What we mean when we talk about AI and machine learning
How we define AI at Xero and why it’s a team effort
Artificial intelligence (AI) and machine learning (ML) have become so embedded in our everyday vernacular that it can sometimes be difficult to stop the conversation and ask for a bit of clarification on what it all means. With many organisations pursuing these technologies, it seems timely to pause and re-introduce these terms in ways that bring engineers, designers, product managers, and scientists into alignment.
Let’s start by defining AI
When we were thinking about how to define AI at Xero, we wanted a definition that would offer a holistic view and inspire collaboration. Inspiring collaboration is critical, because growing our AI capability at Xero isn’t about one team building one thing in isolation. It requires a whole system approach, where every team across the organisation is involved in some way. Understanding that it’s a collaborative effort is the first step to building truly effective AI technologies. So here’s what we came up with:
“Artificial intelligence is the capacity of a system to act appropriately within an environment to achieve a goal.”
It’s a reasonably short sentence, but many of those terms mask complex and potentially contentious issues. You could easily write an essay on each one. For example, how do you measure capacity? Categorising and quantifying intelligence is an area that to date remains contentious and imprecise. This also extends to acting appropriately — is there a right way to achieve a goal and how do we measure ‘rightness’?
And finally the goal itself. Who defines that goal and what motivates them? How do we make sure our AI technologies benefit our small business customers above all? These terms are just as important as the definition itself and require a good deal of thought and planning, to make sure our AI work is not only considered and collaborative, but also aligned with our values.
So how does AI actually work?
Now that we’ve outlined how we’re defining AI at Xero today, let’s look at how it works. I’ve chosen a simplified model that allows us to understand the different elements of AI without delving too far into mathematics. In the stimulus response model, something happens in the agent’s environment (stimulus), and the agent sees that and chooses a response, based on the objective they’re trying to achieve. The space between the stimulus and the response is where intelligence occurs.
There’s a quote, often attributed to Victor Frankl, that sums it up nicely:
“Between stimulus and response there is space. In that space is our power to choose our response. In our response lies our growth and our freedom.”
It’s often easier to think of the agent as human, because that’s the kind of intelligence we’re most familiar with. So what happens when you use your intelligence? Well, your brain might do a few things. First, it accesses your memory. Then it will associate the thing you’ve just seen, with all the things you’ve seen in the past. You’ll then use your natural reasoning skills to make a decision.
For example, you might see a lion and use your memory to recall all the animals you’ve seen before. You then associate what you’ve just seen with all the lions you’ve seen or heard about in the past. You reason that you’re probably in danger, then make the decision to run (although that’s probably a bad idea, because the lion will instinctively chase you).
Now, if we’re talking about Xero and an artificially intelligent system, then each of those functions might be done by software, APIs and databases. For example, a database might stand in for our memory. We might cross reference multiple sources of information by communicating with other internal services (APIs) to associate what we’ve just seen with what we already know. And the application code is responsible for reasoning and making some kind of decision.
That’s it. That’s what makes up an AI product. Are all decision-making systems technically AI? Well, yes. So why don’t people get excited about them? There are probably two reasons. One: they’ve been around forever and have lost some of their appeal. And two: because the sort of AI that fills the current zeitgeist can handle really complex decisions.
Making complex decisions
When we were limited by systems that could only make simple ‘if this, then that’ types of decisions, there was an upper limit to how complex AI could be. Consider a video game, like Pacman or Starcraft, where there are potentially hundreds of actions you can take at any given time. The decisions you make in that game depend on a number of different factors and might change every time you play, so it would be impossible to try and code them in simple ‘if this, then that’ statements.
Instead of simple statements, modern AI uses a mathematical function that can make decisions involving complex inputs and outputs. This means that incredibly complex decisions can be made by one algorithm rather than an endless number of statements. This significantly raises the limit on the complexity of those decisions, or the types of inputs that are required. This is a huge step forward and one of the main reasons that interest in AI and machine learning has increased in recent years.
What about machine learning?
Conversations about artificial intelligence often include references to machine learning, and unfortunately these terms are often used interchangeably, despite having distinct and separate meanings. So what is machine learning? A helpful definition we use is:
“The study of algorithms that improve through experience”
You can think of experience as a collection of observations — basically where something happens (input) and someone took an action. So at Xero, it might be that every time a bank statement line said 7-Eleven, you reconciled it with the contact that’s called 7-Eleven. Every time you saw that input, you took that action. Over time, the collection of those observations becomes experience.
ML algorithms are able to take this experience and distil the relationship between inputs and actions (or the 7-Eleven statement line and the 7-Eleven contact, for example). They do this by effectively memorising how the human is mapping inputs to appropriate actions, or ‘watching’ them complete a task. Modern algorithms such as neural networks have proven to be incredibly effective at learning on a wide range of tasks, such as object recognition and even writing stories!
When you take an ML algorithm and insert it into an AI system, you can unlock new and incredibly exciting opportunities — building experiences that learn over time and have few limits on their complexity. For Xero, that means we can help customers automate many of the tasks associated with running a business, so they can use their human intelligence for other (more valuable and enjoyable) things.
Building AI products is a team sport
When building an AI product, it really is a team effort. While some folks might work on converting experience into better input-action maps (i.e. machine learning), other folks stitch together the systems needed to orchestrate the movement of information and execute the chosen action. Ultimately, this collaboration is more important to the success of an AI product than the algorithms themselves.
Beyond the teams building AI products, every team at Xero and every part of the product contributes to our ability to build intelligent products. Why? Because modern ML algorithms tend to require enormous amounts of data, and the quality of data collection determines the amount of ‘good’ experiences our algorithms can learn from.
Building AI products is certainly not for the faint hearted, but it’s something we’re committed to doing well at Xero. In fact, this commitment to placing customers at the centre of a considered approach to AI, and ensuring we do right by our customers, is part of what inspired me to join Xero in the first place.
