Madhukar Bhatia, Managing Partner at Pentathlon Ventures, doesn’t warm up the room with small talk.
On stage at Pentathlon Day 2025, looking out at a hall full of founders and investors who have all sat through too many “AI panels,” he asks 4 portfolio founders to explain one thing.
How do they leverage AI inside their companies?
The people next to him are very different.
- Dr Amit Kharat, Co‑founder and CEO of DeepTek, an AI radiology company working inside one of the most regulated workflows in healthcare
- Deepankar Biswas, Co‑founder and CEO of ClearTrust, which protects publishers and media buyers from invalid traffic while keeping them aligned with Media Rating Guidelines.
- Hridayesh “Harry” Gupta, Co‑founder and CEO of QuickReply.ai, a WhatsApp marketing and automation engine for eCommerce
- Prateek Jain, Founder and CEO of CustomerGlu, a gamified engagement and retention platform
Murugan asks Prateek to go first.
“AI literally saved my company from dying,” he says. The room goes quiet.
For the next hour they stay on that theme.
Not “AI will change everything.”
Very specific stories about using AI inside the company to survive, produce more with fewer people and turn fixed cost into something more flexible.
Why this conversation matters now
For B2B founders and investors, AI has become the new infrastructure layer.
The efficiency benchmarks are shifting under everyone’s feet. A few years ago, many founders quietly assumed something like “50 people for the first million in ARR” as a reasonable shape of a company. Today you hear credible stories of sub‑ten‑person teams running similar revenue, especially when they design around AI from day zero.
India makes this more urgent, not less. Domain talent is scarce in areas like healthcare, lending, and compliance. Customers are price‑sensitive and often expect global‑grade outcomes at local pricing. Regulation adds complexity and slows some obvious shortcuts. In that context, using AI internally is a survival strategy.
And yet, most conversations about AI in startups sound the same.
Everyone says they “use AI.” Few talk openly about the messy parts. The technical debt from vibe‑coding entire features through a chat window. The OpenAI invoices creeping up every month. The politics of putting AI into workflows that teams currently “own.”
This panel cut through that.
DeepTek: when AI becomes invisible infrastructure
On a typical SaaS panel, a radiologist might feel out of place. On this one, Dr Amit Kharat is central.
Dr Amit Kharat is Co‑founder and CEO of DeepTek. He starts by grounding everyone in the problem.
Radiology and imaging sit at the centre of healthcare and every patient has to pass through imaging at some point. With an estimated 20,000 – 25,000 radiologists serving a population of over 1.4 billion. The question is simple and hard at the same time ‘How do those twenty thousand people keep up?’
DeepTek started in 2017 with classic computer vision.
- Chest X‑rays as the basic, common study
- A model that separates normal from abnormal
- Draft reports for radiologists to edit and sign.
Over time, the stack expanded to more body parts and then to GenAI. Today it looks less like “one model” and more like an internal factory with many agents.
- Specialist agents tuned to CT brain, MRI spine, chest and other modalities
- A formatting agent that turns raw findings into a structured, clinically acceptable report
- A fact‑checking agent that scans drafts for obvious errors: a gallbladder mentioned when surgery notes say it has been removed, a gender mismatch with previous records, left–right errors on single‑lung images.
All of this runs in the background across 1000s of scans a day in more than 1000 centres. The radiologist still signs every report.
In straightforward cases, Amit sees productivity gains up to about 90%. On complex, multi‑finding scans, the gain is still in the 50–60% range.
He keeps emphasising one point that is healthcare is not a space where AI can simply replace experts. Regulation and ethics demand human oversight. That is why DeepTek does not promise “full automation,” it talks about productivity and about turning each scarce radiologist into the effective equivalent of several without compromising quality.
If you strip away the clinical details, you get a model that applies to many founders. AI is the invisible factory under every unit of work.
CustomerGlu: when AI literally keeps the lights on
If DeepTek shows AI as infrastructure, CustomerGlu shows AI as emergency support.
Prateek Jain, Founder and CEO of CustomerGlu, runs an in‑app gamified engagement and retention platform. A couple of years ago, he hit a wall.
The company had live customers and real usage but funds had dried up. Costs were high and revenue was not where it needed to be because they could no longer pay salaries, the engineering team left. Prateek found himself alone with the codebase.
The choices were stark:
Shut the company down
Sell it
Or try to keep going with almost no team
Customer demand was still there, clients wanted the product to keep working. The missing piece was execution.
Prateek has a computer science background but had not been the one deploying this particular stack. He needed to migrate workloads from AWS Lambda to Azure Functions and re‑wire parts of the architecture without breaking everything.
The turning point was not a big “AI strategy” document but a series of late‑night sessions with tools like ChatGPT.
- At first, simple “how do I” questions as a chat
- Then co‑pilots embedded into the IDE to suggest code inline
- Later, agentic tools that could analyse local codebases and suggest refactors
The tools became a stand‑in for a senior engineer sitting next to him.
Fast‑forward to now. CustomerGlu is effectively a two‑person company. They still serve multiple customers. Their revenue per employee and monthly burn numbers would have sounded unrealistic at seed stage in 2020. Today they look workable.
Prateek also offers the sharpest caution on stage. He calls large language models “elite autocomplete,” not magic. In many ranking and matching problems, he prefers a well‑designed decision tree or classical model. He treats models and providers like utilities.
- During experimentation, they route through cheaper or free models, often via aggregators such as OpenRouter
- In production, they pay for more expensive models only where quality clearly pays for itself.
