The regulators have started writing one rulebook, not five

https://www.europeanpharmaceuticalreview.com/news/ema-and-fda-issue-joint-ai-guidance-for-medicine-development/270259.article

The EMA and FDA put out ten joint Principles for Good AI Practice in January, and this month the EMA and HMA published their 2025 AI observatory report to steer the 2026–2028 workplan. The substance for health communications isn’t the principles themselves; it’s that the two largest regulators are converging on a single vocabulary for how AI-touched evidence must be described and monitored. When you write up anything a model helped generate, the standard you’ll be held to is being standardised — better read now than met for the first time in a query letter.

The hyperscalers are moving into the basement

https://nvidianews.nvidia.com/news/nvidia-and-lilly-announce-co-innovation-lab-to-reinvent-drug-discovery-in-the-age-of-ai

Lilly and NVIDIA announced a co-innovation lab to rebuild drug discovery around AI; TCS and Anthropic signed a global partnership the same fortnight. Read it as a forecast of your brief queue. The “AI-accelerated discovery” story is about to be load-bearing in a lot of pipeline communications, and the gap between what the model actually did and what the headline implies is exactly where a careful writer earns their keep. Compressing the first mile is real; it is not the same as finishing the race faster.

Patient-facing accuracy still isn’t there for the messy diseases

https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2026.1847603/full

A June study benchmarked leading models on patient questions about heart failure and cardiomyopathy and found their answers not yet dependable for clinically appropriate, comprehensible guidance on heterogeneous conditions. This is the unglamorous counter to every “just put a chatbot on it” proposal: the harder a disease is to generalise, the worse the model does — and those are precisely the conditions patients search most anxiously. Keep a human between the model and anything a patient reads.

The audience is being rerouted before it reaches your page

https://upgrowth.in/google-ai-overviews-healthcare-traffic-data/

AI Overviews now appear on roughly half of health searches, with click-through down around 60 per cent and some medical publishers reporting 70 per cent traffic drops on the pages an overview answers directly. For any client who owns clinical content, organic reach is being quietly restructured: the prize is no longer ranking first, it’s being the source the AI answer cites. If your content strategy still assumes the click, it is planning for a web that is receding.

The UK built a place to test the things before they ship

https://www.gov.uk/government/news/mhra-expands-ai-airlock-programme-with-a-36-million-funding-boost-over-three-years

The MHRA has put a further £3.6 million over three years into its AI Airlock, the supervised sandbox for AI as a medical device. It won’t touch most promotional work directly, but it’s worth knowing for two reasons: it’s where the evidence that shapes the next round of guidance is being generated, and it gives you a concrete, government-run answer when a client asks “where has this actually been tested?” — instead of pointing at the vendor’s own deck.

That’s it for this edition. Back Wednesday.

— Ned

Follow the conversation: #IrreplaceablesHealth

I keep turning that over because of what sits on the other side of it. The same class of model the agency is pointing at our copy is the one increasingly drafting it. And we know how that model behaves when left alone with a claim: it writes the efficacy beautifully and treats the risk language as something to be got through. Trained on years of marketing, it produces benefit-heavy copy, and if it is optimising for anything it learns that the safety information is the part nobody reads — so it trims it, softens it, moves it down the page (improvado.io). Fair balance is precisely the thing the machine is worst at, because fair balance is a judgement about proportion, not a sentence it can complete.

So here is the shape of the year. The instrument that writes the overstatement and the instrument that catches it are converging on the same architecture. One side drafts the imbalance; the other reads for it. And in the middle is a medical writer who, on a tight Friday, is tempted to let the first machine do more and check the result less.

The line I’d put on the wall is the one the FDA has held for decades and restated without blinking: the company is responsible for the promotional material, regardless of how it was made. There is no AI lane in the regulation, no allowance for “the tool drafted it.” The model is not a co-author you can name in your defence. Accountability did not move an inch when the drafting got faster — it still ends with a human signature, and now that signature is being read back by something that doesn’t get tired at four o’clock.

I don’t take this as a reason to keep AI away from promotional copy. I take it as the clearest argument yet for where our value actually sits. If the regulator has automated the first pass of review, then the fair-balance judgement — the proportion, the placement, the risk carried in the same breath as the benefit — stops being the boring bit at the end of the job. It is the bit the machines on both sides are worst at, which makes it unmistakably ours.

The writers who get nervous about this are the ones who have quietly let the safety information become the thing they do last and fastest. The ones who’ll be fine are the ones who already treat it as the thing they do most carefully. The audit was always coming. It just arrived as software, and it reads every line.

