Every morning, Madhu Karuthedath, 42, the Bengaluru-based solo founder of the employer-brand consultancy The Quick Brown Fox Co., opens what looks like a small control room on his laptop. Several tools gather news, Claude turns the previous day’s work into a prioritised checklist, ChatGPT helps him brainstorm and structure ideas, and runs them through Gemini for critique, and then hands the clearer brief back to Claude for implementation. While the engines work, does he take a break? No. He is waiting for one tool to share the output, checking another, and briefing a third.
In the sitcom How I Met Your Mother, the character Barney Stinson made a catchphrase out of suspense, inserting “wait-for-it” into everything, and most famously “legen-wait-for-it-dary”. Today, between the AI prompt and the output sits a new kind of ‘wait-for-it’. It lasts a few seconds or several minutes, but is rarely empty: An email is answered, an errand is run, a social media post is crafted, or another task is assigned.
For solo entrepreneurs and small teams, several AI assistants can be working in the background simultaneously, while one person briefs, supervises and reviews them. AI has shortened individual tasks, but the gaps are filled with more work, as editing other AI content, context-switching, and verifications have become jobs in themselves. And with AI companies beginning to make generated output identifiable through watermarks, these pauses also make us wonder: Where does AI’s work end, and the human’s begin?
Aditya Malani, 28, an engineer and solo-founder building Speedy Squirrel (a Shopify development company) in Bengaluru, usually has two agents working simultaneously. One fixes or builds a product feature from a previously scoped product requirements document; the other researches a new feature or domain. Meanwhile, Malani reads online or replies to customer-support tickets. He describes the agents as junior developers who require specific instructions. “The challenge right now feels like reviewing AI output and deciding what goes in,” he says.
THE AI 'TEAM'
Then there are those who do not consider AI agents as members of their “team”. To Phalgun Guduthur, 36, an independent product builder, AI is just a computer doing what it is told to. Cursor and Claude build features and analyse marketing while he browses X (formerly Twitter), cooks, or reads about what to do next. “I vehemently say it has made me cram more things into the same number of hours. I actually work a lot more now, as parked ideas move into the time saved,” he says. Akhilesh Bhamburkar, a 33-year-old tech solo-preneur in Nagpur, agrees: “I slog for a minimum of ten hours a day, more than my previous corporate job.”
AI has also created an abundance problem. Through vibe coding, almost every idea now feels buildable. But there’s also a price to pay for building the wrong thing. The productivity gain is showcased as a business’s capacity, but when execution speeds up, the slowest part of the system can become the human in the middle, deciding what happens next.
“I have become my own bottleneck,” says Sankalp Sinha, 33, the Bengaluru-based founder of NextDoor.Company, a startup for map-based recruitment. Faster execution means he must review completed tasks and keep supplying direction. “AI can drive the car, but the ‘human in the loop’ must hold the steering,” he adds.
Earlier, a developer stuck on a problem might search Stack Overflow (a massively popular online community for programmers), notice whether a solution had been accepted or upvoted, remain sceptical and test it. AI output often arrives at speed, but without that social scaffolding and filter, and therefore needs verification. “Thank God for some of the need to verify. Otherwise, there would be a lot more AI slop out there than there is today,” Bhamburkar says.
Priyanka Iyer, a 32-year-old principal engineer in Chicago, tries to remain in “work mode”, all the time. She likens the experience to watching someone work continuously in front of you until their effort tires you out. Guduthur often has three or four tasks running inside the same tool; his concern is that he may be moving too quickly to think through his decisions. Yet he is reluctant to blame AI alone. He feels short-form content and doomscrolling are to blame for the damage to his attention span.
Sophie Leroy, a professor at the University of Washington Bothell’s School of Business, calls the cognitive carry-over from one unfinished task to another as “attention residue”. Her research found that people struggled to transfer their attention fully after finishing one work task, and performed worse on the subsequent task. AI may reduce execution time while multiplying these unfinished mental loops.
CHECKING AND VERIFYING THE OUTPUT
The boundary between assistance and authorship is also now more easily traceable. On 14 August, Anthropic announced that future Claude models would place an imperceptible watermark in generated text globally to comply with the EU AI Act. Google DeepMind has used a somewhat similar system, SynthID, for text in the Gemini app and web browser since 2024, alongside watermarks for AI-generated images, audio and video.
Bhamburkar worries that this may discourage responsible users too. A writer might develop an original idea, ask AI to rephrase parts of it and then edit the result heavily, without knowing whether the watermark remains. Malani is less concerned. Tested code is judged by whether it works and he claims that he rewrites his blogs and content in his authentic voice, because asking people to read it means asking for their time. For Iyer’s team, watermarking would change little because organisational governance already defines where AI can and cannot be used. Each developer still owns their code, with shared responsibility for the final product.
The technology recognises some of this distinction. Anthropic says lightly proof-read human writing may contain too little model-generated language to register, while code offers fewer arbitrary choices and therefore carries less watermarking. A watermark can suggest that Claude was likely involved, but it cannot measure the human judgement that followed.
Sinha draws the line at using AI in the context of experiences that require time. E.g., he would not ask AI to read an e-book for him, because understanding flows from the act of reading, as ideas simmer and return when attention has moved elsewhere. “AI is a reactive tool. It can only react in the moment and not after that. But we as humans can,” he says. Context and judgment, in this case, take the cake.
Machines may not daydream, but their loading screens have developed quite an imagination. ChatGPT tells us it is thinking or searching, while Claude Code might be ‘Crystallizing’, ‘Sleuthing’, ‘Canoodling’, ‘Moonwalking’ or ‘Flibbertigibbeting’. Claude Code users can even customise these spinner verbs by choosing the whimsical vocabulary of the machine’s apparent inner monologue.
In the pre-AI era, one used to wait for a work outcome. There would be anticipation and impatience, perhaps dread. There was a sense of possibility, and time for the mind to wander. The AI waiting room has no such vacancy: every pause arrives with another prompt, another decision, and the cognitive load of constant context switching. Whatever comes next will only make the queue move faster. The future of work is “legen-wait-for-it-delegated”. No rest for the prompted!
Shalaka Kulkarni (@shalakulkarni) is a Bengaluru-based author who writes at the intersection of culture and technology.