new Cowork conversation
- Tell Claude exactly which files to read and what I need
- Ask for a written work plan and outline before any slide generation
- Iterate on the content through conversation until it’s right
- Only at the very end, ask for the actual PowerPoint output
That last point is important. Generating the PPT is the last small step, not the first. Most of the real work — the thinking, the structuring, the data decisions — happens in conversation. If you jump straight to “make me a slide,” you burn tokens on something you haven’t thought through yet, and the output reflects that.
Phase 4 — You manage the master file. Not Claude.
The master PowerPoint file lives in a completely separate folder that Cowork cannot see. Claude generates individual slides; I copy and paste them in manually.
This sounds tedious. It is slightly tedious. It is also the reason I never had to spend an afternoon rolling back a corrupted file. When Claude can’t see your master file, it can’t accidentally overwrite, reorder, or misinterpret it. The manual paste is cheap insurance.
The Tools, Mapped to Their Jobs
| Tool | What it’s actually good for |
| Claude chat (Project) | Planning, brainstorming, making sense of source material |
| Claude in Word | Cross-referencing multiple documents you already have open |
| Claude in Excel | Open-ended analysis and Q&A against live spreadsheet data |
| Claude Cowork (desktop) | Executing specific, isolated slide-generation tasks |
| Claude in PowerPoint | Mostly skipped — Cowork handled this better in practice |
The Honest Summary
AI didn’t make this presentation for me. It made it possible for me to make this presentation in a week instead of three.
The parts that still required me: understanding the source material, judging whether a structure was right, deciding what story the data was telling, and managing the final output. None of that got automated away. What got automated was the tedious cross-referencing, the reformatting, the “turn this table into a slide” mechanical work that used to eat hours.
The traps are real. Handing Claude a messy context and hoping for magic doesn’t work. Treating it like a simple command-line tool doesn’t work. What works is treating it like a smart but context-blind collaborator — one who needs clear instructions, bounded tasks, and a human in the loop who actually knows what good looks like.
If you’re building complex decks and you haven’t tried this yet, try it. Just start with the isolation discipline. One slide, one conversation. Don’t let it touch your master file. Do the planning in chat first.
It won’t feel like magic. It’ll feel like having a very fast, very capable assistant who needs good management. That’s exactly what it is.