Introduction
Workflow design is the meticulous crafting and orchestration of work processes within an organisation or system. As you’d expect, it has progressed markedly over time in line with technological advancements. Once reliant on manual, paper-bound methodologies, it now embraces sophisticated, digital automation.
This discourse aims to navigate the historical progression of workflow design propelled by these technological innovations and to anticipate the transformative implications brought forth by AI technology in the immediate future.
Emphasising human-centricity in workflow design
It is essential to recognise and address the significant impact that user perception and engagement has had historically on corporate strategies. Startups have been effective market disrupters because they freely prioritise customer perspectives over imposing rigid organisational structures or technology constraints on workflow design.
Meeting user needs and expectations is a fundamental challenge of workflow design. By centring on the customer persona during the design phase — comprehending their objectives, preferences, interactions, and incorporating feedback — workflows can be tailored to optimise efficiency, efficacy, and satisfaction for both internal and external stakeholders. Not only does this alignment improve organisational goals, but it also bolsters user engagement, trust, and loyalty.
In the late 1990s, technology artefacts of systems were either split by company structural divisions or by product types, (in-line with 'Conway's Law', which states that the way an organisation designs a system will reflect how it communicates internally), or the non-functional limitations of Queue Management Systems and Batch technology to Mainframe. In addition to this, several archaic quartz schedulers and indexing performance constraints could be found in application databases. This resulted in operational constraints for after-hour changes to not impact the once time-bound working days of 9am –5pm, 6pm –9pm.
This changed around 2010 – 2015 with the introduction of OnDemand scaling with cloud. Architectural and business design was no longer held back by physical server downtime for operations or the need to scale infrastructure. Previously, onsite expansions could take a year when you needed to obtain and commission physical server hardware.
Along with OnDemand cloud, broader networking meant that New York, Hong Kong and London stopped being the only 24-hour finance cities and the world moved globally to being always on.
The digital shift in workflow management
As we entered the 2000s, technological breakthroughs with the first iteration of ‘bots’ facilitated the digitisation of human based fulfilment workflows — converting information and processes into machine-readable formats for streamlined execution.
Despite challenges like technology fragility and scaling issues due to hosted infrastructures' limitation, digitisation heralded improved workflow speeds, accuracy, and cost savings while also opening doors for complex systems supportive of real-time adaptation and inter-platform collaboration. Digitisation also brought with it duplication in location of data input and source systems, and multiple locations of business logic stored close to the graphical user interface (the web front-end). This led to large operational teams and challenges for systems to remain evergreen when the priority became managing regulatory changes into business logic by hard end dates. Businesses only have so much working capital to spend, and they usually chose to spend it reducing opportunities for certain fines vs. potential technology disruption.
Of course, with the 2010s being always on, business logic proliferation became more exposed and engineers and architects world-wide began to play with alternative options on how to manage centralised vs distributed workflows. This is something that we now see again in top forums around the nuances of on- or off-platform AI.
The dawn of Generative AI in workflow design
Generative AI stands at the cusp of radically transforming workflow design and where we focus our time. It holds the promise of generating original, personalised content that reshapes user experiences. With its flexible nature and learning capabilities, generative AI can produce tailor-made workflows that evolve with user feedback. It empowers users to partake in the workflow creation process, thus yielding unique and customer-driven experiences.
To achieve this, we place immutable business logic within a library, allowing it to be seamlessly integrated into workflows that invoke key processes with built-in governance. This approach not only offers scalability and flexibility to the customer but also promotes best practices. Additionally, it ensures that version changes and deployment times are fully traceable.
However, regulatory constraints remain a critical consideration for companies navigating the intricacies of business logic, which must invariably align with consistent, integrity-driven processes. For example, a customer may wish to design a process that provides a refund every time they aren’t happy, but in the dark reality of fraud there could be many opportunities where hyper personalisation like this could be used for evil.
Where do you start?
Clarity on AI tooling such as Machine Learning and Generative AI, together with understanding how risks are mitigated using continuous compliance in the development process is a good first step. Workflow design also needs a clear understanding between the use of Machine Learning vs. Generative AI in relation to composable capability.
Machine Learning involves training algorithms on data to recognise patterns. We then use those patterns to make decisions and provide predictions on new data. It relies on statistical techniques and can be divided into supervised, unsupervised and reinforcement learning. Note the word decision here derives from the Latin word ‘decidere’ - the ability to determine (machine learning uses a framework provided to obtain a result). Remarkably like the quality test frameworks we have today, you run a test with the parameters, and it returns the test result within the options.
Training machine learning to a high-quality outcome requires access to data. In some instances, consent may restrict types of customers who enable the use of their data for these purposes, and more cautious personalities will refrain, while early adopters lean in. Humanity is diverse, which means if we are not careful, we potentially risk training frameworks on a subset of reality. Therefore, so we are not causing harm, we need to use multiple other sets of data from industry and beyond to ensure gender, ethnicity and global customer diversity is applied to the model. (Find more on industry working through Diversity and Inclusion in AI Governance from the National AI Centre (NAIC))
The key thing to note here is that the more data used that covers points of diversification, the more likely we are to get a higher quality in the outputs. We also need to note that when we push an algorithm to production it does not take people’s data with it, it simply applies the learning it has gained for the purpose the model is trained, and it uses it to execute the job (purpose) we ask it to do.
A natural concern is how to mitigate any impact relating to what is being predicted and the potential actions people may take based on the information presented to them. We have these existing challenges today and we therefore use existing mitigations - terms and conditions along with referral to independent legal advice where decisions to be made by the customer and potential impacts are significant. The reality is for operational excellence, the work of classification and action could use existing software tolerance and quality capabilities to enable benefits without introducing risk.
