I’ve used Microsoft Word for decades and run ENVEE Digital/WebSolutionz for many years, so a lot of our documentation has been developed in Word. I wouldn’t say I’m a power user, but we do have ENVEE Digital branded templates which make use of Word’s fields, styles and building blocks to help streamline the process and standardise output as much as possible.
Producing good documentation and project plans for proposals, projects and digital products can take hours, even with good templates, and often this is non-billable time. This puts an upper limit on how much work we can reasonably handle concurrently.
Of course there are always paid solutions to manage this, but as a small business owner I’m reluctant to take on too many subscriptions. The obvious move in 2026 is to use AI more, so I started experimenting with Claude. The first few attempts were impressive – Claude can use my templates and help me write much faster than I can do on my own. Great!
The problem – is that what it generates breaks all of Word’s formatting and comes out looking very “Claude-ish”. If you don’t mind sending out documents that look the same as everyone else’s, that’s fine, but I’m particularly pedantic about presentation and branding. And no matter what I tried, I just couldn’t convince Claude to not break our branding.
The reason – without getting too technical, is that Claude works at the XML layer of a Word document, and Microsoft’s version of XML is frankly a bit ugly. This means Claude sees the value of a style or a field and rewrites that value in the XML, but Word doesn’t work exactly that way.
The end result – is that Claude replaces Word fields and Word styles with plain text and some basic formatting, and sometimes doesn’t recognise hyperlinks or other embedded Word features at all. It looks OK when it’s produced, but if you add pages or move things around there’s no automatic re-numbering, and if you want to change a Heading style you need to go back and manually change every single heading, and sometimes links get formatted red with an underline but there’s no actual link underneath. It’s a classic AI generation that looks OK at first glance but is not particularly reusable going forward. I’ve already spotted documents that have been distributed with this exact issue and the senders haven’t realised, because that first glance always looks fine.
I tried a bunch of different ways to force Claude to honour Word’s inbuilt formatting, and as soon as I forgot to provide the ever-growing list of “do not” instructions it would break something in the document again.
Eventually I accepted that while AI is cool for many things, it can also be pretty dumb, and this was one of the dumb things I couldn’t fight. (Actually I think it’s probably more about Microsoft’s ugly XML, or some kind of inter-tech-bitchiness, but it doesn’t matter).
So I changed direction. Now I’ve broken the process down into two parts:
- Claude helps me to write, and generates my content in plain markdown with specific content markers embedded, which it does very well.
- I run that markdown file through a small Python script I built (with Claude’s help) that runs locally on our server, which inserts the content into a nominated, branded Word template document.
The first step takes as long as it takes to get the content into the shape you want. The second step is instantaneous. The process maintains all of the Word template fields and styles, because Claude never gets its dirty hands on Word directly. And they look great!
I won’t lie, this took days to work through and get set up properly, and I got it very, very wrong several times – until I got it right! 🥳
The good news is – the team can now generate a fully customised document, based on the client consultation transcript and using our branded Word templates, in minutes, at close to final quality. Someone still needs to read and finish off every document, and insert some blocks, and then manually export it to a PDF. Even better – we can now also produce digital products super-fast using the same process, in a similar amount of time.
What did we learn?
I haven’t fully recouped the time I spent on R&D – yet. But the process is so scalable now that it’s going to pay for itself in a pretty short time, and then we’ll be gaining hours per project or product after that.
Is there a direct ROI? Not really – it’s a time-based saving.
The organisations that get genuine value out of AI are those that spend some time working out what it can do, and more importantly what it can’t do, BEFORE throwing a ton of cash at it and hoping for magical (and often non-existent) ROI.
