Save content
Have you found this content useful? Use the button above to save it to your profile.
Diya Jolly at Xerocon 2024

Xero tech chief expands on AI ambitions


Cloud accounting platform Xero has revealed more details about JAX, its generative AI assistant, including its proprietary technology, safety features and what benefits customers may see from the product.

19th Jun 2024
Save content
Have you found this content useful? Use the button above to save it to your profile.

First announced in March this year, Just Ask Xero or JAX is a generative artificial intelligence (GenAI) assistant designed to enable accounting or business users to complete tasks like generating an invoice or editing a quote more efficiently, either using Xero directly or via apps and devices such as mobile, WhatsApp or email.

At its Xerocon 2024 show in London, Xero’s chief product and technology officer Diya Jolly (above) announced that JAX will be available for beta testing in August 2024, with no date yet fixed for its general availability.

A demonstration video accompanying the announcement showed several examples of how the assistant could be used and interacted with using natural language prompts. 

The video showed a Xero accounting partner using JAX to gather information on how much a client is owed by a particular customer, pulling data from their Xero account and offering to present a summary of the largest unpaid orders as a follow-up. 

The user was then able to ask JAX whether the customer usually paid on time – in this case, they didn’t so the system suggested adding a late payment fee to an outstanding invoice. By replying “yes”, the user was able to add the fee and generate an updated invoice, and JAX offered to draft an email to the customer informing them.

Also showcased in the video was the ability for a small business to amend a customer quote on Xero using WhatsApp to interact with the system via JAX.

Once live, both accounting and small business users will have access to the assistant. No details are available about whether JAX will be a chargeable service, such as Microsoft’s Copilot, or included as part of the Xero subscription fee. 

Model systems and safety measures

Speaking with AccountingWEB at Xerocon, Jolly revealed that JAX is built using a combination of pre-existing large language models (LLMs) and Xero’s own proprietary models and technology. 

The proprietary aspect of the technology ensures greater accuracy than artificial intelligence (AI) systems that rely only on general large language models (LLMs). JAX uses a control system to ensure the data that’s fed to the LLMs is only related to the relevant task, such as amending an invoice. 

The new proprietary technology provides data security in line with the broader Xero platform, ensuring the LLMs only receive data that the user has given Xero permission to access and keeping this data safely within the platform.

“Large language models are incredibly powerful but they can be flawed,” said Jolly. “If you take an off-the-shelf model and let it run loose on the data, there’s a chance of it ‘hallucinating’ or making up answers – you can’t have that in accounting. 

“We constrain the data on guard rails in the form of a control system to ensure data fed into its large language model are only related to the relevant task. The way we guide the LLM models is by breaking down the task it has to do into smaller tasks so that it reaches the right answer.”

Testing parameters

For its initial testing and launch phases, Jolly told AccountingWEB that to find out where the tool is likely to have the most impact, Xero will make JAX as broadly applicable as possible, with functionality that both accountants or bookkeepers and small businesses can use.

In the testing phase, Xero will look at whether JAX can handle users’ questions from a functionality standpoint, what questions are asked, and whether the system can get to the answer in one go or needs a couple of goes to get to it. 

“We’ll also look at whether our follow-on suggestions are actually taken, whether they’re really helpful and obviously if they save our customers time,” added Jolly.

Significant leap in the technology

Accounting systems have experienced at least one wave of AI hype before, with several much-hyped chatbots now consigned to the recycling bin of digital history. 

Jolly, who prior to her time at Xero held senior positions at Google and YouTube, was keen to stress that this phase of AI development is likely to result in significant change for accounting and in more general use.

“I think two things have happened [in this latest wave of AI] that make it different and very relevant for this industry,” said Jolly.

“One is that there has been a significant leap in the technology itself in terms of what it can do from a behavioural perspective. Earlier, you had to train AI models on very specific data sets, and then it could perform only very specific actions. Now you can train them on larger data sets, and they can perform a range of actions, which allows you to have AI agents that complete full tasks, sometimes without specific training. 

“The second thing is you can now converse with them in natural language. That allows you to combine AI and accounting, and apply the models to a range of different tasks. If you want insights you don’t need people to trawl through and analyse a tonne of data – if you know the right question, you can ask JAX to extract those insights from the system in natural language.

