The Adult in the Room Approach to Adopting AI

In my work with businesses struggling to adopt AI into their day-to-day work, I see two different strategies and some efforts in between. One extreme is to rush ahead with abandon and embrace whatever is coming their way in terms of models, harnesses, and buzzwords of the day. The other is slow, deliberate, pondering and very risk-averse. There is a middle ground that can be the answer for a majority of businesses. That middle ground consists of understanding the impact and the risk of AI. Initiatives, as trade-offs are made, decisions are taken and outcomes are tallied. Nate B. Jones compares this to riding a bike. You can’t ride a bike super slow because that will not give you the momentum that you need to be stable and upright. As you ride you have to gather some speed to be stable. How fast you want to go after that is dictated by what you really want from AI. As models evolve they become more capable, more refined in their approach, and more able to handle generalized instruction. But there’s also the fact that these models are becoming more of commodity offerings, and the user has a choice to optimize their work with them in different ways to gain maximum leverage without losing key functionality. Functionality the way that the software industry has enshrined repeatability. Not quite adaptable to how AI models operate. This can lead user confusion with the results obtained on each turn, and ultimately lack of trust. Business leaders often dismiss AI results as unreliable and untrustworthy. But the acceleration provided by AI is too tempting to ignore. The root of this dichotomy lies in the expectations. Unlike most conventional software, AI is non-deterministic. You cannot reasonably expect the same answer from asking the same question twice. Then where is the solution to this?

The other issue with breakneck AI adoption is that it is a capability multiplier. It will magnify efficiencies but it will also magnify bottlenecks. The AI tries its best to work around the bottlenecks to get you the goal that you seek, and makes up information to get there quicker. Depending on the kind of information being handled, the consequences can be disastrous or mildly irritating. Some simple modifications like using hooks instead of prompt where money is involved gives you a degree of fidelity, but LLMs will force you to look at the processes in a deeper way than you anticipated. Businesses should be ready to re-examine how they approach their work when adding AI to the mix. This is the hard part. But should you wait till every process is hand optimized to the nth degree? Not at all. Use AI to inspect processes to find points of optimization and dependencies that affect actual business outcomes, then have human experts and users find holes in those findings. Correct the AI with those findings and take another pass. This is what I call process capture.

Let’s take a real example. An industrial organization navigates the quoting process for engineering projects solely through email. Everything is in email including very large engineering drawings, pricing documents, contracts, contractor communications… everything. Finding the right email about the right project takes forever and a half, with very inconsistent and incomplete results. The company cannot function fast enough to avail of opportunities when you have to dig through tons of email going back years to find the details of past performance that they could include in a new proposal. All this information sits in individual mailboxes that make it difficult to extract and make available to the team as needed. There is no CRM and no common knowledgebase. And God forbid, the employee that manages said mailbox took a break. The entire sales operation can come to a grinding halt.

This is a two-part problem: data management and process control. Both of these parts can be dealt with using AI. Let’s break down the solution approach.

In this multi-part series I’ll dig into some approaches to AI adoption for small to medium businesses. So strap in and come with me as we dig into this.

• Step 1: Access and Retrieval
• Step 2: Triage
• Step 3: Tagging
• Step 4: Validation
• Step 5: Organization
• Step 6: Adoption
• Step 7: Evolution