AI: Unpacking the buzz, the bluff & the big questions
About
Tech on the Tyne was held on 27th November 2024 at Sage in Newcastle upon Tyne.
Lisa Ewens - SVP for Small Business Product, Sage
Sage Copilot is embedded into all their customer facing tools to see how they can save their customers time with admin tasks and key processes in their products. Sage Copilot launched in May 2024 and are now live in five countries and in seven products, customers did mention some things have felt gimmicky so they need to make sure they focus on solving real world problems.
How do you upskill colleagues? On their journey at Sage, they had colleagues asking questions including about will AI replace their jobs. They ran quarterly hackathons and ran seven in total with the first was one hundred colleagues attended with a backlog of cases they had to think about including how to use Ai to support QA or help solve customer problems and formed squads and couple of days to ideate and work on prototypes with a prize for best one. The last hackathon had over 500 colleagues with over 900 who participated in total who got more confident and knowledgeable resulting in things they have rolled out such as scripting and tests for QA or onboarding tasks
These hackathons were better than just self-learning and have leveraged partnerships with Microsoft and AWS who provide tools to help build AI solutions. Without the hackathons Sage wouldn't be in the position they are as it really resonated with colleagues and in later hackathons, they didn't just see developers but product managers and designers using tools like Lovable to get results but they don't run hackathons as it has moved into business as usual but could run them again if needed.
Another thing Sage have done is supplement with experts and focus on early careers to bring new talent into their teams. Sage have had graduate programmes before of senior colleagues teaching but now senior developers are often learning from those same people. They bring in 40 graduates per year including students who have used AI as part of their degree and they are less nervous about using it and have more experience in many cases.
Something else Sage have done that is interesting is their smart engineers have built an agentic framework to leverage best of breed LLMs and things like GitHub Copilot to build a framework to use and experiment with these technologies in a controlled way so that things are secure and not delivering inaccurate experiences to customers and helps those colleagues who aren't experienced in building these agents in a low code way, with an aim is to have all Sage colleagues work in this way especially the more experienced developers who may only have desktop experience.
Final thing which has been key part is putting customers at heart of AI development and Sage have worked closely with customers than ever before including early adopter programmes and have people taking to customers about what they are scares and nervous about and keep the human in the loop and are truly solving customer problems that engage their users. They have a growth squad who looks at who is using AI products, drop off rates and satisfaction including detractor scores or positive sentiment so it had been an enriching experience to work closely with customers and have driven colleagues to deliver on customer outcome.
Christian Steanson - Solution Designer, Sage
Sage has an early careers programme where Christian joined and later transitioned to product and that is when Sage Copilot became a thing and they were given a chance to work on it but had no experience of AI but learned and absorbed as much as they could about AI.
What is trust with AI? Trust is the belief in the reliability and ability of someone of something and trust in AI is not believing AI is perfect but it is about it being responsible and safe to use and trust levels in adoption of Ai can vary with context and in accounting trust in AI means everything and the margin for error is razor thin and mistakes can have real world consequences so trust with AI varies with product.
How do you obtain trust in AI systems? Trust takes time and you need to build trust, and users have no reason to trust that and responsible AI at Sage is at the heart of this and having a human in the loop is a great way to build trust. Is driving through the Tyne tunnel twice there and back a duplicate and need to be ignored or is it a situation such as having to go back after forgetting something at home.
What is responsible AI? This is focusing on creating features that are ethical and fair and without reinforcing bias and unfair outcomes. Responsible AI principles include fairness where should treat users equally, transparency is one that comes up a lot to explain where AI is being used and making decisions, accountability and data with AI ethics to make sure handling data in right way and need to ensure data is as high quality as possible. Personalisation should allow users to customer AI behaviour to control where it shows up and continuous improvement is where AI should evolve over time.
Responsible AI matters as it minimises harm, builds trust, ensures compliance and make sure that AI has a good long-term impact and make sure we do the right thing. Transparency is all about explaining decisions and showing where AI is in use and acknowledge where AI is getting things from and make sure human is always in the driving seat. Sage has their responsible AI label to show where and how they are using AI. People don't fear AI but fear untrustworthy AI and we need to build the kind of AI people trust.
