#207: AI - A Decision Maker's View
Looking beyond technology, to business decisions and the role of AI
Over the past few months, every time I sat down to think about AI, I got lost in the range of terms between various kinds of Neural Networks, learning models, algorithms, types of AI solutions, and use cases.
I also realised that increasingly, with all the excitement about of Generative AI, there’s a halo effect in play. When a lot of people say Gen AI, they usually mean aspects of Gen AI but also of all other forms of AI, baked into the background. Which is another reason that rather than looking to distinguish between Generative and other AI, we should assume a level of inclusivity.
It’s taken me a while to sort this out to a point where I can make logical sense of it - where Generative AI fits in, and how we use other types of AI. Here’s a simple overview that does not take any kind of technical knowledge.
Complex Decision Automation
There are many definitions of AI, but from the point of view of business, I like to think of it as complex decision automation. At the end of the day, whether an AI system is playing chess, analysing an MRI scan, or detecting fraud in a set financial transactions, what it is fundamentally doing is helping to make better decisions by automating some part of the decision making process. And after all the core of a managers job is really to make the best possible decisions based on inputs, context and goals.
But let's dig a little deeper, and further segregate these decisions from an AI perspective. I believe that almost all our decisions fall into 4 categories - prediction, classification, generation, and optimisation.
For example, shortlisting candidates for interview is a classification decision. Deciding how much meat to buy for a restaurant is a prediction decision. Designing a slide to best drive home a message is a generation decision. Finding the most efficient way to plan a route is an optimisation problem.
Many of our decisions are also combinations of these categories. Writing a proposal is a combination of prediction and generation. Evaluating loan applications is a combination of prediction and classification. Social media marketing is a combination of classification and generation. And agreeing a care path for a patient is a combination of classification, prediction, and optimisation.
I use these buckets because these are also broadly map onto the types of AI applications, which we can use individually or in combinations to automate decision making.
Prediction drives value and profits. Much of our lives is based on prediction. You could even argue that every decision we make is in some way dependent on an element of prediction. AI systems can be used to make predictions - usually based on past data - ranging for footfall in stores, or probability of weather disruptions.
Classification is by itself a broad range of decision types. It includes ranking, sorting, selecting, de-selecting, odd-one-out, pairing, grouping, and many other forms of classifying models. We use this almost unconsciously every day. When you dress you select combinations of colours using some kind of a pairing or grouping model in your mind. When we order food from a restaurant we select and group from options based on implicit or explicit rules such as taste and health. And so on, in most our daily personal and professional choices and actions we are constantly classifying, choosing, rejecting, running those algorithms in our minds without even knowing it. A lot of AI systems focus on classification which involves data or image analysis, and in many areas AI tools have started performing better than humans.
Generation is the place we bring our originality and creativity to the forefront. Every time we write an email, design a slide, articulate a thought, or summarise a document, we are in the world of generation. Generative AI is a new area which has taken the world by storm because for the longest time this was the one area people assumed could not be done by machines and by AI. Generative AI itself covers a range of activities because the super power of Gen AI is language / syntax and meaning extraction - so you can point it at a number of areas to create, extract, summarise or structure ideas and concepts.
Optimisation is the holy grail of organisations. Every business, department, or manager is looking to find the balance between the most efficient and the most effective way to run complex processes, and functions, be it a supply chain, a vendor selection, or a retail store. Even before AI, expert systems and complex optimisation algorithms and tools have been deployed for scenarios where the number of variables is too large for humans to calculate (fleet optimisation) or where the required speed of calculation is too high (trading).
What AI systems seek to do is to achieve similar or better levels of efficacy in choice making. In order to do this, you might need an appropriate kind of learning model (supervised/ reinforced, etc.) which in turn might need a specific type of architecture (Generative Adversarial Networks, Recurrent Neural Networks, etc.). Let's leave that to the data scientists and AI experts for now.
How Much Automation?
