Sanofi now describes itself, in its own boilerplate, as “an R&D driven, AI-powered biopharma company.” In 2023 it went further, declaring it wanted to be the first pharma company “powered by AI at scale” and launching an app, plai, to prove it.^1,2 The ambition is real and the technology is genuinely useful. But the language tells on itself. “AI-powered” describes how the engine runs. It says nothing about where the car is going.

That distinction used to be the whole point of pharma. “R&D-driven” was never really a claim about research; it was a claim about outcomes — new medicines, longer lives, diseases pushed back. Research was the visible means to a benefit patients could feel. AI does not carry that promise on its own. A patient is not helped by a model. They are helped by a medicine that arrives sooner, works better, or reaches them at all. When a company leads with the tool instead of the result, it has quietly swapped its purpose for its plumbing.

No patient has ever been cured by a company being AI-powered. They are cured by what the AI helped make possible.

The market is splitting in two

Read across the industry and you can see two ways of talking about the same technology. One group makes AI the identity. Sanofi went “all in.” And the agencies serving pharma have caught the same habit — arguably worse, because identity is what they sell. Real Chemistry, which posted double-digit revenue growth it credited to AI and analytics, says its mission is to “bring AI and ideas together” and puts named tools like Compliance Compass in the shop window.^5,6 Publicis Health leans on having trained a model on hundreds of FDA letters in days.⁷ Impressive plumbing, all of it — but a client does not hire a medcomms partner to admire its model. They hire it to move a brand, change a behaviour, or get a message past a regulator. When an agency’s pitch leads with the technology it owns rather than the outcome it delivers, it is making the client’s mistake on the client’s behalf.

The other group keeps the outcome in the subject of the sentence. Novartis frames AI as a “seven- to ten-year long game” embedded across the value chain — the benefit, not the buzzword, is the drug discovery at the end of it.³ AstraZeneca talks about AI in clinical trials, but the claim it makes is about reducing time-to-market.⁴ Pfizer points to computational design potentially yielding breakthrough molecules.³ In every case the technology is the means; the medicine is the message. That is the more honest grammar, and the more persuasive one.

Why “we use AI” is a weak differentiator

There is also a simple commercial problem with making AI your banner: everybody now has one. Seven of the largest drugmakers sit in the top tier of this year’s AI-maturity rankings.³ When every competitor, and every agency pitching them, claims to be AI-powered, the claim differentiates no one. It becomes table stakes dressed up as a strategy. A differentiator has to be something a rival cannot equally assert — a specific result, a faster trial, a molecule that would not otherwise exist. “We use the same general-purpose technology as everyone else” is not a position. It is a parity statement.

The deeper issue is that AI itself has already commoditised. The same handful of foundation models sit underneath nearly every “AI-powered” tool in the sector; the same APIs are a credit-card sign-up away for any rival or any agency. What was a marvel three years ago is now infrastructure — closer to electricity or broadband than to a proprietary edge. And we do not admire companies for being electricity-powered. We admire what they make with the power. Branding yourself around a capability the whole world can rent is not a flex; it is a tell that you have not yet found the thing only you can do. The wonder has moved on, and audiences sense it: announcing that you use AI now lands somewhere between obvious and slightly dated.

None of this is an argument against the work. AI genuinely shortens timelines, sharpens trial design, and catches compliance risk earlier. The argument is about where it sits in the story. Treat AI as the protagonist and you make a promise you cannot keep, to an audience — patients, prescribers, regulators — who care about none of it for its own sake. Treat it as the enabler, and the sentence finishes properly: AI is how we get medicines to people faster and more safely. That is a claim worth making, because it is a claim about them, not about us.

For the marketing and medcomms agencies serving this industry, the lesson is sharper still. Our job is to help clients sound like they know the difference. The agencies that win the next decade will not be the ones that shout loudest about their models. They will be the ones that can articulate, in plain human terms, what those models let a company finally do for the people it exists to serve.

