Types of AI

The consulting firm McKinsey reported in a recent survey that nearly eighty percent of companies are using gen AI – yet a similar number of organisations report no significant increase in profitability. This is the ‘AI paradox’, which is similar to the ‘Productivity paradox’ identified by the economist Robert Solowin in 1987, when the hype about computing was at its height – history repeats!

AI is the ‘catch all’ term for a group of technologies and tools, currently identified under two types of AI:

Horizontal AI used across all functions within a business. Consisting of chatbots (where people converse with a machine or network) and personal assistants that assist individuals with tasks. For both, productivity improvements are difficult to measure.

In April this year, a research paper was released concerning the Impact of ChatGPT based Chatbots. The research covered 25,000 people across 7.000 workplaces located in Denmark. The conclusion from the research was that “…despite substantial investments, economic impacts remain minimal. Productivity gains (measured as time saved) were about three percent on average, so AI chatbots had ‘no significant impact’ on earnings or working hours in any occupation”, which challenges the forecasts of imminent labour market transformation due to generative AI.

An example of AI as a personal assistant occurred recently when a colleague used genAI to develop a PowerPoint presentation, then to format a report and finally to design a four day holiday that follows attendance at a conference. These actions improved his personal productivity, but improvement to corporate productivity could not be measured.

Vertical AI is a function specific use of AI. The McKinsey report notes that fewer than 10 percent of ‘proof of concept’ trials of vertical AI proceed further and then only to support isolated steps within a business process. For example, in Procurement, the most cited use is reported as risk analysis of suppliers.

A choice to solve a problem

Are the current AI types a ‘solution in search of a problem’? To internally build and implement AI applications contains potentially substantial but unknown risks. It requires an investment in knowledgeable people and time, using tool sets that are evolving and ‘one-off’ project documentation and training that is rarely comprehensive. Then, it is an operational manger that must give the ‘go ahead’ and take responsibility for subsequent failures.

As most processes in supply chains are not in a category where AI can (or should) automate all the steps, managers of organisations and supply chains have two options:

• Understand the available toolset and ways to obtain the best return from an implementation. Then invest in a pilot, which may (or may not) lead to a successful full implementation, or
• Learn about the technology and tools, but wait until technologies are developed that best suit the needs and capabilities of your business and supply chains

Considering the second option, a development that can be of interest for supply chain professionals is Agentic AI. According to the publicity this can ‘interpret goals, make decisions and act across workflows’. The term ‘agentic’ refers to the capacity to act independently and with purpose. It uses large language models (LLMs) and machine learning (ML) to autonomously complete complex tasks. Agentic AI can consist of multiple ‘agents’ that perform specific tasks and connect to external tools, which are then co-ordinated to provide a range of responses about demand, supply and inventory.

An Agentic AI trial

Rather than develop a complete ‘solution’, an alternative approach is to initially identify where Agentic AI can add value. This was the approach of a recent trial to improve the processing of Bills of Lading (BL), which currently takes many days within the cargo documentation process. The current BL process is based on emails sent between relevant parties as BLs are printed, couriered, scanned and signed. The requirement was to speed the process and reduce errors.

The trial was a live process. On one day, every export shipment controlled by a forwarder on the Shanghai to Los Angeles route was issued with a digital Bill of Lading – no paper documents or training. Document turnaround time was reduced from 5–7 days to under 10 minutes, with a substantial reduction in costs.

There were no software packs to install and no websites for staff to log into. Employees in the businesses use emails as part of their job, so they were not asked to learn and use new software or technology, nor to make any changes to ERP systems or processes.

Agentic AI agents operated in the background, reading natural language instructions from the emails in both English and Chinese, such as ‘please issue BL’ or ‘transfer to consignee’. These automatically triggered the appropriate actions to issue legally compliant cryptographically secure eBLs. Agentic AI agents also extracted shipment details like container numbers, vessel ID, and consignee information from attached documents and populated the eBL. A compliance co-pilot reviewed each eBL to ensure that parties were compliant with trade regulations and audit ready.

This approach was enabled by the Model Context Protocol (MCP), a framework that standardises how software applications provide context to Large Language Models (LLMs). It enables AI systems to more effectively handle real-time interactions, such as within a transport management system.

Supply Chains using Agentic AI

Although the trial was a success, all the current risks associated with AI still apply, but are magnified when multiple autonomous agents are linked. The consulting firm McKinsey notes “…systems are no longer organized around screens and forms, but around machine-readable interfaces, autonomous workflows, and agent-led decision flows”, which provide opportunities for a system to ‘do its own thing’.

To gain the full potential from Agentic AI in supply chains will require a change in design thinking for the entire process. Agents become the core of a design, with people and Agentic AI viewed as co-workers, that enable custom-built and commercial agents to inter-operate.

The design requirements, operating needs and potential risks of Agentic AI provide opportunities for software companies to design and sell supply chain applications. These would be demonstrated, installed and maintained by IT professionals, so that much of the design and implementation risk would be held by the developer and supplier.

However, the big challenge for Agentic AI in supply chains will not be technical, but people. So, the critical role for supply chain managers and professionals over the next five years is to understand and build trust in the technology, then learn and train with staff to develop the governance protocols around agent autonomy. As the systems interact across supply chain functions, so the organisation structure will change from traditional business units to a Supply Chains group platform.