Back in 2018, which for AI is prehistory, we wrote about the relationship between AI, specifically deep learning AI, and document automation. This went on to become the most popular piece of content on this blog. 6 years on, it still doesn’t feel outdated despite the leaps in capability that AI has made since then.
To put it into perspective, back in 2018, OpenAI, the company behind the popular ChatGPT, was still a non-profit and ChatGPT’s breakthrough moment (late 2022) was more than 4 years away.
Since then, the AI became dramatically more capable, and yet, non-AI software is still significantly better in many use cases.
“Non-AI” computers and software have been good at things that humans are not good at for a long time - think about a calculator performing maths operations in split second, or a regular look up or search for some data. The “Non-AI” computer’s speed and accuracy is incomparably better than humans’. The tasks that “Non-AI” computers are better at are the ones that can be defined with exact rules and logic - may it be multiplication in case of the calculator or finding an exact piece of text in a document.
AI is making computers good at tasks that humans are good at, and traditional computers and software struggle with - creating prose or poems, creating images, recognising objects in images, creativity, following loosely defined instructions etc. Effectively, the tasks that don’t have exact rules and logic.
Conversely, AI is not that good at tasks that have hard rules. You can’t rely on its mathematics, you can’t trust it to reliably find all instances of text in a document, it can make up answers that seem realistic but are completely made up - so called hallucination.
One of the major challenges in the future development of AI is the ability to marry the AI’s human-like capabilities with the ability to follow hard rules and recognise and use hard facts.
Document Automation falls into the category of software that is excellent at doing what humans aren’t that good at - following the defined hard rules without fail, creating documents that are 100% accurate, 100% of the time. It’s very fast, it doesn’t get tired, it doesn’t have bad days, it doesn’t take days off.
Quite often, when you automate documents, you can achieve 100% automation. This means that no human needs to provide any input to generate the document. This type of documents forms more than 50% of documents generated by ActiveDocs. There are instances where human input into document automation is required. This can be either in the form of writing narrative text, perhaps explaining why a specific set of products was recommended, or a review of content or data that can’t follow strict, hard-defined, rules.
And this is where AI comes into play.
If your documents have content that currently requires humans to write it, why not explore using the Large Language Models (LLMs) for writing it automatically.
If your documents require human reviews or validation, why not use AI for it too?
In terms of narrative content writing, in our testing, we could get quite far with the “out of the box” LLMs (zero shot learning). Very likely your organisation already has been creating this content for a while and can generate a number of examples to “teach”, fine-tune, your AI model and generate content that follows the company language, tone, and nomenclature.
Similarly, with the review and validation, if your reviewers previously approved or rejected documents, you can use that as the training data for your LLM and teach it how to review for you.
Noting that, in terms of document review, if you can condense your review process into hard-defined rules and logic, which could be a Business Analysis exercise, then it is possible to build them into your document automation software, and you won’t need the AI. However, AI review can still pick up on issues that you may not have included in your rule-set.
If your document creation is 100% automated, congratulations, you already have the best, fit for purpose, kind of AI.
If your document creation, even though it’s automated, does require human input, consider using AI to remove the need for it. It may not replace all of the human input right away, but it will get you started on the path of full automation.
Using AI as a “co-pilot” for individuals is appealing, and it genuinely can help with some tasks. However, using AI as a tool that is precisely targeted on the remaining human-resource-intensive tasks in your document processes preserves what already works well, and transforms the way your organization works at scale.
You can watch a companion video to this article here:
Martin Srubar
Senior Technology Evangelist
Martin’s engineering background and his passion for great products whether in physical or software form are complemented by his understanding of ActiveDocs applications and how they meet the requirements and fit the architecture of the company’s clients. Martin continues to engage with potential and existing customers, adding market intelligence and customer feedback into the company’s ongoing product development strategy.