References

1. Sanofi, “Sanofi ‘all in’ on artificial intelligence and data science to speed breakthroughs for patients,” press release, 13 June 2023. sanofi.com

2. MM+M, “Sanofi wants to become first pharma company ‘powered by AI at scale’.” mmm-online.com

3. IMD, “Future Readiness Indicator — Pharmaceuticals 2025” and related AI-maturity analysis (Novartis, Pfizer, sector rankings). imd.org

4. AI News, “How AstraZeneca dominates AI clinical trials in 2025.” artificialintelligence-news.com

5. Fierce Pharma, “Real Chemistry debuts AI tool to stay on top of FDA’s torrent of marketing letters.” fiercepharma.com

6. MM+M Agency 100 (2026), Real Chemistry — Medical Communications profile. mmm-online.com

7. Fierce Pharma, “How pharma marketers are using AI” (Publicis Health FDA-letter model). fiercepharma.com

Capability is climbing — unevenly

On raw benchmarks, the models have had a remarkable year. They now match or beat trained humans on PhD-level science and competition mathematics. The success rate of AI agents on real-world, multi-step tasks rose from 20% to 77% in twelve months; on cybersecurity problems, from 15% to 93%.

The same report shows the other half, and it matters more to anyone tempted to hand over a workflow. The frontier is jagged. The systems that win a maths olympiad still cannot reliably tell the time. Robots complete 12% of ordinary household tasks. The models stay weak at sustained multi-step planning and financial analysis — the connective, judgement-heavy work that holds a project together. The intelligence is real but spiky: brilliant in narrow columns, absent in the gaps.

The gap that should worry you

The finding I keep returning to is about people, not models. Generative AI reached 53% of the population in three years — faster than the PC or the internet — and the value is concentrating: the median value per user tripled between 2025 and 2026. Meanwhile the labour market is bending. Employment among software developers aged 22 to 25, among the most AI-exposed roles, has fallen nearly 20% since 2024, even as their older colleagues’ numbers grew.

The distance between people who have absorbed these tools and people who have not is widening, fast. In health communications it won’t arrive as redundancies. It will show as speed: one writer drafting in a morning what another takes three days to produce; one team running a literature scan over lunch while another books a fortnight of associate time. Within eighteen months those are not productivity nuances — they are different cost structures and different careers. If you lead a team, treat AI fluency as a funded, deliberate programme, not a hobby for the keen. If you are earlier in your career, treat it as the most valuable thing you teach yourself this year. Either way, leaving it to chance now carries a price.

In the clinic: time saved, evidence thin

The health findings carry the same double edge. AI scribes — drafting clinical notes from the conversation in the room — saw broad adoption in 2025, with physicians reporting up to 83% less time on notes and lower burnout. A real gain on a real problem.

Then the Index checks the evidence. Across more than 500 clinical AI studies, nearly half relied on exam-style questions rather than real patient data; only 5% used genuine clinical data. Capability is outrunning proof of safety and effectiveness in the messy settings we actually communicate about. For anyone who has to stand behind a claim, that gap between adoption and evidence is the whole job.

Can agents run the workflow? Not soon

The second link explains why. Atlan’s essay on the “enterprise context layer” is written for data leaders, but the argument is ours: to act safely, an agent needs three kinds of context — knowledge (what the terms mean), expertise (how the work is actually done, most of it never written down) and norms (what is allowed, by whom, under which approval). None of it lives in the model. It lives in your SOPs, your MLR rulebook, the client’s brand guidelines, and the heads of the people who remember the 2023 audit.

So agents running health communications workflows unaided is, for now, a fantasy — and the context layer shows why the timeline is long. Encoding that knowledge, expertise and those norms into something a machine can use is not a Friday-afternoon prompt. It is sustained organisational work: mining how the job is really done, versioning it, governing it, catching drift as positioning shifts, and keeping a named human accountable for every consequential rule. Even the vendor selling the fix concedes that humans stay in the loop “for judgment, certification, and exception handling.” In regulated health communications, judgement, certification and exception handling are the job — not the residue left once the agent finishes.