Let’s take complaints. Complaints occur where customer expectations are not met, or we haven’t executed what we said we would. Digital customer experience design would naturally try to remove the likelihood for this to happen at the earliest point in the workflow. However, situations will always arise where we receive a complaint. We can identify through machine learning that a customer is not happy by using a classification of words used in dialogue based on other customer interactions on the same topic and tone. In Machine Learning we classify this as a decision because we have classified data as a result; ‘complaint’ or ‘not a complaint.’
We can make predictions that a particular customer or a topic may have a higher likelihood of a complaint outcome based on previous interaction records that have been classified as a 'complaint.’ The purpose of this prediction is to enable a preventative action - we can notify the banker in advance of taking the call and/or classify the interaction as a priority.
To mitigate the impact of missing a complaint, the initial machine learning model would be designed to overcompensate on the metrics that indicate a complaint, so that more data points are classified as complaints and bankers can then update the data with information. This enables reinforcement learning to a machine learning model that continually improves the accuracy. In addition, the model should have monitoring in production that provides real-time alerting on tolerances as we adjust or change the model. This is referred to as a feedback loop and provides quality assurance information to ensure that we are limiting the impacts.
The significance of this change means we move some of the human effort from classifying and triaging parts of a process to monitoring the production quality of the results of the classification engine. Our workflow design now more than ever should focus on the end-to-end customer flow, including quality assurance and toil removal previously allocated as a separate function and not part of the experience design. By enabling access to a model to train on production data points, we can move to a preventative in-built customer service and fulfilment model.
How is Generative AI (Gen AI) different?
Gen AI is not primarily used for classifying or making statistics-based decisions. It is used for creation. Creating new and unique content and scenarios where novel outputs are required; think images, music, code and stories. I like Will.I.Am’s definition of AI creation: ‘Imagination Regurgitation.’ Gen AI can reassemble information it has already been given in different formats to create something new, but it cannot replace humanity's ability to genuinely empathise, recognise right from wrong or create something completely new.
In workflow design, GenAI will be using information reviewed and signed off by our existing governance process to enable information to be surfaced in the way the customer wants to interact .The important thing here is that existing governance and compliance processes for content creation and code compliance are critical for reducing hallucinations (the ability for the generative artifact to present something that is not right). Again, this comes down to models and design prompts putting in the constraints where they are needed.
Consider a scenario where we previously delivered A and B testing to provide two different landing pages for two diverse types of customers. Gen AI provides the power to scale personalisation and to tailor experiences and interactions in the way each customer prefers.
Where this becomes new and exciting is the ability to orchestrate applications with Gen AI. While that might initially draw large breaths of risk-induced panic, remember the idea is to define the parts where we can be flexible and those where we cannot. This means if we have the unique components and capabilities, we can empower teams to create customised applications using predefined and compliant code modules, in turn making responsiveness to market, brand, and current affairs unique, simplistic and cost effective.
With Machine Learning and Gen AI used safely in the right spots to enable a positive experience, we can add flexibility into the overall workflow, populate data from the customer file with consent and reduce friction to the customer as they interact with us across the bank and beyond.
Towards modular and flexible customer interactions
Gen AI propels a required shift from cumbersome, monolithic workflows to agile, reusable services capable of being composed to suit varying customer interactions. These modular services minimise workflow complexity (see related article), optimise the user journey across multiple touchpoints, and adapt to customer personas — independent of organisational structure.
The benefits to those brave enough to let go of historical reasoning to segment technology is advancement of customer experiences that will change both the game and throughput of work. The movement of where the work lands is a change that requires all of us to remain on the pulse of customer needs by front running research, enabling new data components, new product partnerships and more diverse offerings. There is new work — and lots of it!
Key takeaways for design
Design with a focus on the customer persona, unrestrained by current technological or organisational limitations
Develop minute, discrete data components using cloud technology
Build in compliance and quality assurance
Enable collection of components to form discrete modular capability
Aggregate capability to form versatile workflows
Align immutable business rules to the right layer component, capability, or workflow
Employ Gen AI to augment flexibility; offering workflow personalisation and delivery of actionable insights through customisable interfaces
Key takeaways for compliance
The term ‘decision’ in AI means result, and it is up to us to ensure we have reinforcement learning and auditing built in the right places within the software delivery lifecycle to increase quality (already an existing requirement)
By taking the right-sized service approach, AI tools are not complicated to govern
Conclusion
As a pivotal mechanism guiding work movements within diverse structures, workflow design has always been shaped — and will continue to be shaped — by technological evolution.
It gains from a user-centric approach augmented by the potentialities of Gen AI, offering customisability and responsiveness to user requirements while harnessing the efficiency of small, reusable service modules that redefine customer interaction landscapes. The key to unlocking this potential is approaching tasks with a user-centric and curious mindset.
Adelle McDonald is the Assisted Servicing Technology Lead at ANZ. She specialises in technology transformation, complex distributed workflow design and engineering, problem solving, cloud solution architecture, and complex hybrid cloud implementations. Her most recent focus is on advanced generative AI.
Adelle has over 20 years of experience in the field of technology transformation and has worked for Goldman Sachs JBWere, McKinsey & Company, Dell, National Australia Bank (nab), and Insurance Australia Group (IAG).
She is an adaptive leader with a high degree of commercial acumen. To help people reach their next level and deliver customer-centric regenerative technology solutions, Adelle follows a four-step methodology: empathise, empower, execute, be efficient.
This article contains general information only – it does not take into account your personal needs, financial circumstances and objectives, it does not constitute any offer or inducement to acquire products and services or is not an endorsement of any products and services. Any opinions or views expressed in the article may not necessarily be the opinions or views of the ANZ Group, and to the maximum extent permitted by law, the ANZ Group makes no representation and gives no warranty as to the accuracy, currency or completeness of any information contained.
