The ‘Clerical Work’ Bottleneck in AI Conversations, According to Hiten Shah

In a recent LinkedIn post, Hiten Shah highlights a recurring challenge he’s observed when using AI models like Claude for work-related tasks. Shah points out that before the AI can effectively assist, users often find themselves performing significant preparatory work, a process he terms ‘clerical work.’ This initial effort, he argues, can negate some of the time-saving benefits of AI.

Shah elaborates on the nature of this preparatory phase, stating:

“Before Claude can do the work, I have to feed it the work.”

This involves gathering information from various sources such as Slack threads, email chains, calendar events, project updates, and documents. According to Shah, the user must then compile essential contextual details, including past decisions, involved parties, changes that have occurred, what is considered important, and the definition of ‘done’ for the task at hand.

The Contextual Gap in AI Interactions

Shah’s analysis suggests that a significant hurdle in leveraging AI for productivity is the upfront investment required to provide adequate context. He observes that by the time the AI model is sufficiently informed to begin its task, the user has already expended considerable effort in data aggregation and synthesis.

“By the time Claude has enough context, I have already done a bunch of clerical work for the model,” Shah writes in his post. This observation underscores a potential inefficiency in current AI workflows, where the human element is heavily involved in the data preparation stage.

Addressing the ‘Clerical Work’ Problem

The core of Shah’s argument is that this ‘clerical work’ represents a solvable problem within the AI integration process. He believes that current methods necessitate a manual bridging of the gap between where work data resides and where AI conversations take place.

Hiten Shah suggests that a more streamlined approach is possible. He points to the potential for improvements in how AI tools connect with existing work infrastructure:

“That is the part MCPs can fix.”

Shah indicates that the solution lies in better integration between AI and the tools where work activities are already documented and managed. This would allow AI conversations to start with a more relevant and immediate understanding of the ongoing work, reducing the need for extensive manual data feeding.

The Promise of Integrated AI Workflows

Shah is planning a live walkthrough to demonstrate how to address this issue, focusing on what he refers to as ‘MCPs’ (presumably a term related to managing context or connections). The objective, as outlined in his post, is to enable users to connect AI models directly to the platforms where their work is already happening.

“Connect Claude to the tools where your work already lives, so every conversation starts closer to the actual work,” Shah explains, outlining the intended outcome of such integrations. This vision suggests a future where AI acts as a more seamless extension of existing professional tools, rather than a separate entity requiring extensive manual onboarding for each task.

The challenge Shah identifies is not with the AI’s capabilities once it has context, but with the inefficiency of the context-gathering process itself. By highlighting this ‘clerical work’ bottleneck, Hiten Shah prompts a discussion on how AI tools can be better integrated into the fabric of daily work to unlock their full potential more efficiently.