So the two reports leave us somewhere more useful than the hype. Capability is real but jagged. The context needed to make it act unaided is vast, undocumented and slow to build. The model is augmentation, and it will stay augmentation for years. The teams that win the next phase won’t be the ones that replaced associates with an autonomous agent; they will be the ones that made their people sharper and faster while patiently building the context layer underneath — ready to delegate, task by narrow task, on their own terms.

Get fluent, and get your people fluent, now — because the gap is widening. Build the context deliberately, because there is no shortcut. And keep a human between the model and anything a regulator will read — by design, not as a courtesy.

— Ned

References

• Stanford HAI. The 2026 AI Index Report. hai.stanford.edu/ai-index/2026-ai-index-report

• Lynch, S. “Inside the AI Index: 12 Takeaways from the 2026 Report.” Stanford HAI. hai.stanford.edu/news/inside-the-ai-index-12-takeaways

• Sankar, P. “What an Enterprise Context Layer Actually Is.” Atlan. atlan.com/know/what-is-the-enterprise-context-layer

• Brynjolfsson, E. et al. (2025) — source for the entry-level headcount figure cited in the 2026 AI Index.

Figures cited (agent task success 20→77.3% and cybersecurity 15→93%; household-robot success 12%; generative-AI adoption 53% in three years and median per-user value tripling 2025–26; developers aged 22–25 down ~20% since 2024; AI scribes up to 83% less note-writing time; only 5% of 500+ clinical-AI studies using real clinical data) are drawn from the 2026 AI Index.

There are two generations of AI adoption in health communications. Most agencies are in the first. Some are beginning the second.

Generation one is what almost everyone means when they say they are “using AI”:

• Open ChatGPT or Claude

• Type a prompt

• Get an output

• Review it, fix it, decide what to do next

• Repeat

This works. It saves time. It is genuinely AI. The outputs are real, the efficiency gains are real, and the agencies doing it well have built something valuable.

But in this model, you are still in the loop for every single decision. You set the goal, plan the steps, review the output, decide what comes next, and start again. The AI handles one step at a time. You handle everything else.

The person doing the reviewing, the sequencing, the fixing — they are functioning as the system’s prefrontal cortex. That work is invisible on a timesheet, but it is skilled labour. It is currently being performed by people paid to write, think, and advise.

Agentic AI is the second generation — and it changes that fundamental dynamic.

Instead of responding to prompts, an agent receives a goal and gets on with it. It plans the steps itself. It uses tools. It checks its own work. It reports back when it is finished — or when it genuinely needs a human decision.

The shift is not from “not AI” to “real AI.” It is from AI as a responsive tool to AI as an autonomous system. Generation one and generation two both matter. But they require different things from the people working with them.

What makes an agent different

An agent receives a goal, not a task. Then it gets on with it.

What generation two does that generation one doesn’t:

• Works out the steps itself — no prompt-by-prompt hand-holding

• Uses tools: searches the web, reads documents, writes to files, calls external systems

• Checks its own outputs and adapts when something fails or produces an unexpected result

• Reports back when it has finished — or when it genuinely needs a human decision

The difference isn’t technical. It’s about where the decision-making sits.

A prompt-and-response workflow:

• Human decides what to do

• Human writes the instruction

• AI executes one step

• Human reviews

• Human decides what to do next

• (Repeat indefinitely)

An agentic workflow:

• Human defines the goal and the guardrails

• Agent plans and executes the steps

• Agent flags genuine decision points

• Human reviews outputs and exceptions

• Agent handles the rest

Gartner predicts 40% of enterprise applications will include task-specific AI agents by end of 2026 — up from under 5% in 2025. The transition is already underway.

Who builds agents, and how

Agents are not a product you buy off the shelf. They are assembled from components — and understanding what those components are matters if you want to evaluate vendor claims or build anything yourself.

