Small and mid-size business platform Sage is already well experienced with AI, having woven it throughout products for years, but its new generative AI Sage Copilot solution has been touted as a dramatic step forward in the company's long-term vision of simplifying accounting to make it more accessible to all.
Aaron Harris, Sage's chief technology officer, said the company has long had classic AI deep learning models that perform functions users have relied on for years. These tasks often require the use of many different models working in concert, with Harris noting that invoice processing alone requires the use of 27 different models: 15 to 20 are required to read the data alone, and more are needed to perform the calculations.
Sage Copilot coordinates between these models and acts as an interpreter between them and the human users who are requesting they perform a task. Effectively a "mouth" attached to a larger whole, the LLM works by translating the user request, usually inputted through a conversational interface, into machine language. This then goes to the various AI models on Sage's servers, which then get to work on whatever the user asked them for, eventually sending instructions back to the LLM. The LLM reads back the machine language and implements the command or, in the case of an informational query, translates it back to human language. Harris stressed that it is not the LLM itself that does this work, referencing their well-known difficulties with math, but the other AI models that the LLM interacts with.
"We don't trust them to do math. There's much better ways to do math. … We're not using the AI to do the math on the results. We're using AI to write the [
He added that, rather than being a feature of any one particular solution, Sage Copilot can be used across its products through the use of specialized "agents." The company creates AI "agents" purpose built to do one thing really well, such as interacting with certain types of data, executing specific queries, or engaging with specific software products. Sage Copilot has access to multiple agents, each built to communicate with a different product, whether Sage Intaact, Sage HR or something else.
"The intention is that Copilot can work across the portfolio of Sage products," he said.
Having Copilot act as a coordinator for all the other models also means its automation capacities go far beyond what it had previously accomplished. Harris raised the example of processing an invoice. This act alone requires multiple steps, but many of them have already been automated in Sage, taking out much of the work. Copilot goes a step further by allowing wholesale workflow automation through coordinating among several models that each enable a different automated process. "We can now really accelerate our ability to automate with large language models, and we can use AI to do more of the orchestration. So, giving it the ability to not just process the invoice but to move it on to the approval stage, to understand after it's approved [it needs to] move it through the payment process," he said.
He added that, in the future, "that invoice will have been created for you, automatically." His team had recently conducted a hackathon where it found Copilot can automatically generate invoices based on events happening around the user, such as if it knew they were going on a job using geofencing, and act proactively.
Harris said that certain companies today will attach their solution to a ChatGPT account and call that generative AI functionality. In contrast, developing Copilot was a meticulous process that required a lot of trial and error even before its U.K. release earlier this year. During this time, the development team encountered many challenges that needed to be overcome, such as teaching the model that there can be more than one kind of cash balance.
"I got in and I asked Copilot 'what is my cash balance?' And Copilot said it's at zero. And I dug around and what I learned was that Copilot inferred from my question, 'What's my petty cash?' It didn't actually understand that what I was asking for was the balance of cash across my bank account," he said. "Fast forward a month. We've done a lot of work to train the model the way we wanted to when you ask that question. … [Now] it'll give you a table with your bank accounts and your balances in total. If that is what you're asking, this is the answer."
Getting these kinds of interactions right was vital to ensuring Copilot was easy to use and reliable in its outputs without having to possess a lot of arcane technical knowledge. Interacting with Copilot in plain language allows people to access accounting information and perform business tasks on their own, a major component of the wider goal of making accounting more accessible to a wider base of people.
"Generative AI and copilots, and their natural conversational interface, enable us to bring accounting outside the finance team to the rest of the business. One of the biggest blocks is getting [accountants] to approve things or to answer questions, finance teams are spending time supporting me instead of getting the books closed. With Copilot having a conversational interface, it now becomes much more natural and easy for me to approve a purchase order or to ask a question like 'how am I trending on my travel expenses?'" Harris said.
While Sage wants to make accounting more accessible to the layperson, he added that professionals can be excited too. The big promise, he said, is that it will free them from things they don't want to do. He compared it to having an army of interns at one's disposal who can take care of the numerous mundane demands that pop up throughout the day. The result, according to Harris, will be "faster and smarter decisions."
"Unpacking that, what we're really saying is [this can be] how you, across the whole of the business, understand the patterns of activity in that business in real time to discover when there is a change in performance—when there's something that can indicate an opportunity or risk—that the more strategic specialists can address in real time… We're enabling more decisions to be made confidently," he said.
Harris said the name "Copilot" represents what he felt was a good bet that the term copilot would become generic, versus being permanently associated with Microsoft's product. He said that the term has, over time, emerged as standard in a similar manner as "Band-Aid," "Xerox" and "Google."
The large language model was released in the U.K. in February and is set for release in the U.S. at the end of this year.
Part of a larger strategy
Copilot is one component of the company's larger product strategy to promote continuous accounting, real time assurance and continuous insights. But this, itself, is part of Sage's overall strategy, particularly for the North American market; Mark Hickman, the managing director of North America, said Copilot is "critical to our success."
"As we move forward, [we want] to really be that leader. We want to be ahead of the competition when it comes to AI and how we bring that to market, into that ecosystem of millions of customers globally and hundreds and hundreds of thousands in North America," said Hickman.
To this end, Sage has been busy making new alliances and deepening current ones with companies like Microsoft, Amazon Web Services and PwC. They have collaborated on technology solutions with the aim of eventually driving integration into products like Office and other platforms, as well as on distribution and implementation of said solutions. With Microsoft and AWS in particular, Hickman said they have whole partnerships where they go to market together and close new customers. Given these companies' focus on large enterprises, Sage's focus on small and medium businesses has acted as a bridge to this larger community.
"What we've discovered here is that [Microsoft and AWS], they don't really play in the small to medium businesses with the cloud. So 90% of their new customers are net new customers so they're actually getting into new customers because they're working with us and closing new deals to get into these accounts and … using their amazing, world class platforms and their brands to work together," he said.
This is especially germane as the U.K.-based Sage expands further into the North American market, which makes up more than 44% of the company's global revenues already. The company is making heavy investments in this region, which include technology but also additional staff as well as new facilities. Hickman said Sage will be building a whole new campus in Atlanta to serve as its new base for North American operations (in addition to the office they already maintain in that city), as well as new offices in Portland and Vancouver.
Hickman said Sage's thinking on these new locations came as the company emerged from the pandemic lockdowns, eventually settling on what he called a "hub strategy." Previously, the company had over 80 offices around the world, but he said many of them were small and remote. The company chose a very deliberate strategy where it would instead have large offices in each of the major markets they really invest in, with fewer small satellite offices. This has allowed them to really focus their efforts around these flagship country "hubs." He noted that the new offices in Canada are also a reflection of this hub strategy of "really increasing the investment in the offices where we want to concentrate our growth."
"So the North American businesses, the U.S. businesses, are the fastest-growing [sector] with great growth, and we hope to really accelerate that growth as we move forward," he added.
