Artificial Intelligence (AI) is creeping into everyday life, both visibly in the form of physical robots and invisibly within commercial processes. It is a very nascent technology and unlikely to be used pervasively by your business today, but at the same time it is worth spending some time thinking through how and where it may impact your operations, as well as considering the emerging issues that this technology brings with it.
Why is AI Important
At a macro level it is seen as the new oil for the benefits that it can bring holistically to all humans, for example in medical science, as well as driving a significant competitive advantage when the same technologies are applied to business, recognizing that these systems will move increasingly towards real time.
Different Terminologies
Typically, you will see the following words prefixed to AI that point ostensibly to the same thing: top down & bottom up, weak & narrow, strong & gentle and neural & deep neural. Underlying complexities associated with each one can be very different, but at the end of the day the math associated with each algorithm is by definition exacting, and applications that contain it are laser focused as a result.
Rules cf Probabilistic based Algorithms
Businesses typically up to this point have utilised rules based formulae, also called algorithms that have been defined by programmers to achieve a task. However, generally rules are complex to write, not scalable, and it is very easy to introduce false positives ie results that seemingly follow all the logic, but in fact do not deliver the required or intended end result. Probabilistic algorithms on the other hand are more flexible, easier to develop and scale, noting that in any deployment a combination of rules and probabilistic formulae will be used concurrently, and/or sequentially, to drive the best outcomes possible.
Choosing Algorithmic Formulae Components
Multiple data points always exist, but not all of them will directly influence your desired outcome, so the real challenge is to figure out which points impact the desired and intended result. This is easier said than done and might require extensive use of trial + error techniques, or even data scientists which are in short supply, recognising that each data reference point added by you can dramatically scale data processing requirements and operational running costs.
Black Box cf Transparency
Corporates are having to think deep on digital ethics. As a result company values relating to AI calculation modus operandi, and underlying accountability will all increasingly go towards corporate differentiation. Not only that, but processes will need to be thought through end to end to ensure that any required PII (personally identifiable information) data is managed according to within country laws, including the broader GDPR, US Cloud Act, and legislation from other countries that drive the same underlying intent. As always think through the logical, legal, practical and if applicable digital tax aspects of data location and keep an eye on the evolving Algorithmic Accountability Act and broadening discussions in the US and Europe re data portability.
Unconscious bias as part of these discussions is very topical these days as all of us, especially programmers, can unknowingly introduce it. So what you have is a dichotomy. On one side the driving of competitive advantage leveraging your “secret sauce”, whilst on the other being transparent enough as to how an algorithmic conclusion has been reached ie processing is not within a black box.
Ultimately how bias is handled will be down to the ultimate intent of the algorithm, whether PII data is involved, as well as industry rules and regulations. As can be imagined and seen in the press there are business focused activists and multiple stakeholders who will call a corporate out on bad algorithmic outputs.
Algorithm Management
Are they load n’ go and thereafter forgotten, which is typically how they are generally thought of, i.e. static, or are they managed ? Algorithms can be run in real time or re-calibrated as needed ie hourly, weekly, monthly, quarterly etc depending on the objective and user case.
Often underestimated or missed is that in many cases the algorithm needs detailed handling from the perspective of the process owner, and additionally for rules to be in place for accountability and oversight management. Simplistically, this activity simply changes the “shape” of the data capture area to get a better outcome using a variety of appropriate statistical methodologies (but which one is the real underlying question). A good user interface for the sub-process will be required to achieve this and this needs to be part of any development and deployment process
This needs to be contrasted with “off the shelf” pre-trained models from vendors eg for document recognition. These can be relatively quickly added and integrated to a process, but again think through the design of how, and where, to add and use them.
Algorithmic Results - Getting to 100% Perfection or Close.
This depends on your needs, but typically the 90:10 rule with regards to develop “time” applies. As a very general comment 80% perfection should be “relatively easy” to achieve, but after that it gets much harder very quickly.
Which 20% of data needs checking? Augmentation of a role / function should typically be seen as an end result of AI not necessarily 100% automation, and whether staff are re-deployed as a result of its use depends on how many are involved in the process at any one time, and what it actually entails to figure out those 20% cases that do not fit.
Full time Employees (FTE) Pre & Post AI Process
As with any automation, corporations typically do not currently have everything defined in terms of FTE equivalents and costs thereof, nor do they have in place or measure “success KPI’s” required for a specific function per time period, unless these are associated with volume based larger departments or BPO justifications, as might be found in financial services.
Why? This is partly system related in that this information is not necessarily easy to get hold of, nor is the mindset of having to manage processes at this level of depth. It is really down to a changing mindsight on how to drive it, which partially explains cost variances for the same function within the same industry for running the equivalent process.
Placement of Algorithms
How will this fit into your current system design? Modern systems today enable a process to be fully defined and integrated end to end. This is from data collection, through all ultra-granular transformation flows that go x-application, x-ecosystems to actionable contextual reporting + workflows @anytime @anywhere within the process + simulations + latest payment tech @anywhere.
Algorithmic Digital Enablement cf Legacy
Very simply it is the ability to extend processes with API’s within the process design mentioned above to leverage any of your applications (with compliance controls) or value added 3rd party service providers who may also use AI (banks, treasury for payroll, employee checks, career management services etc) for purposes of value creation. As a result it is important to figure out where PII data is being leveraged and whether there are any implications on privacy and data storage location.
Do you need AI?
No, not necessarily. Most corporates would benefit today by working to eliminate their workflow bottlenecks thru leveraging the auditable repeatable process tools outlined above. In many cases simply having access to these flows provides immediate value by being able to leverage standard processes across all regional entities for more timely accurate information, whilst getting ready for AI deployments.
Where Should you Use AI?
It is critical to re-emphasize that you need to understand what you are trying to achieve or put another way, “how can you define an algorithm without a detailed knowledge and understanding of the process itself”.
AI should not be seen as a generic solution but a means by which existing processes can be made smarter with all the benefits that are implied from this ie making them faster, more intelligent, scalable and interconnected etc. For many it will be a solution looking for a problem, but for some the building blocks for full or partial automation with or without AI are already in place to make changes that will lead to value creation. A game changer!
Ashley Clarke, COO, FlexSystem Ltd
Comentarios