The development of general-purpose technologies which enable AI to deliver better results have been strengthened.
This is due to advances in communications networks, dedicated AI silicon, open-source and dedicated algorithms, and larger storage capacities.
The AI educated talent pool addressing the business need is developing fast, starting at the top by strengthening the Board and propagating right through to the teams implementing and using the technology.
AI is on the Board agenda in many organisations already, others are rapidly catching up.
Prior to implementing AI in their business, the Board need to weigh up the impact and the benefits. We addressed ethical principles in our white paper An ethical framework for the implementation of AI in business process, and following this, we felt that a real world example of a problem solved by AI may help lead the Board through what may appear to be a confusing landscape.
Here we use a specialised electronics hardware manufacturer as an example*.
Their route to market is complex and indirect, they struggle to create accurate forecasts due to uncertainty, last minute movements in projects, lack of access to the decision makers in the entire supply chain and generally external influences outside its control such as weather conditions in the installing country and natural disasters where components are manufactured.
Here we apply good practice principles when assessing a situation leading to a decision based on the best view of the facts.
The Board may ask the following questions of each other and their teams.
1. Need. What problem has been identified that cannot be solved otherwise?
2. Finance. Is the implementation financially viable?
3. Test. How would you test the implementation to ensure that it meets the organisation’s business requirements?
4. Alignment. Does the implementation align with the organisation’s broader strategy?
5. Analyse. What are the risks involved?
What problem has been identified that cannot be solved otherwise?
The company strategy is to deliver the right products to its direct customers who are installers, just in time to enable them to meet final fit schedules and maintenance updates. It currently meets demand by carrying excess stock which results in high working capital utilisation or not carrying the right stock resulting in lost sales. This is not economically viable for the company in the long term and therefore a solution was sought by the Board to increase forecast accuracy through better predictions.
Forecast preparation was a manual task based on historical and future sales predictions. The labour cost of the process was estimated to be around 5 man days a month and its inaccuracy regularly generated up to 6 months of excess inventory at a cost running into seven figures.
The datasets utilised for the forecasting process were both structured and unstructured and were derived from multiple sources:
1. Previous manufacturing schedules were used to create future demand trend projections taking into account seasonal variances.
2. Electronics components stocks and supply chain lead times used to generate trend projections which trigger the purchase of components.
3. Future sales projections based on external interactions with direct customers.
4. The excess/under stock issues were magnified when new products were introduced because there was no historical data on which to base a projection.
5. Before the forecast was actioned, it required a detailed review by senior executives in the company with a broad knowledge of the market and economic conditions affecting it. Sensitivities were then applied to the model.
The company needed to automate the current workflow and create a weighting per product line based on customer and end user sector, which took into account outside factors.
By implementing a combination of AI techniques including supervised and unsupervised learning, reinforcement learning, and deep learning the identified problems were solved. The resulting high level of forecast accuracy enabled the company to meet its availability and on time delivery KPIs.
Supervised learning was applied to the structured data, applying a non-linear regression model such as Lasso or Ridge in order to create the un-sensitised forecast.
A recurrent neural network using natural language processing analysed PR reports and published economic reports in order to determine the effect of prevailing economic conditions and movement in the marketplace or sector. A weighting factor was generated through regression and applied to the forecast automatically.
Reinforcement learning was applied in order to achieve the optimal balance between working capital utilisation and availability of stock over time.
Is the implementation financially viable?
The implementation required two full time data scientists for twelve weeks. There was no disruption to the current process during development, test and implementation as it was planned that old and new models were run along side each other for a period of twelve weeks.
The effect of the AI was highly measurable, initially seen in improvements in stock availability and working capital utilisation after sixteen weeks. Full ROI was realised over a period of around six months, seen as a reduction in inventory to one month and the release of around hundreds of thousands of working capital. Excess stock write off was also reduced resulting in lower stock provisions and greater retention of profit at the financial year end.
How would you test the implementation to ensure that it meets the organisation’s business requirements?
The AI was tested alongside the current system and unusual variations analysed and adjustments made. Unsupervised learning and clustering were used to automatically identify data points which should not be there. A final review by humans involved a gross error check rather than a detailed review. The human impact reduced the manual processing required to create the forecast enabling a focus on high quality customer interactions resulting in demand for more products.
Does the implementation align with the organisation’s broader strategy?
Using AI to improve forecasting was aligned with the company strategy to provide products just in time to the customer and to keep inventory stocks and working capital utilisation to the absolute minimum.
What are the risks involved in implementing AI?
There was a risk that the AI may predict a far higher or lower demand than required. The AI testing model highlighted significant variations prior to financial commitments being made. There were no identified legal risks as the AI improves current process. The ethical consideration towards displaced labour was addressed through training of existing people to effectively interact with customers and users. There was a greater risk associated with doing nothing as stock and working capital utilisation would have continued to increase without a solution.
*the scenario is based on a real company problem. The company is not mentioned here for confidentiality reasons. The solution to the problem is hypothetical.
For further information please contact us at firstname.lastname@example.org
This article is original content which was created on 3rd December 2019 first published online 18th March 2020. It is Copyright 2020 Anekanta Ltd.