Generating a formatted Knowledge Article from a Word document using Now Assist

[SwarnadeepNandy]

Hi All,

I'm looking for guidance and real-world experience from anyone who has tackled something similar with Now Assist / Document Intelligence.

What we're trying to do
We have HR policy documents authored in Word (.docx) and we want to turn them into properly formatted Knowledge articles in ServiceNow. The goal is to take the source document as an input and produce a clean, structured KB article as the output.

The catch
Because these are policy documents, the wording has to stay exactly the same after the transformation. We're not looking for the AI to summarize, rephrase, or "improve" the text — the legal/policy content must be preserved verbatim. What we want the AI to help with is the formatting and structuring (headings, sections, layout into the KB template), not rewriting the substance.
So it's a bit different from the typical "summarize this document" or "answer a question about this document" use case. It's closer to: reproduce this content faithfully, but lay it out as a structured Knowledge article.

Limitations we've run into
When using the Document Intelligence tool, the actions available seem to be limited to three:

1. Ask a Question (Q&A)
2. Extract information
3. Summarize files.

None of these is quite a clean fit for "take the whole document and re-render it verbatim into a formatted article" — Q&A and summarization are lossy by design, and extraction is geared toward pulling specific fields rather than reproducing full content.

So I'm wondering whether we're using the right tool at all, or whether Document Intelligence is simply intended for a different set of use cases.

Is Document Intelligence genuinely limited to just those three actions, or is there a way to extend it / configure additional behavior that I'm missing?

A few specific things I'd love input on

• Has anyone used Now Assist (the Document Intelligence skill / skill kit, or an AI Agent) to generate Knowledge articles from attached documents? How did it work out?
• Is there a recommended pattern for "preserve text verbatim, only restructure formatting"? Generative models tend to paraphrase by default, so I'm interested in how others have constrained that.
• We're also dealing with fairly large documents (~3-5mb) (full policy handbooks, many pages). Any advice on handling large attachments through Now Assist — chunking strategies, size limits you've hit, or whether a different approach works better at that scale?
• Given the three-action limitation above, would you approach this as a Document Intelligence use case, an AI Agent, a custom skill (Skill Kit), or something else entirely? Curious what's worked for people in practice.
• Apart from Document Intelligence is there any other way to handle attachment for GenAI using custom skills or AI Agents? or even for that matter scripts or flows?

Any pointers, lessons learned, or "here's what we did instead" stories would be really appreciated.

This seems quite straight-forward in Copilot/Gemini/Claude, but with Now Assist, it seems quite difficult.

Thanks!

[Mark Manders]

Any reason why you aren't just importing the word document the old fashioned way? No AI required, so no cost, no changing of content, no nothing.

[SwarnadeepNandy]

Hi @Mark Manders ,

Thanks for you answer.

We are currently doing that, but in the era of AI, we want to apply a particular layer of styling on the Knowledge Article content.
For eg: We are applying a unified styling for all the Knowledge Articles for HR, with a particular header & sub-header styling and collapsible sections.

So, what we were trying is upload the document to Now Assist/LLM and as it to style according to some guidelines.

There are ofcourse manual way of doing it, but we are tying to leverage Gen AI to do this styling activity.

Kind Regards,
Swarnadeep Nandy

[SwarnadeepNandy]

And essentially I am trying to understand how easy/difficult it is to handle attachments in ServiceNow from Gen AI perspective.
Which at this moment, feels quite challenging. Hopefully I am missing something or it would be improved later on.