Most pharma AI is borrowed, ours is built

When ChatGPT launched publicly in 2022, I did what most people in our industry did: I watched closely. The conversation in pharma filled almost immediately with warnings about hallucinations. And they weren't wrong. But I kept pondering about a quieter, more structural problem that nobody seemed to be addressing at that time: confidentiality.

Our pharma clients have strict confidentiality clauses. Their unpublished data, their pipeline compounds, their market positioning - none of it should be used for training commercial LLMs. Conversely, banning AI outright in my medical communications agency felt oddly off. I kept coming back to an analogy: before iTunes and Spotify, prohibition didn't stop illegal music downloads. Instead, it drove people toward unsafe options. The music industry solved it by building better and safer alternatives.

So, I too decided to build our very own alternative.

The office server

Early 2024, I decided to install a dedicated AI server in our office. The idea was simple: giving my team the AI powers without risking data the breaches as with most LLMs. My team's enthusiasm was lower than I expected, not because people weren't curious, but because the tools weren't shaped for purpose. Afterall, generic AI is built for generic tasks and medcomms is anything but generic.

That's when I brought in an AI engineer, a PhD researcher specialising in medical imaging. He started program the server for our specific workflows. What happened next I hadn't even the slightest anticipated.

The collaboration worked, not because of the technology alone, but because of the knowledge pairing. I brought in deep medcomms knowhow, human processes, covering the sequencing of tasks, review cycles, and judgement calls that sit between a data package and a polished deliverable. The developer brought the technical architecture to encode those processes into machine workflows.

That combination, I've come to believe, is the real formula for creating useful AI: an experienced AI engineer working alongside someone with deep, granular knowledge of human workflows. LLMs perform best when they follow existing human processes, rather than reinventing them.

The moment it clicked

There's one particular module that changed my perception about AI.

For medical writers, editing a scientific PowerPoint presentation is a painstaking, specialist job. For a 70-slide deck in a complex therapeutic area like haematology, a junior analyst might face 30 or more abbreviations on a single slide, each one needing to be listed, standardised, and verified. References need to be reformatted to a consistent citation style throughout. Not only does it consume as much as two-and-a-halve days of a junior analyst's time, including back-and-forth review with a senior lead, the meticulous editing distracts from higher value tasks: storytelling and scientific rigor.

Our AI does it in less than 15 minutes.

The time required to complete the task was not all that changed, it altered our entire office workflow. Decks no longer go to the medical writing team first, but instead through AI, then to the creative team, before it arrives at the medical writing lead for a high-level scientific review. We've removed the full 2½ days of junior writer's time, along with the review iterations that came with it.

Importantly, our process does not make medical writers redundant, instead positions them as Human-in-The-Middle at the right points in the workflow to check, validate, and approve. They are an intrinsic part of the process, but instead of focussing on details they can now focus on the high-value tasks whereas our AI automates time-consuming, low-value tasks, compliantly, without exposing confidential data, and in ways that mirror human processes.

Bootstrapping an AI venture

CORTiX has not taken external investment, no VC round, no seed funding. A deliberate choice, and it shapes everything about how we build.

When a new project is allocated to our medcomms agency ['mediPr], we question whether it could partially be handled by a reusable AI workflow. If the answer is we can, we design and build an AI agent for it. This way, the development cost is absorbed into a live commercial engagement.

It's a slow approach by startup standards. But it means every tool we build has its practical usefulness validated in a live workflow, while confidential data is handled securely on our own infrastructure. My observation is that too many AI ventures get started without truly understanding human and organisational processes. 

The second funding stream is more unconventional. One of our core modules transcribes audio, a capability that sits at the heart of several other features, including meeting summaries and advisory board reports. Rather than keeping it locked inside the platform, we spun it out as a standalone consumer product: Dub-Dub.ai. The idea is generating independent revenue outside our traditional pharma client base and, crucially, to get real user feedback at scale before we integrate the module into CORTiX's professional offering.

The future of our AI

In three years, I want CORTiX to be serving freelance medical writers, medcomms agencies, and pharma companies directly, both through SaaS and on-premise deployments that keep sensitive data exactly where it belongs.

But the bigger point isn't about our tool, it's about what the industry needs to understand about the difference between adopting generic LLMs and building a for-purpose, domain-specific, AI system.

Using a generic LLM for drafting summaries or reformatting documents is not comparable to designing compliant, domain-specific AI agents. One is a productivity shortcut. The other is an infrastructure. One puts your clients' data at risk. The other protects it by design.

If you work in medical affairs, medical writing, or run a medcomms agency, the question is no longer whether AI will change how we work. It already has. The question is whether the AI you are using was built for purpose or exists with the intention to train LLM models.

We're building it from the ground up. And we're just getting started.