My AI versus the Company AI

How Knowledge Workers Conceptualise Forms of AI Assistance in the Workplace

Last year we worked on a multi-national ethnographic study with Google to explore and define the type of AI-driven assistance knowledge workers wanted, and develop a framework to guide innovation across G Suite's product teams.

This will be a familiar interface for many people. Gmail has over 1.2bn active users. The predictive writing feature 'Smart Compose' is an example of what Google calls ‘Creative Assistance’. There are many similar features embedded across G Suite (including Docs, Sheets, Slides) which try to assist people doing their work.

…But when we started working with Google last year they didn’t have a user-focused framework to guide this innovation – to guide where these interventions should happen and how.

So our objective was to help G Suite develop a strategy for integrating AI across their product portfolio, and, importantly, to do so from a worker perspective: Google’s sales strategy is ‘bottom up’. Partly to create a better user experience, but also in the belief that if they can create assistance that individual users want, then their organisations will follow.

We studied knowledge work across the UK, Germany, France and Switzerland. We combined in-depth interviews with people across a range of industries and roles – from lawyers to designers to administrators – with 4 organisational ethnographies across industrial, apparel and CPG companies.

We started our research by looking very broadly at knowledge work. What is it and how do people experience it?

When we think about knowledge work we focus on the more complex, theoretical tasks: analysing data, writing presentations, making complex decisions.

The everyday experience is very different: Peter, a strategist at a CPG company, put it well in his comments about meetings. Consistently there was a distinction made between “work” and “real work".

We synthesised this with the concepts of ‘core’ work and ‘peripheral’ work.

Core work is what people think about as being their job spec. It is what they get reviewed on at the end of the year. It is how promotions and raises are decided. For Louise, a category manager for an apparel business, it was about being on top of consumer trends.

Core work also relates to people’s sense of professional and personal identity.

Workers are motivated to increase the amount of ideal work in their life.

There is an acknowledgement that at the beginning of your career you may not have much ideal work – but as workers progress they want to increase time spent on ideal work (and this can happen within their job, or outside their formal workplace).

Workers are motivated to protect a strong sense of workplace identity.

Strong work identity makes individuals less vulnerable to stress in ambiguous situations, because their ideas of themselves as professionals are not easily threatened.

Unfortunately our research demonstrated that people are spending less time on core work than they would like. This is a day in the life of Tina who works as an analyst at an industrial company – this was typical in that only 3 hours were spent on her core work (and this is a good day).

When we talked to workers about assistance we framed the conversation in broad, human terms to avoid pre-existing associations with AI narratives or to allow the research to be biased by existing consumer instantiations (like Siri, Alexa or Google Assistant). This included mapping specific tasks against a spectrum of complexity and whether or not they wanted help.

What we discovered was that people wanted assistance to enhance their core work. Importantly, so that they retain agency, so that their skills becomes extended and augmented rather than replaced. For example, if your job is to design and run a workshop, you might want ideas – but you don't want it to be designed for you.

In contrast, they wanted peripheral work to be offloaded to someone or something else. Importantly, it still had to be done right. For example, expenses.

This generated a set of design principles, which were relevant regardless of whether assistance is being provided by human or machine.

But it wasn’t that simple. Specific tasks tended to fit somewhere on the spectrum.

Synthesis: summarising meetings – people wanted help to ensure they didn’t miss anything, but how they prioritised the meeting notes was up to them.

This had implications for the kind of relationship workers wanted with the AI:

Enhance work: Worker maintains control of task

• Reputation not immediately at risk

• Greater permission to fail (or present options)

• Perception that task is complex and requires judgement

• Offload Work: Worker relinquishes control of task

• Reputation is at risk

• Must get it right first time (or solicit feedback)

• Perception that task is simple and should be understood

Interesting insight – work that is peripheral requires higher level of trust to complete

This division was reflected in the sense of ownership and agency the worker wanted to feel over the AI

Peripheral work = 'Owned’ by the Organisation

• Organisation only takes credit for ‘peripheral’ work

• Organisation takes responsibility for failure

The AI stays put when I leaveCore work = 'Owned’ by the Individual worker

• Worker takes credit for ‘core’ work

• Worker can mitigate / manage failure

I take the AI with me when I leave

So what kind of impact has our project had so far?

What kind of AI is Smart Compose?

Of course it depends what you’re using it for. Setting up an internal meeting = peripheral. But generally – it’s core, if it’s a personal email then it's something sent by you. People want to retain a sense of ownership over this for obvious reasons.

We broke down lots of very specific tasks against our framework to demonstrate to Google how to deploy the model. You can read more about this in the paper.

The frameworks outlined in this case study have been socialised across both the executive and product layers of the organisation, helping teams prioritise and develop strategies. Previous to this work G Suite had many successful products that provided AI-driven assistance for knowledge workers, but lacked a foundational framework with which to categorise and evaluate existing products from a user perspective, nor a clear means for understanding where to innovate in the future. Our work has provided G Suite management with a set of adaptable tools to organise and manage innovation across product teams.

Each product team at G Suite (Gmail, Calendar, Sheets, Docs) is now using these frameworks to inspire and guide how they integrate AI into their products, with many examples already live.

There is a specific question we’d like to explore further that emerged from the study. Knowledge work has become increasingly complex, and in that context the idea of personal agency and autonomy can be increasingly tricky to define. However, people are still assessed as individuals. Given this context, what can AI do to delineate roles and help people take credit for core work? Or does the idea of core work itself need to become more collaborative and team-based?

// Tom Hoy & Nanna Sandberg

This presentation was originally delivered at the 2019 Epic Conference in Rhode Island.