The foundation models — the brains of any agent:

• Claude (Sonnet, Opus, Haiku) by Anthropic — strong reasoning, long context, tool use

• GPT-4o / o3 by OpenAI — largest ecosystem, widest third-party support

• Gemini by Google DeepMind — deep integration with Google Workspace

• Open-source models (Llama, Mistral) — self-hostable; useful where data cannot leave your environment

The frameworks — the scaffolding that turns a model into an agent:

• LangGraph — the most widely used in production; models workflows as directed graphs

• CrewAI — role-based multi-agent crews; a working system in under 20 lines of code

• Claude Agent SDK by Anthropic — tool-use-first architecture; powers Claude Code

• OpenAI Agents SDK — released March 2025; production-grade, handoff-based architecture

All of these require a developer to set up. But the landscape is changing.

Can we build our own?

Yes — and the barrier is lower than most people assume.

Three realistic entry points:

1. No-code agent builders — tools like n8n, Zapier AI Agents, and MindStudio let you connect models to workflows without writing code. You describe what the agent should do; the platform handles the wiring. Suitable for: literature surveillance alerts, first-pass claims checking, content routing.

2. Low-code platforms with a health focus — Infinitus Studio launched in April 2026 as the first healthcare-specific no-code agent builder. Agents built on it are reportedly 40% more accurate and deployed up to 90% faster than manually developed systems.

3. Custom-built with developer support — a small agency with a developer or good technical partner can build purpose-specific agents using LangGraph or CrewAI in weeks, not months. The AI agent market hit $7.84 billion in 2025 and is on track for $52.62 billion by 2030. The ecosystem of specialist builders is growing fast.

What to watch for if you go this route:

• Compliance first — any agent handling unpublished data, patient-adjacent content, or regulatory submissions needs proper data governance. Most consumer no-code tools are not compliant by default.

• Start narrow — one structured workflow, not a general-purpose assistant. Claims checking before MLR preparation. Surveillance before content generation.

• Audit trails matter — for health communications specifically, you need to be able to show what the agent did and why. Choose tools that log decisions, not just outputs.

Why health communications is unusually well-positioned for this

Not every industry is ready for agentic AI at the same time. Health communications is, for one specific reason: the workflows are already structured.

The workflows that are ready to hand over:

• Claims checking — structured, verifiable, rule-governed; right or wrong

• Reference verification — binary correctness against a known source

• Literature surveillance — repeatable search and filter logic

• MLR preparation — document-heavy, defined inputs and outputs

• Content adaptation — clear rules for what counts as correct across formats

The MLR market is valued at $13.1 billion in 2025, projected to reach $27.1 billion by 2032. Early AI pilots show 50–65% reductions in regulatory submission timelines. Some companies currently run MLR review cycles of 50–60 days per piece of content. Agents compress that.

These are not creative problems. They are structured processes that currently require skilled labour partly because the tools to automate them safely did not exist.

Those tools are starting to exist now.

What will not be automated

An agent can verify that a claim is accurate. It cannot decide whether that claim is appropriate to make.

Those are different problems.

The judgements that stay with humans:

• Contextual authority — understanding the regulatory, commercial, and reputational environment in which a communication will land

• Scientific integrity — deciding how much uncertainty a physician can usefully hold when making a prescribing decision

• Audience reading — knowing what a room needs to hear, and what it does not

• Accountability — being the person whose name is on the submission, the advisory board output, the label claim

“AI will not replace MLR, but it will transform it into a faster, safer, and more strategic safeguard.”

Source: Pharmaphorum — Accelerating the MLR review leveraging AI

None of these can be broken into verifiable steps and handed to a system. They require professional judgement — and that judgement becomes more visible, not less, once the surrounding process work is handled.

The question that matters right now

The question is not whether your agency will use agents. It will.

The question is whether the people leading it understand enough about what agents can and cannot do to design the handoffs correctly.

What that means in practice:

• Knowing which workflows are structured enough to run autonomously

• Knowing where human review is a genuine quality check — and where it is just a comfort habit

• Understanding what regulatory accountability requires in terms of human sign-off

• Knowing how to brief an agent so its outputs are auditable, not just plausible

Most agencies are not asking these questions yet. That is not a criticism — the technology has moved faster than the governance frameworks. But the gap will close, and the organisations that close it intentionally will be in a different position from those that close it by accident.

