From Risks to Results: Adobe’s Practical Lessons for KMers Working with AI

Whether you are a taxonomist or any other kind of KMer, there was so much to learn, and enjoy, from Rachael Maddison and Connor Cantrell'ssession in KM World conference 2025.

Their session walked us through the real journey of building LLM-based taxonomies, the wins, the challenges, the risks, and the practical decisions along the way.

Here are a few messages I believe all of us can take forward when developing LLMs or AI agents:

1. Manage risks before you start

Think about your data, the model you choose, the training approach, security considerations, and, most importantly, what success actually looks like.

2. Expect, and look for, three types of mistakes

• Inconsistent results

• Inaccurate results

• Unintended results

 Spotting these early can save enormous effort later.

3. Use simple but powerful prompting practices

• Be ready to iterate, almost any prompt can be improved

• Use reverse prompting to help shape better inputs

• Direct the model toward specific content to stay within scope and reduce drift.

It may sound like a demanding journey, but done right, the benefits are substantial. In their case study, manual effort dropped by nearly 90% 

And yet, one question stays with me:

Does AI elevate KMers, or does it mainly accelerate KM tasks, reducing the human effort needed in the long run?

A conversation worth continuing.