— Ned

Follow the conversation: #IrreplaceablesHealth

Source links:

• FDA’s AI-powered crackdown on alleged deceptive drug promotions: https://natlawreview.com/article/fdas-ai-powered-crackdown-alleged-deceptive-drug-promotions

• Pharma ad compliance 2026 — fair balance and ISI: https://improvado.io/blog/pharma-ad-compliance-fda-ftc-fair-balance-and-isi-requirements

• AI-generated pharma content and FDA responsibility principle: https://blog.madebyxds.com/ai-generated-pharma-content-fda-compliance-2026

The agentic MLR reviewer Veeva bought rather than built

What it is: A tool — Veeva Falcon MLR, the agentic review platform Veeva announced on 23 June after acquiring Copli, the company that pioneered it (veeva.com).

Why it’s worth your time: This is a step beyond the pre-checks we’ve covered before. Where the earlier Quick Check Agent flagged issues for a human to weigh, Falcon is pitched as running the review itself — checking promotional and medical materials against the approved label and local regulations — and Veeva is putting a number on it: the potential to remove 70 per cent or more of manual MLR labour within five years. Whether or not that figure lands, the direction is the thing to register. If your review cycle is your bottleneck — and for most teams it is — the vendor that already owns your content stack now intends to automate the slowest part of it. Better understood before it turns up in a release note.

How to use it:

• If you’re on PromoMats, get on the 9 July webinar Veeva is running to show it, and treat it as competitive intelligence whether or not you ever buy.

• Pressure-test the 70 per cent claim against your own mix: an agent will do most for high-volume, low-variation assets and least for the nuanced, claim-heavy pieces where the judgement actually lives.

• Decide now, on paper, what a human reviewer still signs even when the agent comes back clean — before the efficiency case decides it for you.

Watch out for: “Agentic” review automates the checking, not the accountability. A clean Falcon pass is not regulatory clearance, and the named reviewer is still the named reviewer. Treat the labour saving as real and the sign-off as non-negotiable.

A fair-balance pre-check you can run before the file leaves your desk

What it is: A prompt routine — a self-audit you run on your own draft that sets benefit against risk before it goes anywhere near review.

Why it’s worth your time: With the FDA now using AI to screen ads for exactly this failure (see From the Floor), the cheapest insurance is to read your own copy the way the regulator’s tool will. Generative drafts skew efficacy-heavy and quietly under-weight safety language; a structured pass catches the imbalance while it’s still yours to fix. Run against a near-final piece this week, it surfaced an ISI that had drifted below the fold and a benefit claim with no risk in the same eyeline — both things a reviewer would have bounced.

How to use it:

• Paste the draft and ask, in one instruction: “List every efficacy or benefit claim and every risk or safety statement in two columns. Flag any benefit without an adjacent risk, and any place the safety information is less prominent than the claim it qualifies.”

• Ask it to rank the three weakest fair-balance moments and explain why — not to rewrite them. You want the diagnosis; the rewrite stays yours.

• Keep the rule in front of it: the current FDA/FTC line is that risk must carry equal prominence to benefit, in the piece itself, not behind a “see more” or a linked page (improvado.io).

Watch out for: This finds imbalance; it does not certify balance. A model that helped write the copy is not a neutral judge of it, and a clean self-check is a reason to submit with confidence — never a reason to skip MLR.

One to watch your vendor build; one to run yourself before Friday.

Follow the conversation: #IrreplaceablesHealth

A fresh TGaS Advisors brief finds that 100% of surveyed organisations now use generative AI, yet most still lack the governance, data foundations and operational discipline to move past pilots — its president calls the technology “table stakes.” For health communications, this is the quiet end of the adoption race: if everyone has the same tool, owning it differentiates no one. What separates teams now is the part the survey calls execution — judgement, governance and the ability to show value — which is exactly the part you can’t buy off the shelf.

Pharma arrives at Cannes ready to show its AI “receipts”

After several years of panels speculating about what AI might do, this year’s festival reportedly pushes brands and agencies to show what they actually built — including, more revealingly, where it didn’t work. One Pharma Lions juror notes that “humanity is unexpected” emerged as a standout theme across the case films. The signal for our field is that the credential is shifting from having an AI story to having an honest one; being able to say what failed is becoming the more persuasive position.

Is AI the fix for MLR, or the reason it breaks?