This is the first post in a focus on agentic AI in health communications — covering what agents are, how to build them, and what the commercial landscape looks like. If you are working on this, or thinking about it, follow along. There will be a new section devoted to agentic solutions.

— Ned

#IrreplaceablesHealth

Sources

Gartner via DEV.to — 2025 was about chatbots, 2026 is about agents

GlobeNewswire — MLR Review Software Market Report 2026

Pharmaphorum — Accelerating the MLR review leveraging AI (Part 1)

Anthropic — Building Effective Agents

Lindy — AI Agents vs Chatbots

LangGraph — Best Multi-Agent Frameworks 2026

PEC Collective — AI Agent Frameworks Compared

Nirmitee — No-Code AI Agent Builders Healthcare

McKinsey — Agentic AI Advantage for Pharma

That is the honest starting point. What follows is not an argument that editors are irreplaceable. Some are. Some won’t be. The purpose here is to be specific about which is which — and what the people on the wrong side of that line can do about it.

What is actually being automated

The pre-MLR check is the clearest example. Tools are now scanning drafts before they reach human reviewers — flagging unapproved claims, checking references against source material, identifying consistency errors, tagging language that doesn’t match the approved claims library. Tasks that used to sit with a skilled editor for hours.

Indegene and Vodori are doing this at scale. Some pharma companies have reduced MLR submission timelines by 50–65% through AI-enabled workflow redesign. The content volume coming through those workflows is up 29% year-on-year. Headcount is not keeping pace.

This is not speculation. This is current operations at the companies that employ most of the people reading this.

Emma Hyland at the Veeva Commercial Summit on AI, content volume, and what it means for review. 1:26.

The work being automated is the mechanical layer: grammar, consistency, style guide compliance, reference checking, claim-flagging against an approved list. It is also, for many copy editors, the majority of their billable hours.

Tools doing this work now: Indegene AI for MLR · Vodori Pepper Flow · Veeva Vault PromoMats · Writer.com · Grammarly Business

What is not being automated

The layer that survives is judgment at the point of consequence — where being wrong has a real cost.

In MedComms, that means: Does this claim hold up against the clinical data? Is this the right framing for this efficacy result with this audience? Is the language compliant not in a generic sense, but in the specific regulatory context of this product in this market?

AI tools hallucinate. In a general context, a hallucinated fact is embarrassing. In a pharma communications context, it is a material risk. The failure modes are bad enough, the accountability clear enough, that human judgment stays in the loop — and will for the foreseeable future.

There is also the layer that is harder to articulate but equally real: knowing when to break the rule. AI applies rules. It does not transcend them. It produces averaged prose. A voice that belongs to someone — a writer, a brand, a scientific communicator with a particular way of handling uncertainty — is something AI consistently flattens. The distinction matters in any piece of communications that is trying to do something other than merely comply.

The Chartered Institute of Editing and Proofreading has developed a practical framework for mapping which editorial tasks AI can and cannot do reliably. It is worth reading in full.

“Will AI Take Our Editing & Proofreading Jobs?” — a balanced 3-minute take on where the real threat sits and where it doesn’t. 3:03.

The pivot: four paths, none of them easy

There will be fewer copy editors in this industry in ten years than there are now.

The Bureau of Labor Statistics projects 1% growth in editor roles through 2034. Before you set that number down: it’s wrong for this industry, and it’s probably wrong for most. BLS projections are built from employer surveys and historical employment trends — rearview-mirror statistics by design. They didn’t predict the collapse of newspaper newsroom employment, which shed more than half its workforce between 2008 and 2020. They don’t incorporate the AI displacement analysis that Goldman Sachs, McKinsey, and Oxford Economics have been publishing since 2023 — all of which identify content production and editorial tasks as among the most exposed to automation. And they lump “editors” into a category so broad — book editors, magazine editors, podcast producers, food bloggers, corporate communications writers — that it captures nothing useful about any of them. The number is measuring a different profession with a methodology built for a different era.

