Have you ever used an AI tool and just wished it would get to the point? You find yourself thinking: “I just didn’t want to know that much about it.”
Or maybe you fell victim to the clickbait cliffhangers: “If you like, I can also tell you the secret most people miss, the history of the field, or a surprising connection that makes this truly effective for people like you.”
You can’t resist, and down the rabbit hole you go. Hours disappear into that chat box.
And if you can’t resist, what chance for your employees?
Organizations need to figure this out. It starts by recognising there’s a problem. According to Workday research (January 2026), while 77% of employees *claim* higher productivity due to AI, organizations are losing approximately 40% of those gains to a “productivity tax” – time spent correcting, refining, and auditing low-quality responses. My guess is that even more is lost due to sheer AI-generated overload.
In other words, people might think they’re being super-productive with AI when in fact it’s a net drag, albeit a fun one.
Why are large language models so verbose? LLMs are designed to be a certain kind of helpful by giving not just a direct answer but also context, options and possible next steps. This shows up as long explanations and invitations to explore further. It can be useful when learning or generating ideas – but unless you grab the steering wheel, conversations quickly expand beyond the original question.
Before going further, consider a prescient point made by Stafford Beer – one of the first people to use computers seriously in business – way in the 1970s. Beer saw that even back then, businesses were using computers all wrong. They’d enthusiastically cry: “Look at all the DATA we can produce!”
Data is almost universally assumed to be ‘good’, but it often fails to convey information: stuff you can grasp, decide about and act on.
Beer argued that computers should help with intelligent filtering and summarising: reducing noise, triaging information, and presenting what matters. For proof that they often don’t, look at all your incoming channels: email inbox, Slack or Teams channels, WhatsApp messages, task notifications from multiple enterprise and personal apps and on and on…
In practice, organizations often use computers in ways that produce more demands, more options, and more cognitive load rather than less.
The default LLM pattern – verbosity plus optional “bonus content” and “cliff-hangers” – amplifies data, but not necessarily relevant, actionable information. Even when the content is useful, it increases the volume of material that must be processed by busy managers.
This helps explain the “productivity tax”: time and cognitive effort are consumed managing excess material rather than benefiting from usable insight. Multiply this across an organization and you can see that for all its power, AI can actually be a net productivity drag.
This is not an argument against AI. It’s actually an argument to zero in on one of its biggest opportunities: AI has huge capabilities to summarise and filter.
If used that way, it can be a major amplifier of productivity. Those who are getting the most value from AI are managing this explicitly – treating interaction with AI as a problem of information summary, not just getting more data.
Default LLM behaviour amplifies inputs (too many words, options, tangents). Organizations need to ensure that people get selective elaboration and intelligent filtering.
Otherwise, it’s down the rabbit hole they go.
© Andy Bass 2026