Two conversations happened at NEXT Pharma 2026. Only one was on stage.

8 June 2026 | London

NEXT Pharma 2026 wrapped in Dubrovnik recently. 400 attendees, 60 speakers, two days of stages, plus the table-side conversations that often matter more than the keynotes. I want to talk about what's changed since last year, and what hasn't.

The conversation has moved on

A year ago, the room was asking whether AI belonged in pharma at all. The agenda this year had a different shape. Sessions ran on agentic workflows, scaling AI from pilot to production, the human and AI sweet spot, and the build, buy, or partner question. There was a packed session titled "Why 95% of AI Projects Fail" and another on AI accountability. The framing has shifted from "should we" to "how do we, and who owns what".

That's a real maturation. The technology is being treated as part of how the work gets done, which means the harder questions are surfacing. Who's responsible when the model is wrong? Whose job changes? Does this scale across markets, or only inside one team? These are the right questions, and the industry deserves credit for asking them honestly.

Underneath, the older problem kept coming up

Sit at the stand for two days and the same line surfaces in different rooms: we can barely get the content we have out the door, never mind add new content from new tools. Teams are looking at agentic AI as a way to multiply content velocity, and that's a good instinct in theory. In practice, a lot of teams haven't yet mapped where their existing bottleneck is.

Adding speed to a queue doesn't move the queue. It just lengthens the line of things waiting for the same gate to open. In pharma, that gate is medical, legal and regulatory review. Until that gate widens, every new content tool adds to the pile in front of it.

This came up in our session and independently in others. Many of the people I spoke to had not yet asked themselves the foundational question: where exactly does our MLR process slow down, why does it slow down, and what does it cost? Before adding any new tool, that's the question worth doing the work to answer.

Build, buy, or partner

One of the most-discussed sessions of the conference was the three-sided debate on this. My read on each:

Local builds

Teams using Claude Cowork, live-coding tools, and internal AI assistants to build their own dashboards and small productivity utilities. This is good. It's the modern equivalent of the colleague who built the brilliant KPI sheet that everyone used. Those tools die when the person moves on, because nobody else maintains them. Fine for local productivity, dangerous for anything critical.

Global builds

Pharma companies building their own AI platforms at enterprise scale. The pull is understandable: control, customisation, no vendor lock-in. The cost is that you've just turned a pharma company into a software company. That requires a roadmap, a maintenance team, version control, regulatory expertise across the markets the platform operates in, and a willingness to keep up with a fast-moving model landscape. Pharma is not set up for that, and it shouldn't try to be.

Partner

The third path that doesn't get as much airtime as it deserves. A partner builds and maintains the infrastructure; the customer brings the operating context. The tension here is real. Software needs to be adaptable enough to fit how a customer actually works, while staying coherent enough to deliver the benefit of scale. Tip the balance one way and you end up with bespoke versions of the same product for every customer, with nothing improving at the platform level. Tip the other way and you build something so rigid it doesn't fit any one customer well.

For critical infrastructure like MLR, the partner model is the one that makes sense. The work doesn't get less serious over time, and no internal team is going to keep a tool current with the ABPI Code, the PMCPA, evolving local regulations, and the model layer underneath. That's a full-time job for a specialised team.

The audience gap nobody talked about

One observation that didn't make the agenda. The crowd at NEXT is not a baseline. It's an enriched crowd: people who are curious about AI, comfortable with AI, often already using the best tools available. Their colleagues at home, in many cases, have not touched any of it.

The change management story is bigger than the technology story. New tools only work if everyone uses them, not just the enthusiasts. Which is also why critical-path tasks like MLR are a useful place to apply AI. Everyone has to do MLR anyway. If you make the process more frictionless, the onboarding threshold is low. You don't need a behavioural change to get adoption. The work pulls people in.

What I'm hoping to see at NEXT 2027

Less conceptual. More what worked, what didn't, what jobs actually changed. A year from now, a lot of these AI platforms will be embedded. I'd like to hear about the outcomes, the staffing shifts, the friction between global and local that AI either resolved or made worse. That's the conversation that moves the industry forward.

Fix the foundation first. Everything else compounds from there. That's the work MLRcle is built for.