The signals from inside this one point the other way. Major pharma companies cut more than 22,000 roles in 2025 — and that was before the £300bn patent cliff fully hits. When clients cut headcount they cut agency spend, and agency spend cuts fall on production roles first. When internal pharma teams adopt AI tools for content production, agency spend drops 25–70%. The AMWA is running webinars on how medical writing careers are shifting. They wouldn’t bother if people weren’t anxious.

But the people who remain in this one will be doing work that matters more. And the people who move now will be better placed than those who wait. Here are the four pivots that are real.

1. The validator role

As AI generates first drafts at scale, someone has to own the quality framework — not fixing individual sentences, but designing and maintaining the standards AI output is checked against. This is an editorial function. It requires everything a good editor already knows, expressed as governance rather than a red pen. The AI-aware editor is a role being created right now inside large pharma content teams. It is not a downgrade. It is a different application of the same expertise.

2. The prompt architect

Writing the instructions that generate consistent, on-brand, compliant first drafts. Editors already know what good looks like — this is expressing that knowledge as a system rather than a correction. It is a closer translation of existing editorial skills than it first appears. Tredence’s prompt engineering overview is a useful primer; skip the generic sections and focus on domain-specific prompt design. The people who will do this well in MedComms are editors, not engineers.

3. The scientific editorial specialist

The intersection of scientific accuracy and regulatory compliance is the one layer that gets harder to automate, not easier. If you can read a clinical paper, understand what it does and does not show, and judge whether a piece of communications faithfully represents it — that skill is worth more per head now than it was two years ago. Envision Pharma Group’s 2026 analysis is specific about where human scientific judgment remains critical in the AI-augmented workflow. There will be fewer of these roles than there were general copy editors. They will be better paid and harder to automate.

4. Content strategy

Editorial judgment applied upstream — to which content should exist, not whether it’s correct. Harder to transition into without a track record, but the natural destination for senior editors whose production role is being compressed. The Digital Content Next analysis of AI and editorial authority is clear on how the upstream role is being reframed in professional publishing. The MedComms version of this is not yet well-defined, which is both the risk and the opportunity.

The honest closing

Most writing about editors and AI tries to reassure. This is not that.

The mechanical layer is going. The judgment layer is not. But there is less room for people who live only in the mechanical layer, and the transition is happening faster than most editorial teams are acknowledging internally.

The question for anyone in this profession right now is not whether AI is a threat. It is which part of your skill set lands on the right side of the line — and what you are building toward the other side.

Further reading and tools

On the MLR workflow shift

• Vodori: The future of AI in MLR review

• Pharmaphorum: Is AI the missing link in fixing MLR?

• Indegene: AI for MLR excellence

• Envision Pharma Group: 10 AI game-changers for pharma and MedComms in 2026

On what editing actually is — and what AI can’t do

• CIEP: What the future of AI means for editors and proofreaders

• Inkbot Editing: Could AI ever replace copyeditors?

• Jane Friedman: The hidden costs of AI copyediting tools

On the pivot

• ePublishing: The rise of the AI-aware editor

• Tredence: Prompt engineering skills guide

• Digital Content Next: How AI reshapes editorial authority

• Goldman Sachs: Generative AI could raise global GDP by 7%

Yang’s forecast is blunt: somewhere between 20 and 50 per cent of America’s 70 million white-collar workers displaced in the next several years. Not eventually. Now. A CEO he spoke to described three tranches of redundancies — 15 per cent now, another 20 per cent in two years, another 20 after that — as if he were announcing a building refurbishment. Sell anything that consists of people sitting at a desk looking at a computer, one investor told him. The stock market will reward headcount cuts and punish the companies that don’t follow suit.

His advice, in the end, is practical to the point of being bleak: cut your expenses, build your savings, make a plan, because a lot of people are about to lose their jobs and you could easily be one of them.

I can already hear the health communications response to this. We’re different. We’re regulated. You can’t just run AI output through MLR and call it a medical claim. Regulatory agencies have views. HCPs can tell when the science has been garbled. The structural constraints of this industry are real, and they will slow things down.

This is true. It is not sufficient.

