There are many lessons being learned on a daily basis in use of AI and successfully implementing digital transformation projects. These lessons are a good thing like any other lesson in life, as every lesson makes us better. The near-term advances in AI are on the agentic front that is rapidly moving towards multi-agents, physical AI, Multi-modal AI and the full-blown AI factories. There are many decisions to make every step of the way on these programs. Teams are rapidly figuring out that domain knowledge is critical, human oversight is critical to train the models correctly and the selection of use cases for AI are critical. So where should a team start? Its no different than how we start life. We crawl, we walk, we run and then we learn how to sprint or how to run a marathon depending on the goal. However, we don’t have years to accomplish this like we do in real life to stay ahead of the competition.
Let’s take an example of supply chain agents that will be intelligent enough where a group of agents can work like a team. To train the models, plenty of domain knowledge will be required. As an example, to keep the production line going, agents will need to have the domain knowledge to track inventory, which in real life resides at multiple places. Parts reside in Manufacturing bins, in customer support warehouses, in quality control cribs and so on and so forth. A shortage can impact any of these business functions adversely. The human in this process today makes intelligent decisions such as when to pull a required part from manufacturing bin to support a customer or vice versa to avoid shortage on the production line or work with the quality engineers to free up good parts from the quality control cribs. In doing so, the human is talking to his/her counterparts in customer support, supply chain and quality to make the correct decision. All this intelligence now has to be infused into the agent. You can see how quickly this gets complicated and hence the crawl-walk-run analogy.
The human domain expert will need to identify use cases that are simpler in nature first to train the agent, for example, train the agent to look at all relevant areas to track the inventory. The next step is to train the agent to ask the right questions about available inventory. The
next step after that is to follow the process to transfer the required quantity to the production line and keep building these various steps one at a time. Let’s see now what could go wrong if the intelligence infused is not fully developed. An agent can randomly transfer the parts to
the production line from the customer warehouse causing a short fall, where now a customer in need of the same part must wait, causing customer dissatisfaction. The checks and balances will need to have oversight of the human expert until the intelligence is fully developed and matured before the next case gets looked at.
A key assumption made here is that the process to be followed is robust and will work the same each time. Is this really true though? Or have we as humans contaminated them to satisfy the need of the hour over time? The answer almost always is that the processes are contaminated and any kind of contamination in the intelligence infused into the agents could have a catastrophic possibility, depending on the severity of contamination.
Let us take another example of supply chain. Agents in Team A place a Purchase Order (PO) less than or equal to $250K so they learn to crawl. The agents before this step in Team B, raise the Purchase Request (PR) and they were newly introduced in the system. If the process for them is not fully vetted, it can result in three scenarios.
Scenario-1: The PR agents provide the required PR to the PO agents when they should not have, and the PO agents have issued PO’s erroneously, causing cost to the business and perhaps hardship to the supplier with no raw material ordered for the parts.
Scenario-2: The new agents have a typo in the decimal point, and no POs are getting issued due to the restrictive guard rail of $250K.
Scenario-3: The agents over time have become very smart or have been hacked and they have now figured out how to circumvent the guardrails and place PO’s that did not need to be issued. Unless the business and IT have rigorous processes for testing and vetting and periodic audits show that the process is not broken, the AI agents can cause damage. Imagine how rigorous the process must be for scenario-3, as it may just be one agent that has gone rogue and causing the damage. Unless the process is set to figure out who that agent is, mitigating the risks can take a long time.
These simple examples illustrate how disciplined the program teams will have to be before they can reap the benefits of AI and how smart the Program Manager will have to be, to thread the various steps of the transformation so progress can be made fast. Thus, the secret of the steroid AI, is to have a human in the loop and disciplined processes to start running.