There is a number that has been circulating in Indian pharma boardrooms for about eighteen months now. It says, in various forms, that roughly half of Indian pharmaceutical companies are still at proof-of-concept stage with AI. The number is approximate, the methodology is informal, and the boundary between proof-of-concept and production is contested. But the directional reading holds: by the end of 2025, in the country that supplies twenty percent of the global generics market, AI adoption is somewhere near the middle of the curve. Not at the start. Not at the finish. Mid-way.
The phrase "mid-way through" is doing some real work in that sentence. It means three different things at once, and the boardrooms that are most useful to be in are the ones that hold all three at the same time.
It means a meaningful number of Indian pharma functions are running AI in production today. Not in slide decks. Not in pilots that have been alive for eighteen months. In production. Generative models drafting first-pass responses to FDA queries. Foundation-model-assisted case triage in pharmacovigilance. Retrieval over the company's own SOPs, queryable in plain language by regulatory reviewers. These are not edge cases; they are the most boring AI workflows in the world; they are also the most expensive ones we could possibly automate, and they ship.
It also means a much larger number of Indian pharma functions have started a pilot, run it for six to nine months, learned what they would do differently, and stalled before going to production. The stall is rarely about the model. It is about the data, the governance, the validation, the MLOps, the procurement process, the tolerance the function has for an AI failure that lands on the regulator's desk. The stall is mid-way honestly because the next step is harder than the first one, not easier.
And it means a third set of functions, in the same companies, have not started. Manufacturing yield prediction at the formulation stage. Pharmacovigilance signal detection across multilingual case narratives. Real-time deviation classification in batch records. These are workloads where AI would matter, and where the engineering and data work to get there has not begun.
Why the framing matters
The reason "mid-way through" is useful, rather than just descriptive, is that it predicts what the next twelve months will be about. A market at the start of adoption is dominated by demonstrations: can the model do this at all. A market at the end of adoption is dominated by competition: who deploys this better than whom. A market in the middle is dominated by something quieter and more uncomfortable, which is the work of getting the second half of pilots to production.
That work is engineering work, not vendor-selection work. It is the work of MLOps discipline, evaluation infrastructure, data governance, human-in-the-loop design, and the patient negotiation between a function head and her IT counterpart over what counts as "ready to ship." It does not have a single named buyer. The buyer is the Head of Regulatory and the Head of IT and the Chief Information Officer and the Head of Quality, sitting in a room together, deciding whether a system that has worked for nine months in a sandbox can be allowed to take a single byte of customer data.
That room is the room AIforPharma is being convened to host.
The stall is mid-way honestly because the next step is harder than the first one, not easier.
— The view from the room we are trying to host.
Three forces that make 2026 different
If 2024 and 2025 were the years of "can AI do this?" inside Indian pharma, 2026 looks different on three fronts. None of these is a prediction in the marketing sense. Each is an observation about what has already shifted, that the second half of the curve is being held back by, and that the next twelve months are likely to test.
One: foundation models are now good enough for the regulatory document
The single largest unlock of the past two years, for Indian pharma specifically, is that the current generation of frontier language models, GPT, Claude, Gemini, are good enough at reading and drafting against pharmaceutical regulatory documents that the gap between "model output" and "the version a regulator would accept" is now a review problem, not a generation problem. This is a quiet but enormous change. A year ago, the bottleneck was the model. Today, the bottleneck is the human review loop, the eval, the citation discipline, the audit trail.
The implication is concrete: the procurement conversation that mattered in 2024 (which model, which vendor, what does it cost per token) matters considerably less in 2026 than the engineering conversation (what is our eval set, how do we route uncertain cases, how do we trace every claim back to a source document). Indian pharma teams that have figured this out are already moving from proof-of-concept to production. Indian pharma teams that are still negotiating model contracts are running the wrong meeting.
Two: the governance question moves from "can we?" to "how do we?"