What we need to also consider is the extent to which we automate the decision making process. Let's think about 4 levels:
At the top is a completely AI driven process. Let's call it 'Pure AI'. At this level, AI is entrusted with the entire decision making – currently useful for reasonably low-risk scenarios, that don't impact humans, such as quality control on a manufacturing line, or fine-tuning energy consumption of a building, by controlling light and heat based on occupation of the building.
The second level is 'Human in the Loop': In this level, AI does the analysis but a human reviews it before it goes out – typically used for scenarios where AI performance is high but humans are impacted – for example reviewing loan applications, or analysing scans in radiology.
The third level is AI Behind The Scenes or Partial Automation: The human does the work but has AI tools supporting them. For example a customer service agent who has access to tool that allow them to search for solutions to client problems, or a business using predictive maintenance models for assets. Partial automation is where we might only automate a part of a decision. For example, a decision might require prediction and generation – such as selecting and creating an image for a holiday brochure. We we might use predictive AI to help us with identifying which type of image would appeal most to our target audience but then use photographs shot by humans. Or we might do prediction ourselves and use Generative AI to create the images based on our direction.
The bottom level is Marginal AI, that is baked into products and tools such as face recognition on a mobile phone, or cybersecurity pattern recognition and alerting – used as triggers / support for other processes. For example you still have to make a payment yourself via your phone but face recognition is baked into the payment process.
Considering the 4 types of decisions and the 4 levels of AI inputs, it’s possible to construct a table like this - with examples of AI intervention and support in decision making.
(You might find that in reality the bulk of your decisions are combinations of these 4 areas, and it’s hard to classify them into any one box as neatly as this.)
It’s useful to think of a range of typical business decisions that we make everyday and explore how AI might be used to partially or completely automate these. Here are 5 scenarios:
Project planning - ChatGPT can be used to generate a standard and detailed project plan based on the type of project. Of course ChatGPT works when the reference base is large - so it would work for a (say) building an ecommerce website or executing an SAP project. It might work less well for something very specific - for example for a product that’s used by just your own company. This is an example of a type of generation decision. As per this piece, it can also generate more detailed plans when you connect ChatGPT to a project management system, and it can also create your risk log and cost estimates.
Planning a workshop - AI assistants can help find the best dates for a series of working sessions for a workshop, and optimise across the calendars and locations of multiple people. Gen AI can recommend a workshop outline, and also some ice breakers and activities. You can record workshop activities and have AI transcribe, translate (if required), and summarise the outputs.
Building a hospital: complex environments such as this will require a vast number of decisions. Some of these include predictions, about the relative capacity to plan for in an outpatient department, versus (say) phlebotomy, or intensive care. You might increasingly see the use of GenAI in areas like architecture and design. The layout of the hospital will require a series of complex optimisation models based on the expected journeys of patients on their visits to the hospital, and the experiences of the staff. And the entire supply chain selection of tools, products, and supplies involves a range of classification opportunities.
Decisions aren’t just Decisions
An interesting thing happened when I observed my own decision making. I looked back at my calendar and all my meetings for a week and tried to classify and describe my own decisions. I also asked my wife, who heads a ecommerce product group for a well known retailer to think about her decisions. Here are some interesting things that emerged:
A large amount of time is spent in getting us to decisions - either we’re gathering information, asking our teams to find out, or scanning the environment for the right signals, or weighing up the options. For example, as while designing an innovation space for TCS we had to make choices about space layout, technologies to acquire, and how to structure an engagement. Each of these is based on incomplete information so each decision is preceded by information gathering and scenario building.
Sometimes as a manager, there are 3 layers to decision making. First, sometimes the decision to make a decision itself is the need of the hour - we’ve all been in rambling discussions and it takes a leader in the room to call for a decision rather than unending debates. Second, what parameters to consider for a decision is also a decision. For us, whether or not to put in a specific technology into our innovation space would be driven by what percentage of our clients it was relevant to, whether it would be a salient demonstration of our capabilities, and of course the cost of the technology. And of course, the third layer is the decision itself, which also includes managing its outcomes, and communicating it effectively to all concerned.