Promotional content volumes are climbing roughly 29% year on year while review capacity stays flat, and a large share of approved material is never used by field teams. Generative tools make producing more content trivial; they do nothing for the bottleneck, which was always review, not drafting. The uncomfortable implication is that AI can make the MLR problem worse before it makes it better — unless teams use it to send reviewers less and better, rather than simply more.

The reality check on “agents”

Amid the agentic-AI enthusiasm, the sober note worth keeping: by Gartner’s reckoning most vendor “agents” aren’t genuinely agentic, and it expects more than 40% of agentic AI projects to be scrapped by 2027 on cost, unclear value and weak risk controls. Before you buy anything badged “agentic,” the question to ask is plain — what decision does it actually make on its own, and what happens when it makes the wrong one. In a regulated field, an agent that can decide is also an agent that can decide wrongly, and the accountability stays with you.

That’s it for this edition. Back Friday.

— Ned

Follow the conversation: #IrreplaceablesHealth

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

Your review pipeline is drifting with the conversation’s mood

If your team runs promotional copy, clinical summaries or MLR-prep outputs through an LLM in a single conversation thread, the scores are being skewed by whatever came before — negative items more than positive, by a factor of 1.6. The fix is unglamorous and effective: a fresh context for every item. Most off-the-shelf review workflows don’t do this, which means the more you batch, the less you can trust the output.

A third of medical chatbots breach their own safety thresholds

A systematic sweep of 6,233 medical GPTs found that a third or more violate operational safety thresholds, and over half of the action-enabled ones carry no privacy disclosures. These are exactly the numbers regulators and legal teams will reach for. If you are advising a client on deploying an AI-assisted medical information tool, you now have an audit framework — and a set of findings — to point at when the question of due diligence comes up.

The FDA is piloting AI in early-phase trials

The agency’s request for information on AI-enabled optimisation of early-phase trials is worth reading not for what it decides — it decides nothing yet — but for how it frames the questions. This is an early look at how the FDA will think about AI in the trial process, and by extension the communications around it. The thinking that shapes the guidance is happening now, in the open.

Ogilvy Health puts its AI strategy on the record

Ogilvy Health has gone public on its adoption pace, its nano-influencer plans and its own proprietary bot, ‘Ria’. Read it as a competitive benchmark, not a playbook: the tool is agency-owned and won’t transfer to an in-house team. What does transfer is the signal about where a large healthcare agency is willing to say, on the record, that it is placing its bets.

The AI upsell is coming at renewal

If your operation runs on Salesforce or a Veeva-adjacent stack, the AI add-on costs land at your next renewal whether you wanted them or not — and Gartner is already warning that enterprise agreements may not hold. For any communications operation whose budget predictability depends on its CRM, this is a line item to raise before it raises itself.

I’ve been hearing that qualifier a lot. A physician writing about the quiet shame of AI in medical writing put a name to it: people confessing their AI use in whispers, prefacing it with “just a little,” as though the size of the admission changes its nature. It maps almost exactly onto what I see in health communications. The adoption is happening everywhere. The talking about it is happening nowhere.

Which is strange, because the numbers say the opposite. A recent industry overview puts adoption somewhere between 67 and 76 per cent — most organisations either using or testing AI across their workflows. Read that and you’d picture a field that has settled the question and moved on. The survey crossed the threshold. The people did not.

The gap between those two facts is where most of us are actually living. Officially, AI is a productivity story with a tidy percentage attached. Unofficially, it’s a thing you do at your desk and don’t quite mention at the team meeting, because you’re not sure whether disclosing it makes you look efficient or replaceable. The web developers have started saying this out loud — that AI is generating more than half their output while their employers quietly reassess headcount. They’ve decided the silence costs more than the admission. We haven’t, yet.

I don’t think the shame is about the tool. It’s about the unspoken sum the qualifier is trying to manage: if a little AI is fine and a lot is suspect, then admitting how much you actually use is admitting something about how much of the work was ever the hard part. That’s the real conversation, and “just a little” is the sound of us avoiding it.

So here is the question I’d put to the floor. If two-thirds of us are already doing this, who exactly is the whisper for?

— Ned

The MHRA starts drawing the rules for AI that won’t sit still Dame Jennifer Dixon, writing for the MHRA’s strategy series, sets out the genuinely hard part of regulating AI that adapts in the real world rather than staying a fixed “product”: not just whether a model is safe and accurate at launch, but whether it stays so as it drifts, and whether it is usable, acceptable and fair in practice. Her sharper point is one health communications should sit with — NHS local evaluations are often weak, with strong incentives to over-claim results. That temptation will land on anyone writing up an AI tool’s performance, and the credibility bar is about to rise.