In 2024, the dominant governance question inside Indian pharma was binary. Can we use a hosted foundation model on regulatory data? Can we put pharmacovigilance case narratives through an AI system? Can we let an LLM draft a response to a CDSCO query? The answers were inherited from US and EU peers, layered with local interpretation, and almost always conservative.
In 2026, the question is no longer binary. CDSCO has begun issuing guidance on digital tools in clinical research; the FDA's draft and final guidance on the use of AI in drug development is being read carefully here; the EU AI Act's implications for clinical AI systems are being mapped against existing pharmacovigilance and GxP frameworks. The governance question is not "can we?" but "how do we?" How do we document the eval set? How do we treat human-in-the-loop versus fully autonomous? How does the audit trail satisfy a 21 CFR Part 11 review? These are answerable questions, and the companies that answer them best will be deploying AI in places their peers are still asking permission to enter.
This is exactly the shift that the consortium is intended to host. The "can we?" conversation could happen between a pharma compliance team and an external auditor. The "how do we?" conversation requires the pharma function head, the AI engineer who built the system, the quality team, and (sometimes) the regulator herself, all in the same room.
Three: the procurement question moves from IT to function
There is one structural change in Indian pharma's AI adoption that is undersold and important. AI procurement has, until recently, been an IT decision. The IT organisation evaluated platforms, negotiated contracts, selected tools, and rolled them out to functions. This worked when AI was infrastructure. It does not work, anymore, when AI is workflow.
What has changed is that AI workloads with the highest production value are workflow-shaped, not infrastructure-shaped. They live inside a specific function (regulatory, pharmacovigilance, medical affairs, manufacturing), they use that function's domain language, they require that function's domain experts to evaluate and improve them, and they are deployed against that function's KPIs. The IT organisation is necessary but not sufficient. The Head of Regulatory or Head of Pharmacovigilance is making the decisions that used to be made by the Head of IT.
This is part of the reason the AIforPharma application asks the applicant which function they lead, and what AI question is on their desk right now. The function is the unit of decision, and the room we are building reflects that.
What it means for the next twelve months
Putting the three forces together, here is what we expect the next twelve months to look like inside the kinds of Indian pharma companies we are building the room for.
First, the gap between leading and lagging companies will widen, faster, in 2026 than it did in 2024 or 2025. Leading companies have already crossed the production-engineering threshold for one or two workloads; the next workloads are easier because the platform and the eval discipline already exist. Lagging companies, the ones still negotiating their first procurement contract, are starting from zero on a curve that is now steeper.
Second, the most useful conversations will not happen at large industry conferences. They will happen between specific function leaders at specific companies who are running specific workloads against specific datasets. The pattern that closes the gap is peer-to-peer, narrow, and recurring. Conferences broadcast; consortia convene; the difference matters more in a market mid-way through adoption than in a market at the start.
Third, the questions that get asked will get more concrete. A year ago the question was "how can we use AI?" Today it is "how do we get our ANDA response system to handle the cases it currently routes to human review?" That question has a real answer; the answer is engineering work and eval work; the engineering work and eval work has been done somewhere else by someone the asker has not yet met. Closing that gap, deliberately, is the unit of value the consortium is designed to produce.
Why we are convening this
AIforPharma is being convened in 2026 because the structure of the conversation has changed and the existing rooms have not changed with it. Pharma trade associations were built for a different cadence and a different mandate. Tech conferences host a much larger and much less specific audience. Vendor user-group meetings host the function but never the regulator or the peer competitor. None of those rooms is wrong; none of them is the room a Head of Regulatory needs to be in when she is trying to deploy a generative AI system into her submission workflow next quarter.
We are convening because that specific room does not yet exist in India, because we have been having most of these conversations bilaterally for a year, and because the consortium model, three to four editions a year, single-day, single-room, by application, is the right size and the right cadence for the conversation that 2026 will turn out to need.
The first edition is forthcoming. The application is open. The room is small on purpose.