AI and the Letterhead
I keep seeing posts about how AI can build products, replace teams, and automate whole departments. So I used it for something far less impressive: we needed a few documents put onto our letterhead and signed. Nothing strategic. No clever workflow. Just standard admin that needs to look right and go out the door.
I uploaded the letterhead into ChatGPT, pasted in the body text, and assumed (hoped) it would take seconds. It didn’t. The formatting broke, the font changed, and the logo disappeared entirely. The only thing that came across properly was the address.
It was frustrating, not because I expected perfection, but because of the gap between what we’re told AI can do and what just happened on my screen. If it can reason through code and analyse contracts, how does it struggle with what feels like structured copy and paste?
When I stepped back and looked at it properly, the explanation made sense. The “template” wasn't really a template. It was just a Word document that looked right to us. There was no real structure behind it: no defined header logic, no embedded layout rules. It only worked because humans already knew what it was supposed to look like. To AI, it was just styling.
That’s the bit that gets glossed over. AI looks impressive when the environment is clean and properly structured, but most businesses aren’t built like that. Ours included. We operate on documents that have been amended over time. We use templates that evolved rather than being deliberately designed, and processes that make sense simply because we’re used to them, not because they’re logically tight.
AI just reflects that back at you. It didn’t “fail” in some dramatic way; it just highlighted that what felt simple wasn’t actually simple. It was familiar, and I was filling in the gaps without realising.
There’s a constant narrative about AI replacing people and hollowing out teams. What feels more accurate is that AI exposes how disciplined your operations really are. If the foundations are loose, the output will be loose. If the inputs are vague, the results will drift.
That’s not as exciting. But it’s probably more accurate.