Demand Agents: Human Assisted AI

There has been a definite shift in the AI narrative over the past six months. Until last summer, companies weren't willing to consider where AI could replace human employees. That changed when tech firms realised that job cuts were the only way to generate a meaningful ROI on their spiralling data centre investments. Yet, in the months since, few companies have actively replaced human employees with AI, despite numerous announcements from CEOs to the contrary. Despite the hesitation, AI adoption is still a top priority in many C-suites.

Level of Accuracy

Companies planning AI projects need to learn the limitations of the technology as well as the benefits. A lot of the debate around AI focuses on hallucinations: the tendency of LLMs to make stuff up. This has long frustrated AI developers and is indeed a bad habit most of the time, but not always. There are plenty of research or creative scenarios where you want LLMs to say new things. There are even more scenarios where 90% accuracy is good enough. For example, in content generation use cases, where the AI writes a first draft for a human specialist to edit.

When identifying use cases for AI, it's important to identify where absolute accuracy is required and where it isn't. After all, the technology can do many things. It can generate marketing plans, draft content and analyse campaign results. It can identify the best leads and send a personalised email to follow them up. It can handle routine customer service queries in real time. That doesn't mean it can or should do everything. There is still an important role for humans in the AI workplace.

Human Support

The most relevant concern is exceptions. How do you handle the stuff that AI can't answer? It's all very setting up a human in the loop procedure to check AI output. However, you then need escalation processes that allow a human to take over. This is seen most clearly in customer support scenarios. Customers often want the bot to handle routine contractual queries or common troubleshooting. It's normally quicker than human live chat and less stressful than your typical offshore call centre. However, brand reputation generally lives or dies based on the quality of the escalation process.

It doesn't matter how helpful the AI customer service agent is. Customers still need the ability to escalate complex scenarios to a human support rep, or they won't bother dealing with a particular brand at all. Good customer service is all about solving the one-off problems that no one has seen before while keeping the customer happy. Yet, many customer success teams focus on the everyday situations, and forget about the occasional crisis. The ambition should be to have the best human support team possible, and then using AI to avoid over-burdening them. If recent tweets are anything to go by, even AI pioneers such as Klarna have learned that lesson.

AI Assistant

From a business perspective, the most important question is when can AI usefully support human employees. Can it handle the routine administrative tasks that human workers don't want to handle? It already does in many areas. Can it surface the business-specific insights to enable data driven decision making? It's been doing that for years already. Can it overcome blank paper syndrome by providing the starting point for new ideas? It does have access to most of the internet.

AI is increasingly delivering value, but more as an assistant than as a co-worker. There is nothing wrong with that, even if tech firms do have higher ambitions. At the macro-economic level, productivity has been stagnant for over a decade. Incremental changes to internal technology may have substantial benefits, but only for a few individuals. Even in the current state, AI can act as a research assistant and as an editor. That in itself is a substantial time saver.

AI is becoming increasingly capable, but it's still not a mature technology. In the rush to automate, do not forget the human element. The goal of an automation project is to deliver a better or more efficient output, but not for its own sake. The end goal must always reduce workload and benefit the customer. That does require constant learning. Yet, just like a successful human, AI models are always learning. Unfortunately, it's very difficult to determine what they are learning. Data scientists are still trying to solve that particular problem.