“You can then interact with them on any device or any kind of app, as we’ve shown with WhatsApp. All of this was impossible before. You couldn’t just tell the AI: ‘Hey update an invoice or a quote’ and it would just do it. Those things are new, and I think that’s different in the past couple of years.”

Ultimately, the proof of this and other AI systems will be in whether accounting users decide to trial them, whether they find value in the systems and if they trust the accuracy of their output. On the first point at least, Jolly seems confident that Xero’s users will take up the challenge.

“I think people will try them because they want more efficiency,” said Jolly. “If there is more efficiency, there is an incentive. That’s why accountants moved to Lotus 1-2-3 or Excel all those years ago – because there was more efficiency and more insights.”

Replies (8)

Please login or register to join the discussion.

Rob Swan
By Rob Swan
20th Jun 2024 08:40

"The user was then able to ask JAX whether the customer usually paid on time...."

TBH... IMHO... Utterly underwhelmed!

If you're running a small business or accounting/bookkeeping you SHOULD know that anyway AND you should have had a 'conversation' (with the offender) and/or put appropriate measures in place - or just accept it. If not, you're probably not doing a very good job and maybe you should go spend some time learning how to run a business properly.

Sounds (so far) like Xero are promoting a 'Nothing Burger' to keep up with the trend. And no doubt the gullible will be charged a hefty fee for this 'feature'.

Why can't Xero (or anyone else) focus on producing good bookkeeping/accounting software with great features - so far I've yet to see any, (beyond a purpose written spreadsheet). We know they have a long list of user requests/requirements which don't get much attention.

Thanks (7)
Replying to Rob Swan:
By TB93
20th Jun 2024 09:36

There are plenty of great, if not necessary feature requests that have been outstanding for years with little to no attention, it does seem crazy that they leave the job half-finished and then move on to the next one.

In fact, they sometimes remove features that work and replace them with something that doesn't, the Pay with Wise feature for example being replaced with the open banking option. This new version doesn't allow an easy reconciliation to the supplier invoices being paid and the Xero support staff even know it doesn't work.

Thanks (4)
By TB93
20th Jun 2024 09:34

I wonder if we'll be able to ask JAX why the IT consultant paid £5,000 to Luxury Hot Tubs Ltd and claims it's all business

Thanks (4)
By Joe Alderson
20th Jun 2024 09:40

I don't think any of this improves the customers experience or the ability and knowledge of the accountants and their capability to do a good job. If anything, making people more reliant on AI will end up having the opposite effect in my opinion.

Thanks (5)
By johnfrancis
20th Jun 2024 10:52

All these comments are very negative - but we are in the early days of a new and sophisticated technology. If it has a lot to learn, perhaps so have the rest of us. I won't be the only reader of these columns to remember preparing accounts with a sharp pencil and a rubber - then I spent a week's wages on a pocket calculator. Of course I couldn't use that in my ACA exams, and Visicalc was not yet on the horizon.

Ten years from now we shall accept something that looks a bit like "AI" as an essential tool, and it might even make navigating cloud accounting packages a bit quicker and easier.

However, I sympathise with the posters who suggest that essential improvements ("feature requests") are neglected. I've sat on user groups over the years, and my experience is that software developers are not in the business of listening. They are anxious to deliver what they think their users ought to want, and are not particularly interested in fixing the bits that "just about manage" to work. Most of us have been just about managing to work with accounting and tax software for many years.

Thanks (3)
Replying to johnfrancis:
Pile of Stones
By Beach Accountancy
20th Jun 2024 19:24

At least there's the comedy value.

When Xero finally release a feature that the user base has been asking for for years (e.g. ability to search on the unreconciled items bank screen) you would have thought they'd invented nuclear fusion, and found the Holy Grail and the Loch Ness Monster on the same day.

Thanks (1)
Replying to Beach Accountancy:
Rob Swan
By Rob Swan
21st Jun 2024 09:40

I believe Sage are hunting Nessie ;)

Thanks (0)
By stepurhan
21st Jun 2024 11:48

The thing that gets me about so many of these announcements is how little of what it does sounds like AI.

How much a client owes? That's just a debtors report. What are the largest orders? Export to a spreadsheet and sort by value. Whether a customer pays on time. Export invoice lists with raised and payment dates and calculate the difference between the two.

It is simple data analysis. Nothing that requires the machine to demonstrate "intelligence".

Thanks (2)