Digital twins move from simulating the trial to steering it

In a sponsored interview — read it as a vendor’s pitch — Unlearn describes pushing digital-twin models out of trial simulation and into the whole decision chain: planning, blinded data monitoring, analysis. The substantive signal under the marketing is real, though: its PROCOVA prognostic-covariate-adjustment method now carries EMA qualification and aligns with FDA guidance, and Phase 3 trials using digital-twin analysis are described as imminent. For anyone communicating trial design or evidence, the synthetic-comparator conversation has stopped being hypothetical — and “take the regulators seriously, early and often” is the line worth stealing.

Does AI spell the end for HEOR?

The Healthcare Economist poses the question every value-and-evidence team is quietly asking: if a model can draft the literature review, build the economic model and assemble the dossier, what is left that is irreducibly human? The answer worth holding is that payers buy credibility, not output — and a model fluent enough to generate a cost-effectiveness case is equally fluent at generating a plausible wrong one. The work that survives is the judgement that defends the assumptions, not the keystrokes that produced them.

The token bill comes due

TechCrunch charts the scramble now that the metered cost of running large-language-model workflows at scale has stopped being a rounding error. For agencies and medical teams that have quietly routed literature reviews, drafting and slide production through these tools, the lesson is unglamorous but overdue: budget for usage as you would any other input, because the per-token economics now scale with your volume, not your headcount. The free-tier era of “just try it on everything” is closing.

Reword the question, change the diagnosis

A new study finds clinical LLMs remain acutely sensitive to small prompt changes — the same question, phrased two ways, can yield different answers. For patient-facing or clinical content, that fragility is the whole risk: reliability that depends on exact wording is not reliability at all. It is the standing argument for keeping a human between the model and anything a patient or regulator will read, made once more, with data.

That’s it for this edition. Back Wednesday.

— Ned

Follow the conversation: #IrreplaceablesHealth

A colleague’s IT team “rolled out AI to everyone” last week — meaning the assistant in the browser, the one that summarises your tabs and rewrites your emails. The medical writers were to use it like everyone else. Reasonable, free, already there, and a quiet category error. The general-purpose assistant is built to be passable at everything for everyone; our work is narrow, exact and unforgiving of the near-enough. Harvard Business Review made the wider point this week: healthcare’s specialised dialects sit beyond what a generic model handles dependably. The danger was never that the tool is bad — it’s that it’s good enough to be trusted by people who can’t see where it fails, and in our field the failures stay invisible until someone qualified looks. When the rollout email lands, the useful reply is a question: trained on what, grounded in which sources, checked by whom?

Source link → Harvard Business Review: https://hbr.org/2026/03/healthcare-uses-specialized-language-it-needs-specialized-ai-too

The unfashionable question, asked out loud

For two years the conference AI sessions ran to a script: a demo, some adoption statistics, a closing slide about the future arriving whether you like it or not. That’s changing. STAT reported on panels where scientists are finally asking the unfashionable question aloud — not “is it impressive?” but “does it actually work?” We sit downstream of these tools; when a client’s research team runs a literature scan through an AI platform, the output flows into our copy and our name goes on the claim at the end. So practitioner scepticism isn’t something to borrow sheepishly. It’s cover. Make the question respectable in your own rooms: validated against what, on whose data, and who notices when it’s wrong?

Source link → STAT: https://www.statnews.com/2026/05/21/are-ai-scientist-tools-actually-useful-ai-prognosis/

When the builder is more cautious than the brief

The brief wanted copy on how the company “used AI to design” its lead molecule — the implication, unmistakably, speed. The trouble is that the person best placed to know was more careful than the marketing. In an interview with STAT, the chief executive of an AI-driven biotech drew the line: AI accelerated the early design — the candidates worth making — while everything after, the assays, safety, trials and regulators that decide whether a drug exists, runs at the old speed. The clever bit is narrow; the slow bit is most of it. When the person who built the model won’t overclaim, the writer downstream certainly shouldn’t. My test now: would the scientist who did the work nod, or wince? Write for the nod.

Source link → STAT: https://www.statnews.com/2026/05/26/ai-biotech-bighat-biosciences-ceo-on-ai-drug-development-hype/

— Ned

Follow the conversation: #IrreplaceablesHealth

We started with a single sentence — when this happens, do that — and climbed all the way to agents that set their own steps. This last instalment is the most important, because it’s the one that decides whether all the rest helps you or quietly hollows out your work. The question isn’t what can Zapier do? By now the answer is “a startling amount.” The question is what should you let it?

