Implementing an AI system can be time-consuming, expensive, and complicated. Fortunately, in virtually all cases, if well thought-out and executed properly, these implementations yield massive ROI. However, it’s important for businesses to understand that AI systems aren’t plug and play implementations. AI systems require extensive testing, training, and a bit of trial-and-error.

Think of an AI system as a fleet of workers. You wouldn’t hire a group of people and immediately ask them to serve customers. You would train them first. There would be initial mistakes. Some workers would be let go or reassigned. Others would rise to managerial positions. Implementing an AI system is no different. As a result, there is some risk involved.

For example: While you’re training an AI system to perform a task, you’re not yet generating ROI. If a process repeatedly breaks, the final implementation is delayed, and no business value is being created. This results in lost time, costs, and labor. At IPsoft, we work with our clients to minimize the risk associated with AI system implementation.

Training an AI System Before Implementation is Crucial to Mitigating Risk

In order to train Amelia, IPsoft’s cognitive AI system, we work with our clients to enrich Amelia with data, such as a company’s business vocabulary, its employee handbook, and historical customer support data. AI systems are very good at absorbing this information quickly and being able to put it into practice almost immediately.

Here’s where things get slightly more complicated. We also train her to understand our client’s industry-specific semantic network, and any logic framework they need Amelia to follow in her decision-making. We upload a company’s pre-trained classifiers and analytic modules that will help her conduct business according to industry regulations. At this point, Amelia is armed with a company’s data and a strategy for how she can use that data to perform business processes and/or transactions.

As with any complex technology, implementing an AI system is a multi-faceted process.

From there, teams will work to auto-annotate Amelia’s conversational framework so she can understand utterances and common phrases – essentially, the teams help Amelia speak as humans do. This is where trial and error becomes massively important to AI system implementation. For example: What happens when a customer says, “In a few days…?” There are several ways of interpreting a phrase like this (Two days? Three days? A week?), so the teams will train Amelia to read that phrase as the business requires. After a few weeks of tweaks and fixes, Amelia will speak in comprehensible, natural-sounding phrases, including the use of idioms, humor, and in some cases, even sarcasm, if that’s what a company wants.

Once Amelia is ready to begin handling queries on her own, she will inevitably get stumped by unique or custom questions and issues. If Amelia is unable to resolve an issue, or finalize a transaction, she will escalate the conversation to a human agent. As this happens, she will stay connected with human employees to determine how the issue was resolved. She will then learn from her human colleagues, applying that knowledge to any similar conversations so as to avoid unnecessarily escalating future interactions. There is no standard for how long it will take Amelia to master these specific escalations. In fact, some escalations can never be handled by any AI system because of industry regulation. This is why it’s crucial that businesses have a very clear understanding of the use cases they’d like to tackle, and then set expectations to remain within that scope; if you begin an implementation expecting to accomplish password resets, and then two weeks later you decide you want to train an AI system to offer financial advice, you can do so, but you run the risk of prolonging the implementation.

How to De-Risk Your AI System

We have successfully trained Amelia in only few weeks for certain use cases, typically ones that can be applied across an entire customer or internal employee base. More complicated use cases require more training time. After the first use case has been implemented, we encourage our clients to train Amelia themselves for secondary and tertiary use cases, but that’s only after our clients have seen exactly how to get Amelia up and running on their own.

As with any complex technology, implementing an AI system is a multi-faceted process. As a result, there is some risk involved, particularly if you have strict time or budgetary restraints. However, once your system is up and running, and you’re familiar with how to make modest changes on-the-fly, the risk is reduced and the ROI is plentiful.