View RALI news and insights to keep up to date with the latest on trend developments relating to future leadership capability and experience requirements and the future world of work.
About 35% of current jobs in the UK are at high risk of computerisation over the following 20 years, according to a study by researchers at Oxford University and Deloitte. Go to http://www.bbc.co.uk/news/technology-34066941 and type your job title into the search box below to find out the likelihood that it could be automated within the …
With $372 billion to his name, Amazon founder Jeff Bezos says Americans have been taught to view billionaire success in the wrong light.
The announcement of a new AI model or feature fills me with a mixture of anticipation and dread. I’m excited to see what “new thing” I might be able to do, but I also know that some of my existing processes and systems might break as a result.
Every AI tool comes with the disclaimer that it can make mistakes. You might set up a workflow that is perfectly fine one day and perfectly awful the next.
It’s an ongoing challenge. Unlike traditional software, once an AI system is up and running, it’s not set-it-and-forget-it. Even a system that seems to be working fine needs an occasional check-in or fine-tuning.
The predictable AI failures
AI can stop working in some very predictable ways. Maybe it’s connected to an external tool (like your email), and the connection expired. Some AI tools require your computer to be on/awake for tasks to run. You might get an error message, or the output never shows up.
These failures are easy to address because it’s often clear what went awry. You reconnect and try again.
I rely heavily on AI scheduled tasks and have a notification system built in. If something breaks, I get an email telling me to review the process. That way, I realize immediately that something went wrong and can fix it.
The unpredictable AI failures
The harder failure is when an AI process runs as expected, but the output gets worse. Models and functionality change almost daily. You may have a well-crafted prompt that gave you what you were looking for last month, but the results are noticeably different today.
People in the industry call this “drift.” Most setups have no way to catch it, since technically nothing broke.
I use Claude and struggled with Opus 5.0. For the way I use AI, the output was “off” in a way I couldn’t quite put my finger on. I found myself returning to Opus 4.6, which was the last model where I felt comfortable. (As an aside: the update to Opus 5.5 went much more smoothly.)
The most reliable way to catch drift is to consistently review AI output, no matter how long it has been working as expected. Give every review the same critical eye you gave it the first time you tested an AI process.
Drift is inconvenient, especially when the whole point of an automation is to save you time. But you can’t count on a model to behave the same way from one month to the next. I look at it this way: a 5-minute review is still faster than the 20-30 minutes it may have taken me to do the task manually.
Factor ‘AI maintenance’ into your routine
AI tools save me a lot of time. But keeping my prompts and processes working smoothly takes time, too. That time has to come from somewhere.
Some fixes can’t wait. When a notification tells me something failed, I often need to look at it right away. Depending on what broke, it can throw off my whole workflow until I fix it.
Other changes I have to set aside time to review thoroughly. When a new model is announced, I don’t switch over right away. A big change in output could derail an afternoon, so I take time to test it first.
Upkeep is part of the cost of any AI system you set up. And if a process takes more time to maintain than it saves you, it’s fair to ask whether it’s worth keeping.
For quite some time in the already long series of articles I’ve been writing on what is and what should be corporate AI, I have been discussing concepts such as agents, working with ontologies and world models that represent the company, connecting th…
Roughly 11 million workers may need new occupations, and many lack the skills or credentials to get there.
As AI agents gain spending power, business leaders face a costly question of who pays when they go rogue.
Canada needs ‘quantum wins’ that don’t rely on winning a spending contest against the largest economies, Edmonton company says
Hello and welcome to Modern CEO! I’m Stephanie Mehta, CEO and chief content officer of Mansueto Ventures. Each week this newsletter explores inclusive approaches to leadership drawn from conversations with executives and entrepreneurs, and from the pa…
“I don’t know if I came across super harsh or not. At some point, I lost sense of how I was showing up.”
That was my client after having a hard conversation with one of her leaders about getting the resources she needed in her rol…
Jobs are always evolving, particularly in the age of artificial intelligence. But if the work you’re doing bears little resemblance to the role you were promised, and especially if no one has explained or discussed the change with you, you might be ex…
