Inside a Company That Runs Almost Entirely on AI Tools

Kaleidoscope Analytics is an eight-person B2B data startup based in Indiranagar, Bengaluru. They have no dedicated content writer, no full-time designer, no HR department, and no marketing ops hire. They do have ₹3.2 crore in ARR, two enterprise clients, and a tool stack that would have been science fiction in 2020.

This is not a story about AI replacing jobs. It's a story about what actually happens, day by day, when a small team decides to treat AI as a core operating infrastructure rather than a productivity experiment.

7:30 AM: The Morning Brief Nobody Writes

Founder Arjun Shetty starts his day with a briefing document he didn't write. The night before, a scheduled workflow pulls from their CRM, product analytics, support tickets, and Slack, runs it through a summarization prompt he's refined over six months, and generates a one-page morning brief: what happened yesterday, what looks anomalous, what decisions need attention today.

"It's like having a chief of staff who worked all night and wrote me a memo," Arjun says. "Except it's Zapier and Claude and about ₹200 a month."

The brief isn't always right. About twice a week it flags something that turns out to be a data artifact. Arjun has learned to read it critically, not credulously. But it surfaces things his eye would have missed — a client whose usage dropped 30% three days ago, a support ticket category that spiked without being escalated.

9:00 AM: The Sales Call That Started With AI Research

Their sales lead, Meghna, is on a call with a procurement head at a Pune manufacturing firm. Before the call, she spent 20 minutes with a research prompt that scraped the company's recent press, LinkedIn announcements, and any public financial data, then generated a one-page briefing: likely pain points, recent strategic moves, who in the organization might be a champion.

Is this perfect? No. The AI occasionally confuses two similarly named companies. Once it generated a confident briefing about a product line the company had discontinued. Meghna verifies the facts she plans to use. But the baseline research that used to take her 90 minutes of manual Googling now takes 20 minutes of reading and verification.

"I go into every call knowing more than I used to," she says. "That changes the conversation."

11:00 AM: The Design That Nobody Designed

Kaleidoscope's pitch decks, one-pagers, and client reports have a visual identity that looks like a mid-market design agency produced them. Nobody at the company is a designer. Their process: write the content first, paste it into a structured prompt that describes their brand guidelines, and use a combination of Canva AI and manual adjustments to produce layouts that would cost ₹40,000 per deck from a freelancer.

This is the area where they're most honest about the limitations. "Anything that needs genuine creative originality — a brand refresh, a conference keynote, a campaign concept — we still hire someone," says Arjun. "But the 80% of design work that's assembly and formatting? We don't need to hire for that anymore."

2:00 PM: Engineering With AI Copilot, Not AI Engineer

Their two engineers use GitHub Copilot and Claude for code generation. This is probably the area with the most nuance.

"Copilot is fantastic for boilerplate," says Kiran, their backend lead. "Anything I've done a version of before, it finishes my thoughts. Anything genuinely novel or architecturally complex — it suggests things that look right and are subtly wrong. You have to know enough to catch that."

The productivity gain is real: Kiran estimates he ships roughly 30-40% faster than he did two years ago. The risk is also real: a junior developer without strong fundamentals using AI-generated code they don't understand is a technical debt factory. Kaleidoscope doesn't have junior engineers yet. Arjun says he thinks carefully about what that hire will look like when it happens.

4:00 PM: The Support Ticket That Answered Itself

Their support workflow runs on a Claude-powered triage system. Incoming tickets are classified, matched against a knowledge base, and for about 45% of queries, a drafted response is generated that a human reviews and sends. For another 30%, the response is sent automatically to confirmed low-risk query types that have been stable for six months.

The remaining 25% — complex, ambiguous, or high-emotion tickets — route to a human immediately. "The AI is very good at the easy stuff," says their customer success lead Divya. "The hard stuff — a client who's frustrated, a bug that broke their workflow — you want a human there. Empathy doesn't automate."

What's Broken

Here's what Arjun says doesn't work:

AI for strategy. They tried using AI to generate strategic options for a pricing change. The outputs were generic frameworks they could have found in a business school textbook. The actual strategic decision required judgment about their specific market and customers that no AI had context for.

AI for performance management. They briefly experimented with AI-generated performance summaries. The summaries were technically accurate and felt hollow. People want to hear assessments of their work from a human who knows them. That experiment lasted six weeks.

Multi-model workflows without a human checkpoint. They had one workflow where AI-generated content flowed into AI-generated formatting into AI-generated distribution without a human reviewing the final output. It went out once with an embarrassing error in a client report header. Now everything with a client-facing endpoint has a human sign-off step.

The companies that are getting this right are using AI to handle volume and velocity while keeping humans on judgment and relationships. The companies that are getting it wrong are removing the human from places where judgment and relationships are exactly what's required.

What They'll Never Go Back On

The morning brief. The research workflow. Code completion. First-draft generation for anything — emails, reports, proposals. These have fundamentally changed the capacity of eight people to operate like a team twice their size.

"I sometimes think about what it would take to hire the equivalent of what our AI stack does," Arjun says. "It would be five or six people, maybe ₹1.5 crore in salaries annually, plus management overhead. Instead we're paying maybe ₹3 lakh a year in tool subscriptions."

The math is uncomfortable for anyone whose job is doing the things those tools now do. Arjun is honest about that. "We're not a proof point that AI creates jobs. We're a proof point that AI lets a small team operate at scale. What that means at the macro level is a question I don't think anyone has answered yet."

---