If you think about these complexities, it makes the question of AI intervention more complex. In the first point above, can AI improve our information gathering? It’s been noted by a few people including the great John Cleese, that creative people are able to live with ambiguity for longer - in their world, the right time to make a decision is as late as possible. Whereas for a project manager, an early decision gives the most comfort. Could AI help us optimise the time of decision making?
And to the second point, can AI help define the parameters for a decision? The way most AI systems work today, the parameters a part of the training data. It’s possible that from a wide universe of parameters, an AI can select the right ones for a decision. But as of today, I’m not aware of any systems that can look at parameters outside of their training. So one of the interesting challenges is, in a fast changing environment how can we add parameters to our decision tools?
Leave the Technology to the Technologists
Just to be clear, there’s a lot more technical complexity to AI implementation. For example once you know the exact use case, and kind of AI you’re looking for, you may then evaluate tools and platforms, and classify the kind of architecture, the kind of learning model, and so on. That’s when you might get into neural network architectures - from GANs to Transformers, or ponder the difference between supervised, and unsupervised deep learning. In this post I’ve put that to one side, because I wanted to remind ourselves about the core aspect of decision making, and the role that AI can potentially play in it.
I watched this excellent Netflix Documentary called Count Me In. It’s about drummers in rock and jazz. There’s something about drumming that’s primordial. Besides, I also have a fascination for observing people who are really good at what they do and can also be articulate. I found this inspiring. From Nicko McBain (Iron Maiden), to Chad Smith (Red Hot Chilli Peppers), John Bonham (Led Zeppelin), Stewart Copeland (The Police), Keith Moon (The Who), Charlie Watts (The Rolling Stones), Cindy Blackman Santana, Stephen Perkins - you’ll find them all here. Watching this also made me explore some completely new music that I had not really ever been into. Cindy Blackman also makes some illuminating observations between Rock and Jazz drumming.
Health AI: In this study of 80,000 women in Sweden, AI found more cancer cases, and the false positive rate was 1.5% for both. The question to ponder isn't how good/effective AI technologies are, but how fast are they improving? The speed of change means that any judgement on performance needs to be retested more frequently than we are accustomed to. Benchmarks have a reducing shelf life. (Guardian)
State of Gen AI: A McKinsey Study reports that Gen AI adoption is unsurprisingly taking off. But it also says that a majority of businesses use AI in only one function. (McKinsey)
Trust in AI Decisions: Here’s an interesting case study of the extent of trust in an AI system that made over 17,000 allocation decisions, half with an easily understood algorithm and one with a more opaque one. The trust wasn’t higher in the more easily understood algorithm, and the article provides some food for thought in what creates trust in AI systems. (HBR)
Gene-Edited Microbiomes: You may know that our microbial systems are both extremely unique to individuals and also hold the key to a lot of our health. Scientists at the Innovative Genomics Institute in Berkeley are looking to genetically modify microbes to influence how diseases are identified and treated. For example, they are looking at reducing methane emissions in cows. In future the similar techniques might be used to detect and treat cancer. (MIT Tech Review)
Post Generation/ Perennials: Mauro Guillen, the Vice Dean of the Wharton School argues that with longevity and multiple generations at the work place, it’s time to think beyond individual generations. The collective melting pot of multiple generations is what he calls the perennials, who are all learning from each other. Having this kind of generational mix might be key to successful workplaces. Guillen has a book coming out on this soon (Fast Company).
Generative AI / Software 3.0: What makes Gen AI different? Its natural language interface, its pre-trained models, its foundational capability. But are we in a bubble? And what are the risks? Is differentiation going to be a challenge with everybody using the same tools? This is an interesting discussion published by Goldman Sachs.
Sustainability: A cargo ship called Windwings has sailed from China to Brazil using wind based power, which will reduce its fuel consumption and emissions by 30%. The technology is provided by Bar Technologies. Cargo ships account for 2.2% of global emissions. There’s a lot of opportunity here. (QZ)
That’s all for now, see you soon!
Subscribe for free to receive new posts as they’re published.