Dazed and Confused by AI in Your Business?

Read this article on LinkedIn.
Don’t let AI daze and confuse you like the iconic Zeppelin track. Particularly for non-technical business leaders, AI remains vexing, fraught with more confusion than clarity. On the one hand, implementing AI can deliver impressive results, on the other hand, we need to treat it like any other major tech implementation. Apply sound business judgment to navigate through the haze. Here are 8 principles to follow:
1. Don't believe in miracle cures. AI can do a lot for your organization, but many vendors overpromise and underdeliver. For example “zero human agentic coding factories” is not viable for most enterprise production applications. Bear in mind MIT’s Project NANDA finding that 95% of pilots never added real results. AI has engendered a tremendous amount of magical thinking about what’s possible. It’s not all smoke and mirrors, but the chasm between demo and enterprise-wide adoption is vast. Crossing it requires discipline and commitment.
2. Follow the money. Focus AI efforts on the things that will create the most value for your business. AI can be used to make a key competitive strength stronger, such as Aaron’s experience at AWS deploying agents into workload migration and automation to accelerate growth and improve customer experience. Or AI can be used to shore up a critical weakness, such as Morgan Stanley’s Wealth Management making over a hundred thousand pieces of proprietary research accessible to its advisor teams. Pick the use cases that will move the needle. Automating a spreadsheet in finance might be helpful, but it’s not going to redefine your business. Here’s a useful framework for deciding where your time is best spent.
3. Set goals, measure outcomes. Introducing anything new into your business needs to be done systematically. Without a robust process, you create movement without momentum. For example, some customer support bots have gotten worse with the shift to agentic agents. In the old rules-based regime, chatbots performed adequately for most customer requests while today’s generative models can fail at the most basic tasks. Recently, I found myself in a catch-22 with FedEx on a tariff dispute: the support chatbot sent me to two different department chatbots with each respective bot sending me to the other for redress. Don’t spin up an AI project without a systematic way to measure if it’s actually an improvement on what you were doing before.
4. Look both ways before crossing the street. Deploy with caution. The purpose of AI deployment is to scale your expertise. Make sure your expertise (documentation, formal and informal proprietary knowledge) is protected and that your agent doesn’t unwittingly end up giving it away to your competitors. If the largest scale AI players like Anthropic are deeply concerned that other AI players are distilling the expertise in their models through undesired access and “distillation,” shouldn't you be? More on how to protect yourself in this article.
5. Check the homework. Don't rely 100% on machines to audit agentic output. In the pilot stage, manually review samples of the prompt response logs and engage in regular hybrid testing once the model is deployed. Simple user feedback loops can help here as well. Do not deploy unsupervised agents without feedback loops. You would not have employees unsupervised, don’t think agents are any different. Recently, while driving through Utah to the Las Vegas airport, Google Maps pulled up directions to the Las Vegas airport in New Mexico, over 600 miles away, and a much less likely destination for anyone driving in the region. This oversight demonstrates that errors that escape pure machine reviews can be significant.
6. Partner, but don’t get stuck in a bad relationship. Every organization will need to build an AI supply chain that includes other parties. As you engage, intentionally plan for the flexibility to alter your tech stack and vendor base as needed. Explicitly ask any major partner upfront where and how the ability to change things, including with them, is built in. Nothing is more permanent than a tech-debt-laden deployment with no clear off-ramp.
7. Commit for the long haul. Any agentic system will need time to adapt and learn continuously in order to truly be useful to your business. You need to plan for continued maintenance and support, whether internal or external. Like most enterprise technical implementations, 90%+ of the present value of total cost of ownership is beyond initial deployment. Apple’s Siri was introduced in 2011 to much fanfare, but generally poor reviews. 15 years later they’re still improving it.
8. Empower your people. Your people are the backbone of your organization. The fear of being replaced by AI is palpable. Empower your people by giving them the tools and skills to thrive with AI. Communicate a future state that involves multiplying impact, productivity and professional growth for those that embrace the technology. While messages of doom and gloom are persistent, the more likely future is an AI-assisted workforce rather than an AI-only one. Rather than pre-empting the latter, structure your organization to thrive in the former.