Karen Ainley - SVP Accountants, Sage
One of the biggest challenges with sole traders is a revolution of digitisation of tax for income tax and anyone who is a self-assessment who has a trade had to use software, so Sage are building an agentic workflow to automate and work in an effective way with an MTD agent for making tax digital which can help get set up for clients and submit quarterly updates and add automation into the update itself. This agent can include checklists and ways to request any missing data and anomalies where can review this information and can pick what elements of a checklist they want it to do and things to check and can save searching through data and find areas to focus on including how much control you give is up to you using the settings for this.
Clients can also be assigned tasks with an email with task from an agent such as uploading missing bank statements and can tell Copilot what you want the email content to be including how formal and detailed it is and can change this in settings to use the tone and language you want. They can send an email or a WhatsApp message along with notification in software to perform that task. Once information is gathered can send information to HMRC or can send an approval request to client to enable update to be sent to HMRC.
Using AI and automation to streamline tasks but this has been built by working with customers and testing and learning and see what works and doesn't but need to have customer and accountant in control otherwise will alienate them, have to have tasks they are happy to do or the ones they want Sage Copilot to do.
Q&A
How are Sage tackling phishing regarding getting bank statement and how do you know this is genuine and how do customers know it is a genuine email? Email sent provides a link with MFA or multi-factor authentication and it wouldn't be an unauthenticated link, but it is something they think about email ending up in spam folders. However, the key thing is getting customers back into unauthenticated experience. How do you know the statement is genuine, spot if a document is fake can train a model to spot nuances and be able to identify things and way things are displayed. How to spot anomalies is something they need to watch out for, and it is a big challenge and AI tech is evolving and need to keep up with technology of AI such as deep faking and others so fight against it needs to match up. Direction connections to bank or e-invoicing and partner with bodies to reduce fraud. Every single invoice from 2026 must be a specified format and things must go through a specified system for example we will be doing e-invoicing from 2029 in the UK.
How do you get AI to avoid errors to make sure avoid problems? Sage is investing in different machine learning models and see if something doesn't look right and they have outlier detection at experimenting and investing in this but is early days of this and you could do some things offline but need connections to use LLM.
What is there is an inconsistency with what has been done? It goes back to human in loop, and they never do anything on behalf of the user, don't make a change and the human is liable as they make a conscious choice and decision. Can have someone upload an invoice with a business expense that included something they shouldn't be there can detect this with a model. There was an art competition that used AI to win but who won it was deemed in the end was person who submitted it, the model and everything the model was trained on was the true winner.
How do you present the idea of doing hackathons or to make a business case to do this? Before the hackathons at Sage, they had a centre of excellence with Ai for research and development and already had an appetite for this. What ideas do they have to upskill the organisation and it was there boss who was the budget holder so was able to take this on themselves and other teams and functions were onboard and it was three days out of your day job to play around with things and build things with AI, and it wasn't just engineers but everyone who got involved. They are still on the journey and are looking at how they can build things first and whilst they have stopped the hackathons they have looked at some of the things they need to do to get things done and how to solve things and get on that journey and want to get agentic first and have looked at their agentic processes and looked at what has worked and not worked with AI and then see what they can do with something and then if it works bring it to customers.
Have Sage considered anything for anyone using repurchased accounts for checking invoices? They have looked at ways customers can solve their own problems, not checking what they are doing but helping them with their own outcomes. Client takes ownership of information, but there are checks and balances but the responsibility of filling out tax return is on the business owner. Payments from business to customer will be helped by e-invoicing as will be able to balance up figures between that and tax return. In Portugal each spend has an id and they know what you have spent but is automating personal and business tax.
Has Sage developed their own models, or have they fine-tuned pre-built models? Sage do have their own small models but for larger ones they fine tune data with existing models. When looking at deterministic and non-deterministic models and using more generic models didn't know about specific invoices so used small models to do more specific things.
Do they use AI to filter job applications? They have seen documents and presentations that have been written by AI but as long as someone had thought about it and it is relevant to Sage so if someone can demonstrate what they know and anyone who isn't using AI they would be suspicious of including lack of evidence of using AI at all but a lot of other businesses do filter CVs with AI.