For the past few years, we have been bombarded with news items telling us that artificial intelligence (AI) is transforming the workplace, potentially making some roles redundant, but at the same time bringing greater abilities to those who stay in work. It’s only now, though, that we are starting to see an approach that fully employs AI’s abilities, and it’s all about context.
Much of the early wonder inspired by AI apps was generated by their ability to produce images and videos in response to a prompt. And this certainly made it possible for even those of us with very limited graphical abilities to generate passable imagery. But most business use of IT isn’t about pictures – it involves words and numbers.
In these areas, AI has still delivered, but primarily in the form of sophisticated web searches.
Where previously we might turn to a search engine with a question, now we are more likely to ask an AI to get a specific answer, rather than just a set of links. Even when using Google, say, it’s increasingly the AI response at the top that is taken as the result. Similarly, when writing a report, many will now at least start with some AI generated content, rather than sitting down and writing from scratch. (My writing, however, remains 100% human.)
There’s nothing wrong with this, as long as we’re aware of AI’s limitations, both in its uncontrolled sources and its tendency to make up responses if it can’t find a suitable answer. A powerful reminder of this was the example of freelance writer Amanda Guinzburg, who asked ChatGPT to summarise her most successful articles to put in a pitch. She was surprised to see that the summaries bore no resemblance to the articles, even though, when questioned, ChatGPT responded ‘I am actually reading them – every word.’ Eventually ChatGPT admitted to making things up, because it couldn’t get to the original text on Substack. If Guinzburg had access to contextual AI, things might have been different.
At the moment, much of what we see written about contextual AI involves, for example, augmented reality glasses and wearables which will monitor what you see and hear to be able to deliver information in context. But the greater possibility for business is likely to be contextual AI that can access your information bunkers.
Microsoft’s Copilot has begun to take a contextual approach by tying in to work produced using Microsoft 365.
Whether it’s a Word document, an Excel spreadsheet, or your emails, when you ask for input from Copilot it can provide assistance that goes beyond generic material from internet sites to work effectively with your own source material. It’s claimed this will reduce the chances of AI hallucination, though even contextual AI can make mistakes.
Similarly, I’ve started using an AI package called Littlebird which has not yet reached version 1, but shows how contextual AI has huge promise for business applications. It monitors the user’s system, collecting information that would otherwise remain in individual silos. So, for example, I use emails and WhatsApp for communication, keep a large amount of information in OneNote, manage tasks in Things and do a lot of work in Word and Excel. If I had tried the same thing as Guinzburg using Littlebird, it would have been able to provide accurate summaries, because it could access the original source documents as I produced them.
A simple example of contextual AI
As a simple example of how contextual AI can help, I recently wanted to email some potential authors for a new book where I have the role of series editor. My starting point was to ask Littlebird how I had previously summarised the content of the book. (I admit, I had forgotten.) I asked ‘How did I describe the possible Military Science book in an email?’ and was given the summary I’d provided, along with other key points in the email. I then used this summary to get the app to search for potential authors, draw up a list with emails and suggest the content of personalised emails, including information from my notes on the series in OneNote and a document giving background to the series.
Ideally, I would then have been able to tweak the emails and get Littlebird to send them, but as yet it can’t send emails on my behalf. Even so, it was a useful aid, because I didn’t have to deal with each of the separate information silos, but rather could use the software to pull together selections from my overall information base, dependent on context.
It’s early days, and like all AI, Littlebird is still capable of hallucinating. I earlier asked it to summarise a table from a PDF I had stored in OneNote. It successfully accessed the table, but added a column that didn’t exist and made up a significant part of the content that wasn’t directly connected to me. Using contextual AI should not be about taking the human out of the loop. But it can make humans a lot more efficient, and offers far more promise than apps that are limited to data that is accessible on the internet.