Robert Stoecker 02.01.2022

Challenges of human-AI collaboration and how Forto addresses them

In a bid to improve efficiency and boost profitability, Forto made use of a machine learning algorithm to predict, forecast and support decision-making. Introducing such a complex and often distrusted technology into an organisation is, however, accompanied by multiple hurdles, such as human-AI collaboration. Robert Stoecker studied the challenges of human-AI collaboration at the London School of Economics (LSE) and here provides some insights about how Forto addresses these challenges.

Increased use of AI in prediction- and decision-making

Artificial intelligence (AI) is being increasingly used for predictions, to improve forecasting and aid decision-making in everything from medical diagnoses to recommendation systems suggesting which Netflix show you should watch. Machine learning (ML), as a subcategory of AI, is predominantly used in such prediction scenarios. ML is being introduced in organisations to improve, amongst others, efficiency, quality and accuracy in decision-making processes as well as optimising internal business processes and automating repetitive tasks, thus freeing up workers to focus their attention on more creative jobs.

Forto’s use of AI

The freight forwarding landscape is troubled by a truckload of problems, such as analogue processes, volatile prices, exception management, and stakeholder management, to just name a few. To address such levels of uncertainty, Forto’s data science team implemented a ML-based prediction and decision-making algorithm in the first half of 2021. The algorithm predicts which booked shipments are likely to experience a margin loss. More precisely, based on historic shipments (training data) the ML algorithm flags shipments that are likely to generate a credit note or a decreased actual margin against the planned margin. Both use cases empower the ML algorithm to support the human counterpart in decision-making, decrease repetitive manual work and improve exception management. In the long-run, the ML algorithm aims to discover operational process inefficiencies and identify, amongst others, customers, partners, and trade routes causing inefficiencies and profit loss. The underlying success factor of the ML algorithm is based on the data gathered from Forto’s multiple systems over the course of the past years. It is here where the work of Unsal Gokdag, a Senior Data Scientist at Forto, and the rest of the company’s data science team, to “cleanse” the data to ensure high-quality training data pays dividends.

Human-AI collaboration and the challenges this presents

Effectively using a complex information system, such as a ML algorithm, is accompanied by a multitude of problems. The most prominent challenge when implementing AI into an organisation is likely to be data-bias, as Amazon found out at its expense when it was discovered that its ML-based recruiting engine did not like women. Recent academic literature has also highlighted the not-insignificant challenges associated with human-AI collaboration. AI can perform human feats and thus is fundamentally different to other technology as it invades the previously exclusive human domain by shifting the locus of choice, power and control away from humans. In other words, AI is becoming good–and getting better all the time–at performing jobs once exclusively done by people. This is creating tension and raising fears that AI will replace human workers across the economy. This perception is fuelled by a fear of the unknown as complex ML algorithms are difficult to understand for the vast majority of people. This represents a serious challenge to human-AI collaboration.

Successful AI implementation goes hand-in-hand with explainable AI

To reap the benefits of AI over the long term, the goal of any organisation is not to replace human workers with AI, but to create a supportive collaboration between them. To do this, an AI algorithm’s output must simultaneously be trustworthy, accurate and explainable, and thus understandable by the people coming into contact with these outputs. The figure below illustrates how good use of AI is based on a mix of data science and human science to attain explainable (and responsible) AI. Figure: A schematic view of Explainable AI — Adadi & Berrada 2018

Measures that allow Forto to address challenges of human-AI collaboration

The outcome of studying Forto’s use of AI and its implementation focuses on human-AI collaboration. The study validates the importance of explainable AI, and also points out other vital measures that are needed for the successful implementation of an AI system:

1. Accuracy-Explainability tradeoff to build trust amongst employees
   The academic literature on AI tends to illustrate that accuracy and explainability are opposing characteristics. The logic behind this is that with increased explainability, an algorithm’s output loses accuracy. However, Forto’s case illustrates that in the practical world, accuracy must first be achieved to create a proof-of-concept and build trust. People are naturally hesitant to blindly trust complex algorithms. Elaborate testing of Forto’s ML algorithm to improve its hit rate (accuracy), ultimately led to trust and thus acceptance by Forto’s employees. It’s only then that the need for explainability drastically increases. The result of increased explainability is effective human-AI collaboration, which can then lead to inter-departmental procedural changes as human employees can better act upon the ML results.
2. Business and tech alignment
   Forto’s use of AI showed that alignment between the tech side, in this case the data science team, and the business side of the organisation is vital. Forto has mastered this through first, clearly defining conjoint use-cases during the creation and subsequent implementation of the algorithm, and secondly, via close interdepartmental feedback loops once the algorithm was applied
3. Provide an environment which nurtures human-AI collaboration
   Lastly, studying Forto’s ML algorithm illustrated that implementing a complex information system, such as a ML algorithm, is not a one-time event. Organisations should provide a suitable environment that fosters human-AI collaboration. Forto was, and still is, patient enough to allow human-AI collaboration to develop with recurring interactions that evolve over time

This example of an academic study of Forto’s ML algorithm illustrates how AI can be more widely applied in freight forwarding companies. It also shows that complex technology must be met by open-minded employees willing to interact with an inherently complex algorithm. The use cases of AI in the logistics industry are multifaceted and when used appropriately, offer the benefits of greater efficiencies and higher profits, among others. However due care and attention must be given to the employees who will fall under AI’s sphere of influence before, during, and post-implementation.

If you are looking for a challenge in the new year, check out Forto’s career page.

Unsal Gokdag, a Senior Data Scientist at Forto, gathered and cleansed historical shipment-based data to feed and train a ML algorithm to make accurate credit note and margin loss predictions. Cleansing the data to use as training data for the ML algorithm is vital, otherwise no valuable results would emerge– garbage in, garbage out (GIGO).

Robert Stoecker was a Senior Business Operations Manager at Forto until late summer 2020 when went on to study for a MSc Mgmt of Information Systems at The London School of Economics (LSE). For his MSc dissertation, Robert and the LSE collaborated with Forto to study the challenges of the ML algorithm, its implementation, as well as the explainability versus accuracy trade-off.