Why Pharma Needs Purpose-Built AI, Not General Models

Jaseem Mahmmdla, Co-founder and CEO, Kognitic

I spend most of my weeks talking to pharma teams about competitive intelligence. Commercial strategy. Portfolio decisions. The conversations have changed in the last year. Not the questions. The assumptions behind them.

A year ago, the first question was usually about coverage. How many trials? How many indications? How does your dataset compare to Citeline or Cortellis? Those were fair questions, and they still come up.

Now the first question is almost always about AI. Specifically: “We already have an internal AI tool. Why would we need yours?”

I understand why the question gets asked. General-purpose AI models are impressive. They have changed how people interact with information across every industry. In pharma, teams are using them to draft summaries, explore therapeutic areas, and pull together background research faster than ever before.

But there is a difference between using AI to explore a question and using AI to support a decision. That difference is what I want to address.

The market shifted faster than the infrastructure

The pace of clinical data generation has accelerated dramatically. In oncology alone, the volume of trial readouts, endpoint disclosures, regulatory milestones, and conference presentations creates a competitive intelligence challenge that did not exist at this scale five years ago.

Most pharma organizations responded by provisioning general-purpose AI tools internally. This was a reasonable first step. These tools gave teams a faster way to orient themselves, summarize documents, and generate first-pass analyses.

What they did not give teams was a structured intelligence layer. And that is where the gap opened.

A general-purpose model can summarize what happened at a conference. It cannot normalize the endpoint data across the twelve trials that reported in the same indication that week. It cannot reconcile differences in patient populations, biomarker selection criteria, and comparator arms that determine whether those endpoints are comparable. It cannot trace every number back to a specific registry record, abstract, or publication.

Those are not AI problems. Those are evidence of infrastructure problems. And they require purpose-built systems to solve.

What I mean by purpose-built

I want to be precise about this because the term gets used loosely.

Purpose-built does not mean narrow. It does not mean a tool that works only in one indication or workflow. It means a system whose architecture was designed from the ground up for the specific constraints of clinical and competitive intelligence in pharma.

Those constraints are real:

• Entity ambiguity. A single-sentence biomarker name might indicate disease severity. In the next sentence, the same term might refer to a drug target. The clinical meaning is opposite depending on context. General models have seen these terms millions of times. They have not been trained to disambiguate them with the precision required by competitive intelligence.

• Implicit clinical logic. Trial eligibility criteria specify biomarker requirements through qualification scores rather than explicit labels. “PD-L1 Tumor Proportion Score should be greater than 50%” never says “positive.” A clinical scientist reads that and understands the implication. A general model may or may not, depending on how the question is framed.

• Nested negation. “Patients should not have any EGFR-sensitizing mutation to qualify for enrollment” is one layer of negation. “Patients will be excluded if no EGFR sensitizing mutation is found” is two layers deep, with opposite clinical implications. One word and a restructured clause change the meaning entirely. General models handle simple negation reasonably well. Nested clinical negation is where they produce confident, well-formatted answers that are wrong.

These are not theoretical risks. These are the kinds of errors that propagate into competitive landscapes, portfolio assessments, and board presentations without anyone catching them because the output looked right.

We tested this directly

Our data science team, led by Sanya Chetwani, recently published a peer-reviewed paper at EMNLP 2025, one of the top conferences in natural language processing. The paper introduced BIOPSY, an end-to-end pipeline we built for extracting structured biomarker intelligence from clinical text.

We benchmarked BIOPSY against GPT-4o, one of the most capable general-purpose models available, on 5,000 real-world oncology abstracts.

GPT-4o achieved an F1 score of 0.73. Our purpose-built pipeline scored 0.86.

We then tested the pipeline on 2,000 neuroscience abstracts without any additional training. It scored 0.87. GPT-4o scored 0.74 on the same dataset.

I want to be clear about what these numbers mean and what they do not mean. GPT-4o is an excellent model. It performed well in zero-shot conditions, indicating its general capability. But a 13-point gap on a clinical NLP task is significant. In practice, it is the difference between an output that is plausible and one that is precise enough to structure into a competitive landscape that leadership will use to make portfolio decisions.

That gap is why we invest in purpose-built architecture. Not because general AI is bad. Because the decisions our clients make require a level of precision that general models were not designed to deliver.

The four questions I ask every pharma team

When I meet with a new team evaluating their intelligence infrastructure, I ask four questions. They are not about AI. They are about the output.

1. Can you trace it?

When a number appears in your competitive landscape, can you click through to the source? Not a link to a search result. The actual registry record, abstract, or publication number was extracted from. If the answer is no, the landscape is built on trust, not evidence.

2. Is it normalized?

When two endpoints appear side by side, has someone reconciled the patient population, biomarker criteria, comparator arm, and response criteria? Or are the numbers sitting next to each other because they happen to share the same metric name? If the answer is the latter, the comparison is not defensible.

3. Is it current?

Not “was the answer generated recently?” Current means the underlying data layer is continuously updated. Trial registrations, regulatory milestones, endpoint disclosures, and conference readouts are structured into the landscape as they happen. Not at the next model refresh. Not when someone remembers to re-run the prompt.

4. Can the team use it together?

Competitive intelligence is not an individual exercise. CI, Medical Affairs, BD, Commercial Strategy, and Clinical Development all need to work from the same view. If one analyst built the landscape, the rest of the team should not need to reconstruct it from a blank prompt. They should be able to open the same view, inspect the same evidence, and see what changed since the last time they looked.

These four questions separate a general AI tool from a purpose-built intelligence platform. In my experience, most teams can answer yes to maybe one of them with their current setup. The organizations that can answer yes to all four are the ones making faster, more defensible decisions.

What this means for the market

I believe general-purpose AI will continue to get better. Models will get faster, more accurate, and better at handling specialized domains. That trajectory is real, and I respect it.

But I also believe that the pharma organizations making the best competitive decisions three years from now will not be the ones that waited for general models to close the gap. They will be the ones who invested in intelligence infrastructure designed for the specific constraints of their industry.

Purpose-built is not a limitation. It is a commitment to the precision that these decisions demand.

That is what we are building at Kognitic. A decision engine that structures the clinical trial landscape, normalizes published evidence, and delivers competitive views that travel from the analyst’s screen to the boardroom without losing their sourcing, their structure, or their defensibility.

The question is not whether AI can help pharma teams. It already does. The question is whether your intelligence infrastructure is built for the decisions that actually move portfolios.

---

Jaseem Mahmmdla is Co-founder and CEO of Kognitic, the Commercial Decision Engine for Life Sciences. He co-authored the BIOPSY paper with Sanya A. Chetwani, Technical Lead, Data Science.

Schedule a Landscape Audit. to see how Kognitic structures the competitive landscape in your therapeutic area.