By Karthi Sepulohniam (Subramaniam), Head of Asia Pacific & Australia, Partners for Growth
Partners for Growth AI Series – Episode 2
Featuring Suneeta Mall , Head of AI Engineering, Harrison.ai
The best AI lessons come from the people shipping it.
This is the second instalment of our AI Series, where we sit down with technology leaders across our ecosystem to understand how they’re using AI in ways that are practical, non-obvious, and genuinely changing how they operate.
For episode 2, I sat down with Suneeta Mall, Head of AI Engineering at Harrison.ai, to understand how they’re building these systems, how the team thinks about working alongside AI without losing their own judgment, and where the real bottleneck lies in getting AI into hospitals. Harrison.ai is different in that they build the models themselves, from the ground up, to solve one of the most consequential problems in healthcare: there aren’t enough radiologists, and the ones we have are drowning in admin.
From Second Pair of Eyes to First Draft of the Report
Harrison.ai‘s products have evolved through a clear arc. The first generation – regulated medical devices for chest X-ray, CT brain, and CT chest – acts as a diagnostic aid. The AI reads the scan, detects pathologies, localises findings, and presents them to the radiologist as a second opinion. These products are already live across more than 40% of NHS Trusts and Health Boards in the UK, and are available to more than half of radiologists in Australia.
The next step is more ambitious: with the recent launch of Harrison.Rad 1.5*, the company’s radiology foundation model, the product has moved from flagging findings to drafting reports – producing the first version of a clinical report for a radiologist to review, modify, and sign off.
As Suneeta described it, the aim is to absorb the administrative burden so radiologists can spend more time on patient care and less time on documentation. It’s a shift from AI as a safety net to AI as a working colleague.
“Off the Shelf Hasn’t Served Us”
One of the clearest points Suneeta made was that general-purpose AI models don’t cut it in radiology. The domain knowledge is too specialised, too nuanced, and often not available in the general training data that large language models are built on.
Harrison.ai builds its own models in-house. The team specialises in model architecture – drawing on published research, then modifying and extending it to fit the specific requirements of clinical imaging. They train on their own compute cluster, running multi-stage training processes designed to strengthen the clinical signal and suppress noise in the data, because off-the-shelf models don’t hold up in a clinical setting. Harrison.Rad 1.5 passed the mock UK radiologist certification exam (FRCR 2B Short Case) with a median score of 86.5 against a 73.2 pass mark, clearing over 50% of exam sheets, while every other model tested, domain-specific and frontier alike, passed none. https://arxiv.org/html/2607.05880v1
The foundation model, Harrison.Rad 1.5, is then extended into what Suneeta described as agentic workflows – a system of specialised components working together to produce a draft report.
Execution Is No Longer the Bottleneck
While AI has dramatically compressed execution timelines, the constraint has moved upstream. Execution is no longer the bottleneck – the hard problem now is deciding what to invest time, resources, and effort into. The team can build big features faster than ever, but choosing the right ones requires the kind of strategic judgment that AI can’t provide.
This is consistent with what we’ve heard from every company in this series: the speed gain is real, but the challenge is staying in front of it with clear thinking about where to point those productivity gains.
Cognitive Surrender: The Risk Nobody Talks About
The most thought-provoking part of the conversation was Suneeta’s framework for how her team works alongside AI without losing their own judgment.
She referenced recent research from Wharton suggesting we’ve moved beyond Daniel Kahneman’s two systems of thinking – fast and slow – into a third mode: augmented thinking. The risk, as she sees it, is that once people become comfortable with this third mode, it becomes easy to accept what AI gives you without challenging it.
Her team’s approach is deliberate. First, treat AI as an associate, not an expert. The default posture is to look for weaknesses in what it presents, not to accept it at face value. Second, before going to AI for a cognitive task, formulate your own view first – then use AI as a thinking partner to challenge and refine it.
Suneeta was candid that this tension cuts both ways. There are times when AI pushes her thinking in directions that she wouldn’t have reached on her own. My key takeaway from Suneeta on this is that the AI augmentation is real, but so is the risk of surrendering to it, and the discipline to resist that requires conscious effort.
The Real Barrier: Getting AI Into Hospitals
When I asked about the biggest challenge in deploying AI in healthcare, Suneeta’s answer wasn’t about the models. The models, she said, are quite capable already. The barrier is everything around them.
Healthcare systems have evolved organically over decades. Electronic medical records are varied. Imaging systems differ from hospital to hospital. The way radiologists dictate reports and how those dictations get documented into systems – all of it varies. There’s no single standard, and that lack of interoperability makes it remarkably difficult to drop an AI solution into a clinical workflow and have it just work.
Harrison.ai‘s response has been to build what they call the Harrison Open Platform – a standardisation layer designed to sit beneath their AI products and provide a consistent foundation across different hospital environments. It’s infrastructure work that isn’t glamorous, but without it, even the best models struggle to reach clinicians at scale.
AI That Improves AI
When I asked about non-obvious use cases, Suneeta shared one worth highlighting.
Harrison.ai is building systems where AI identifies the weaknesses in their own foundation model. At the scale they operate, putting radiologists in the loop to evaluate every output isn’t feasible. So instead, they’ve built what Suneeta described as probes – secondary AI systems that learn to recognise where the foundation model underperforms. Those probes then push the model to learn more effectively in its weak areas.
It’s a technique known in the research community (reinforcement learning), but seeing it applied in production at a clinical AI company is rare. The result is a feedback loop where the AI effectively teaches itself where it needs to get better.
The Takeaway
Harrison.ai occupies a unique position in this series. They’re not just adopting AI tools – they’re creating them, for one of the highest-stakes domains imaginable. What stood out was the rigour beneath the ambition: the discipline around cognitive surrender, the honesty about deployment barriers, and the recognition that even at the frontier, the hardest problems are human and systemic, not technical.
*Disclaimer: Harrison.Rad 1.5 is a research-only foundation model, not a medical device regulatory-cleared for clinical use. Harrison.ai is seeking regulatory clearance for medical devices powered by these models in various markets.
This is Episode 2 of AI: Where Innovation Meets Execution, a series that profiles how technology companies and investors across our ecosystem are deploying AI in practical, non-obvious ways. If you’re building something interesting with AI and want to be featured, reach out to us at karthi@pfgrowth.com.
Partners for Growth (PFG) is a global private credit firm specializing in custom debt solutions for high-growth companies. For over twenty years, PFG has provided growth debt financing to tech, fintech, healthcare, and tech-enabled companies to accelerate their path to profitability or finance specific assets at pivotal stages of growth. Since its inception, PFG has partnered with more than 250 companies across 15+ countries. If you’re a founder or CFO exploring minimally dilutive capital, we’d love to hear from you: pfgrowth.com/connect/
Featured companies may be clients or portfolio companies of PFG. No compensation has been provided for inclusion. This content is for informational purposes only and does not constitute investment advice or a recommendation.
