In AI, one term often used is ‘grounded.’ This means that an AI tool or function has been set to reference only a specific body of data that you have provided.

“That material can include anything from Word documents, PDFs and spreadsheets to webpages, Teams meeting notes and emails,” says Bruce Penson, Managing Director at Pro Drive IT. “In Copilot,” he explains, “a notebook is a special workspace that you’ve grounded in a way that equips it to help you with a particular task.”

Digesting data

According to Penson, a notebook and standard Copilot prompt are generated by the same AI large language model (LLM). But whenever you submit prompts to a notebook or ask questions, that LLM is focused only on the discrete set of information input at the setup stage. That gives you authority for analysing that data, or using it as a launchpad for meeting strategic, creative or project-based goals.

“You can prompt a notebook in the same way you’d prompt a more general Copilot chat,” Penson says, “and you can use it for drafting or research in the same way as the standard approach. However, a notebook has some additional options that are great to explore.”

One of those, he notes, is a feature called Audio Overview, through which the notebook can provide live commentary. That may mean something as simple as reading aloud its responses to your prompts. But more interestingly, it can also talk about any data it is analysing in the conversational style of a podcast. “That’s very handy for helping you to absorb large documents, if you’ve fed something weighty into the notebook’s dataset,” Penson says. “It gives you an on-tap sounding board for digesting complex information.”

However, a notebook really comes into its own providing a venue in which multiple users can collaborate.

On the same page

“Accounting and auditing work tends to be very document heavy,” Penson says. “You’re typically looking at lots of PDFs and a huge number of spreadsheets. Plus, you may need to reference areas of HMRC’s website, along with webpages from relevant regulators, such as the Financial Reporting Council. On the client side, you’re likely to be dealing with working papers and trial balances, or helping to finalise lettings agreements or property investments. A notebook enables as many team members as required to pool their skills on the same project, in a location where everyone can keep track of progress in real time.”

Penson notes another advantage: referencing only documents within your dataset, you can guarantee the notebook will provide accurate citations. The self-contained nature of the workspace means the notebook is not going to be marred by bad data linked to the documents’ subject areas that can wash in from other parts of the web, or from incorrect previous outputs generated by AI.

“That will help you achieve the right levels of compliance and traceability,” Penson says. “In an audit team, you need to ensure that everyone is researching and referencing the same set of sources. You don’t want people going off on their own paths, for example, by using another AI to generate responses based on different sets of documents. You want everything to be 100% consistent – not just in terms of the source documents, but the notebook responses that your firm will end up referencing in its paperwork.”

Follow a framework

For Penson, when it comes to project work, a notebook beats the standard Copilot chat function in two important ways.

In a standard situation, he notes, you need to re-explain the context for every new thread. Conversely, a notebook is already grounded through your dataset and ready to pick up wherever you left off. Plus, if you receive additional reference documents to factor into your work, all that’s necessary in a notebook is to enter them in the dataset and re-run previous prompts. If you are running a chat, though, you will have to copy those earlier prompts over to a new thread and upload the extra documents. “Even then,” Penson warns, “you may not get quite the same structure in your responses – which takes us back to consistency issues.”

For accountants thinking of taking the notebooks path, Penson has good news. “They’re pretty straightforward to use,” he says. “If you’ve picked up experience with basic AI chat tools, you probably won’t find notebooks too much of a challenge. That said, I’d always recommend training. Microsoft has some great resources, and there are plenty of YouTubers who’ve published videos that will get you started.”

With that initial training in hand, he advises, play around with a couple of trial notebooks to get a feel for how they work. Don’t use them on vital pieces of live client work until you are comfortable.

Penson has two tips for setting up a notebook and putting it to use:

1. Ensure its dataset is comprehensive, but not overloaded: “Keep it as focused and precise as possible,” he says. “Don’t add too much information, because AI loves to ramble.”
2. Follow a recognised framework for smart prompting “We use the Goal, Context, Source and Expectations (GCSE) method, which Microsoft recommends,” Penson explains. “Alongside the GCSE framework, there are many others. Much like the dataset, the idea is to ensure your prompts are focused and productive, rather than vague or woolly.”

Once you are up and running, smart prompts will pave the way for smart working. “If there’s one task you’re repeating for multiple clients,” Penson says, “you may want to save prompts that have worked particularly well for reuse in new notebooks.”