In his story, leveraging AI internally was not something you bolt on once you have a full team. It was a way to keep the team as small as reality demanded.
ClearTrust, QuickReply and disciplined use of AI
If DeepTek and CustomerGlu are the two ends of the spectrum, ClearTrust and QuickReply sit in the middle where a lot of SaaS companies live. They show how to combine upside with discipline.
ClearTrust – AI tools, human‑owned code
Deepankar Biswas, Co‑founder and CEO of ClearTrust, protects digital marketers from fake traffic and bad leads. The threat has changed.
In earlier campaigns, they could detect around 30% of traffic as fraud. Lately that percentage has gone up, and the nature of fraud has become more subtle. Residential proxies, slow “poison” traffic that looks like human behaviour and corrupts performance data over weeks and the bots that mimic normal browsing patterns.
To stay ahead, ClearTrust uses AI across several internal building blocks:
- Synthetic data generation to simulate different kinds of fraud at scale
- Trap generators that look at blended customer and synthetic data to find new patterns
- Log‑analysis tools and simulators, often drafted using copilots and refined by engineers
The important part is what they do not automate. Their CTO pushed back early against letting AI directly commit production code. Today, ClearTrust uses tools like Cursor and Claude to explore options and draft code and final code that goes into the core product is reviewed and typed in by humans.
Overall, they see roughly a 50% efficiency gain in designing new traps and tools, without handing their codebase over to opaque generations.
QuickReply – AI as translator between product and engineering
Hridayesh “Harry” Gupta, Co‑founder and CEO of QuickReply.ai, runs a WhatsApp marketing and automation platform for ecommerce brands. For his team, the most visible AI benefit has been in product communication, not just in code.
Their unlock came when a GenAI‑native product manager joined. Instead of:
- Long calls
- Figma files
- Bullet‑point specs
they started to use prompt‑to‑UI tools like bolt.new and v0.dev to generate interactive prototypes. Product could type out the desired behaviour and quickly produce flows, edge cases, and states that developers could see and click through.
That shift did not show up as “X% more lines of code.” It showed up as:
- Fewer cycles between product and engineering
- Clearer specs
- Faster releases with fewer misunderstandings
In this setup, AI is a translator between product, design and engineering, not just a code‑writing assistant.
Across both ClearTrust and QuickReply, the common thread is clear boundaries. AI can think, suggest, and draft. Humans still own the core production systems and the final decisions.
Beyond product: proposals, research, support, and the cost line
Once you start listening for it, AI shows up everywhere in these companies. Not just in what they sell, but in how they run.
A few examples from the panel:
1. Proposals that used to take weeks
DeepTek sells into hospitals and health systems. Their “proposal” is not a one‑pager. It is:
- 100–200‑page tender responses
- Clinical references and regulatory language
- Multiple pricing models: per study, perpetual, bundled lifetime, on‑prem vs cloud vs PACS
They now let AI agents:
- Pull prior answers
- Insert the right clinical references
- Assemble the first draft across all these variations
Humans then review and edit. Time to a solid draft has dropped sharply, even though the bar for accuracy is high.
2. Market and literature intelligence
In fast‑moving domains like medical AI or fraud, founders described using GenAI to:
- Scan recent papers
- Summarise regulatory changes
- Map competitor feature sets and positioning
Often, the tools surface material the team had not yet seen, or compress hours of reading into a couple of tight briefs.
3. Support with a real “human in the loop”
ClearTrust talked about internal GPT‑style assistants tuned to each customer profile:
- Support reps paste in a question
- The assistant drafts a contextual answer, using that customer’s config and history
- The rep edits and sends
This is not “let’s slap a bot on the website.” It is AI as a researcher and first drafter, with a human still owning tone, nuance, and responsibility.
4. A new cost line: “cost of intelligence”
Nobody on stage called it that, but the pattern was obvious. AI tools are now a real spend category:
- Model calls and tokens
- Copilot seats
- Agents running on cloud
Founders joked that AI feels like the razor‑and‑cartridge model:
- Easy and cheap to start
- Painful if you do not watch usage and scope
Pricing models are moving from per‑seat to per‑usage and, over time, towards credits, tokens, and outcome‑linked pricing – something Pentathlon has been writing about separately.
The more disciplined teams are already:
- Routing different tasks to different models based on cost/quality
- Setting strict scopes for each internal tool
- Using no‑code workflows (Make, Zapier‑like tools) wherever a simple trigger‑action beats a fancy agent
It is a quiet but important shift: AI is becoming a line item next to cloud and salaries, not a side experiment.
What this says about the next wave of B2B
Walking out of that room, the impression was clear. For these founders, and for a fund like Pentathlon, the question is no longer “Should we use AI.” It is “Where does AI belong in our operating system and where does it not”.
They are building companies around a few shared beliefs. Smaller teams can do more when AI is treated as infrastructure, not as a feature. Business models will have to account for usage‑based and outcome‑based pricing, not just seats. Founders who understand their domain deeply and then layer AI on top will outlast those who lead with demos alone.
If you are a B2B founder, there are a couple of uncomfortable but useful questions this panel leaves you with.
If AI disappeared from your company tomorrow, what would break? Would it just be your pitch deck? Or would proposals stall, QA slow down, support queues pile up, and product cycles stretch?
And right now, is AI sitting mainly in your marketing copy. Or is it quietly running under your workflows, shaping how many people you need, how fast you move, and how resilient you are when the next shock hits.