I made a version of this argument in the first essay I published here, which was about the framework knitters who called themselves Luddites. They, too, had quality controls: guild structures, expert buyers who understood the craft, customers who had complained about inferior product before. None of it was enough, because the economics changed who was making purchasing decisions and what they were optimising for. The market for hand-finished hosiery didn’t disappear — it shrank, and the people serving it had to accept that the volume work had moved permanently somewhere else.

The regulatory apparatus of health communications is a more robust quality control than the hosiery guild. MLR is a forcing function that doesn’t yield to market pressure in the way that consumer goods can. But watch what’s happening at the client side. Large pharma companies are signing enterprise AI agreements — Bristol-Myers Squibb’s deal with Anthropic being the most visible recent example — and framing them as infrastructure investments. When AI becomes infrastructure, it changes the procurement question. Not “is this the best possible output?” but “given what we can now produce internally, what exactly are we paying agencies for?”

That’s not a regulatory question. It’s an organisational one. And organisations tend to answer it with headcount reduction.

The specific character of the Fuckening in health communications is that it will arrive on delay, and then arrive all at once.

The regulatory layer means you won’t get a letter this quarter. You’ll get two or three years of enterprise AI getting embedded in client workflows, junior roles quietly not being backfilled, scope creep in what “AI-assisted” means on a deliverable, and then a procurement conversation that has already concluded before you were invited to it. Yang’s CEO describes three tranches of cuts because the maths only becomes undeniable in stages. The regulated nature of pharma communications means the stages will be longer. It does not mean there won’t be stages.

The Luddites didn’t see the end coming because the early machines were genuinely inferior — and the early objections were genuinely valid. By the time the machines were good enough, the infrastructure to deploy them was already in place, and the economic logic was irreversible.

So what would Yang advise, adapted to this specific professional context?

He would say: cut your expenses and build your savings now, while you still have income and while the role still exists in its current form. Not because you’ll definitely lose your job, but because the people who survive disruptions of this scale are the ones who had optionality — who could take a risk, turn down a bad deal, wait for the right thing — rather than the ones who needed next month’s salary the moment it was threatened.

He would say: make yourself genuinely useful in the parts of the work that AI cannot replicate. Not “AI can’t replace expertise” as a general reassurance — that’s the quality argument the framework knitters made, and it’s true and insufficient in equal measure. Specifically useful: you know what a medical director is actually worried about; you know how an MHRA reviewer reads a data table; you know what the trial results actually imply as opposed to what they appear to imply. That’s the supervisory and judgment layer. It’s where the premium will concentrate.

And he would say — and this is the part that the reassuring voices in health communications tend to skip — that being right about the quality argument will not protect you from the market’s eventual indifference to it. The weavers were right. It didn’t save Nottinghamshire.

The Fuckening will come for health communications. It will take a little longer than it takes for tech and finance and marketing. Use the time.

— Ned

Sources

[1] Andrew Yang — “The End of the Office”, February 2026

[2] The Irreplaceables — “The Weavers Were Right”

[3] Bristol-Myers Squibb press release — BMS Announces Strategic Agreement with Anthropic

History has not been kind to them. "Luddite" is now shorthand for someone who fears technology out of ignorance or sentiment. The actual Luddites were neither ignorant nor sentimental. They were skilled framework knitters who understood exactly what the new machinery would do, and they were largely right.

The frames didn't just make stockings faster. They made them cheaper, and in doing so they destroyed the economic basis of an entire craft. Not gradually, over generations, in a way that allowed for adjustment. In a decade. The Luddites' mistake wasn't misunderstanding the technology. It was believing that being right about the quality argument would protect them from the market's indifference to quality arguments.

Hand-knitted hosiery was, in fact, better. Machine-produced hosiery was cheaper. Cheaper won.

I think about this a lot when I look at what's happening in health communications.

The standard reassurance you'll hear — and I've heard it, and perhaps offered versions of it myself — goes something like this: AI can produce text, but it can't replace expertise. It doesn't understand regulatory context. It can't navigate MLR. It doesn't know what a medical director is actually worried about when they push back on a claim. The deep craft of health comms is safe because it's the deep craft that matters.

