In 2018, we are focussing on designing services powered not just on data, but on Machine Learning and Artificial Intelligence.
This has been our clear direction of travel over the last few years, and it now makes sense to focus our efforts in this area.
We have been working with data-powered systems since 2010 – witnessing the key shift from Analytics to ML/AI and from passive reporting systems to active experimentation systems – which requires a change in project and design methods that we will outline below.
Our current work involves two main types of projects:
Helping clients amplify their current AI services. Here we understand what the model does, how it learns and how it benefits users, and our role is to help users engage effectively with the services derived from this model. For these projects, the AI technology (models, prediction engine) is in a current version that we don’t expect to change. These projects are about providing a new, useful, effective interface to that system.
Helping clients develop the AI itself. As part of concept development, we are asked to develop a service built on AI. The AI specification is part of our work, alongside the service it underpins. We work with the technology and business analyst teams to understand the needs of users, to identify the potential of the available data, and to specify how a learning system can benefit the service. We work closely with developers and data scientists to scope the algorithms and model, including data to use for a technical prototype or version 1.
How we design AI systems
All systems are defined by its inputs as much as its resulting outputs, and it is this principle we use to organise our thoughts when designing for AI.
Inputs: there is no intelligent system without the right data
- Interface design for AI is as much about creating learning/input/feedback loops for the machine as it is in issuing instructions from the machine to the users.
- We work with clients to specify the data required to power and train the system and we define feedback surfaces and feature extraction opportunities to help users exploit the potential of the system.
- Depending on the channels available, we suggest a stream of inputs into the system – such as video feeds for pattern recognition.
- Our Performance Framework tool matches strategic and business needs to available and prospective data. It is important the the clients understand what data is immediately usable, or just speculative.
- Increasingly, clients need help in understanding the licensing rights for the data that is supplied to them. We help draft agreements to provide them with ongoing access to the data they need.
- How can humans help? For the foreseeable future, much of our work will involve creating input channels for humans/co-workers of the machines to train and steer the development of intelligence.
Intelligence: making the invisible visible
- The system is active – an ever-changing and learning engine. Much is written about the need for C-level staff to engage with their AI and ML services for both business and ethical oversight, but these systems are hard to understand. We, therefore, design interfaces to express the changes in the system so that the non-technical leadership of the company can better understand what it is doing.
- We enable product teams to carry out experiments on the data and the algorithms – these may not affect the real product but will be another thread that teams can pull on to understand the system better.
- Making the AI visible beyond the basic product and service also endears the enterprise to an increasingly robust regulatory environment for technology services.
Outputs: encouraging interaction
- Many outputs of AI are not intelligent in themselves. Some, such as Chatbots, have a semblance of sentience but are clearly not ‘intelligent’. We design outputs at a level of sophistication that matches the intelligence of the system.
- Alternatively, seamful design allows the strengths and weaknesses of a system to be highlighted. For instance, showing probabilities in predictions, along with the provenance of the data that is powering that insight.
- AI interfaces should also suggest multiple options that users can abstract and not just a single instruction. Rather than a didactic serving of information, the machine should serve options that train the intelligence depending on which are selected.
- Whereas complex data visualisation typified the early age of Big Data by unwittingly celebrating complexity with illegible patterns, the AI age must aim for clearer communication – often using verbalised data because words are universally understood.
These examples are taken from real projects and will become firmer principles as we progress.
If you want to talk about how your company is using AI and explore how we can help, get in touch.