This is probably true. It is not necessarily reassuring.

The framework knitters were also in possession of deep craft. They had spent years — apprenticeships, journeymanship, mastery — developing an understanding of materials, tension, pattern and structure that no machine of 1811 could replicate. They were right that machine-knitted hosiery was inferior in construction. The market found that it didn't care enough about the difference to pay for it.

The question isn't whether your expertise is real. It is. The question is whether the people buying communications services will be able to tell the difference between AI-assisted output and expertly crafted output — and whether, if they can, they'll be willing to pay the premium.

Here is where I think the parallel gets uncomfortable.

Health communications has, for structural reasons, always had quality controls that the hosiery market lacked. MLR is a forcing function. Regulatory agencies have views. HCPs can tell when something is clinically imprecise. These are real constraints that don't disappear because AI can generate plausible-sounding sentences.

But the Luddites had quality controls too, in their own way. Guild structures. Expert buyers who knew the difference. Customers who'd complained about inferior product before. None of it was sufficient, because the new economics changed who was doing the buying and what they were optimising for.

Watch what's happening at the client side. Large pharma companies are signing enterprise AI deals — Bristol-Myers Squibb's Claude deployment being just the most recent visible example — and framing them as infrastructure investments. When AI becomes infrastructure, it gets embedded in procurement decisions. The person buying communications services starts asking a different question: not "is this the best possible output?" but "given what we can produce with our internal AI tools, what are we actually paying agencies for?"

That's not a technology question. It's an organisational one. And it's the one the framework knitters didn't see coming until it was too late.

So what did survive?

Not the craft as it was practised. The hand-knitters who continued to produce high-end, hand-finished hosiery found a market — a smaller, premium market that explicitly valued what machines couldn't replicate. Some framework knitters moved into supervisory roles in the factories. Some moved into quality control and inspection. Some moved into design, where the judgment layer above the machinery remained human.

None of these were the same job. They were adjacent jobs that preserved some of the expertise while accepting that the core production economics had changed permanently.

For health communications professionals, I'd read that as three things worth paying attention to.

The premium market for craft exists, but it's smaller than the current market for competent execution. If your value proposition is "we write better than AI," you need to be operating in contexts where that difference is detectable and valued. Specialist therapeutics. Genuinely complex science. Regulatory submissions where a wrong word has consequences. Not everything qualifies.

The supervisory and quality layer is real and will grow. Someone has to know what good looks like in order to catch what bad looks like. That's a genuine expertise premium — but it's a different role than the one most senior health comms professionals currently occupy. It requires knowing the territory of AI failure modes as well as you know the territory of clinical science.

The judgment layer — the part that requires knowing what a medical director is actually worried about, or what an MHRA reviewer will focus on, or what the trial data actually says as opposed to what it appears to say — is genuinely defensible. For now. The honest qualifier is "for now."

The Luddites lost, but they weren't wrong. The Industrial Revolution did eventually create more jobs than it destroyed — but not in Nottinghamshire, not in that generation, not for those workers. The economists who point to net job creation as evidence that the Luddites were mistaken are, with respect, answering a different question than the one the Luddites were asking.

The question the framework knitters were asking was: what happens to us?

That's the question worth sitting with. Not "will health communications exist in ten years?" It will. Not "will there be work for people who understand clinical science and regulatory context?" There will. But in what form, at what volume, and compensating what skills?

The weavers were right about the quality argument. They were right that their craft was superior. They were wrong that being right about that would be enough.

I find that clarifying rather than depressing. Not because the answer is reassuring — it isn't, particularly — but because it's at least honest about what the question is.

That's what this newsletter is for.

— Ned

Sources

[1] Smithsonian Magazine — What the Luddites Really Fought Against

[2] IntuitionLabs — Big Pharma & Hyperscaler AI Deals: 2026 Industry Tracker

[3] Anthropic — Claude for Life Sciences

[4] Bristol-Myers Squibb press release — BMS Announces Strategic Agreement with Anthropic

[5] Knowable Magazine — What happens to the weavers? Lessons for AI from the Industrial Revolution