Two piles

Sort the work into two piles. The first is dull, rules-based, repeated and reversible: gathering, filing, formatting, copying, reminding, first-pass sorting. This is the pile to automate without guilt. Every hour you reclaim here is an hour returned to the work that actually carries your name. The Signal pipeline lives entirely in this pile, and it should.

The second pile is judgement, taste, context and accountability: deciding what a finding really means, whether a claim is fair, how to phrase something a worried patient will read, what to leave out, when the data doesn’t support the headline everyone wants. This is the pile you do not hand over — not because the tools can’t produce something plausible, but because plausible is precisely the danger. A confident, well-formatted, subtly wrong paragraph is worse than a blank page, because it invites you to stop looking.

The test

When you’re unsure which pile a task belongs in, ask three questions. Is it reversible — if the machine gets it wrong, can you catch and undo it cheaply? Is it low-stakes — does a mistake cost minutes, or credibility? And is there a human at the gate before anything becomes fact, becomes public, or reaches a regulator? Three yeses and you can automate it. A single no and you keep your hands on it.

Why this is the whole point

This series has been practical — feeds, filters, AI steps, agents — but the thread underneath it is the reason this newsletter exists. The value of a health communications professional was never the typing, the filing or the formatting. It was the judgement: knowing what’s true, what’s allowed, what’s clear, and what’s safe to put in front of someone who will act on it. Automation is very good news for that person, because it strips away the part of the job that was never really the job.

The people who struggle in the next few years won’t be the ones who refused to automate. They’ll be the ones who automated the wrong pile — who handed over the judgement and kept the typing. Do it the other way round. Let the machine carry the queue. Keep the part only you can do.

That’s where we’ll leave it. Build the boring half. Guard the irreplaceable one.

Zapier for health communications is a practical series. This is the final part.

— Ned

We’ve climbed a ladder. Plumbing that moves things. Logic that branches. AI steps that interpret. Copilot that builds from a description. The top rung is the one everyone’s talking about and few are using well: agents. This is where automation stops following a fixed recipe and starts deciding its own steps — and where, for anyone in a regulated field, the questions get serious.

What an agent actually is

A normal Zap is a fixed track: trigger, then these steps, in this order, every time. A Zapier Agent is different. You give it a goal in plain English — “triage each new Signal item: judge relevance, summarise it, score it, and flag the strong ones for me” — and the agent works out how to get there. It reasons, picks which of its available tools to use, takes several actions, and can loop back if the first attempt falls short. The same agent can read a feed, write to a table, post to Slack and pause for a human, all from one instruction, across Zapier’s 9,000-plus connected apps.

The shift is from automation you choreograph to a teammate you delegate to. That’s genuinely powerful. It’s also exactly why you need to slow down before handing it anything that matters.

The capabilities worth knowing

Three features make agents more than a gimmick. Memory lets an agent carry context across runs, so it isn’t starting cold every time. Bring Your Own Model lets you choose which underlying AI powers it. And it can keep a human in the loop by design — pausing to wait for a named person’s approval before it does anything consequential. That last one isn’t a nicety in our world. It’s the whole basis on which an agent can be trusted at all.

The guardrails — and why they matter here

To Zapier’s credit, the platform has built a safety layer, and it maps almost exactly onto health communications anxieties. The guardrails scan for personally identifiable information across more than thirty types, detect prompt-injection attempts, and flag toxic or harmful language and negative sentiment. In plain terms: it tries to stop an agent leaking patient or personal data, being hijacked by malicious text hidden in its inputs, or producing something it shouldn’t.

Take the prompt-injection point seriously. An agent that reads external content — emails, web pages, documents — can be fed instructions buried in that content, designed to make it act against you. In a field where the inputs are clinical documents, KOL correspondence and regulated copy, that is not a theoretical risk. Guardrails reduce it; they do not abolish it.

Where the line sits

So here’s the honest position. An agent triaging your Signal queue, drafting first-pass summaries, flagging what looks strong for your review — that’s a fine use. The work is reversible, low-stakes, and a human sees everything before it goes anywhere. Let the agent move work to the gate.

An agent that drafts a claim and sends it, adjusts approved copy without review, or touches anything destined for a regulator without a person signing it off — that’s the wrong side of the line, guardrails or not. The technology is ready to propose. It is not ready, and may never be ready in our field, to decide unattended. Keep a named human accountable at every consequential step, by design and not as a courtesy.

Next: 16. Zapier: what you don’t hand to the robot — the closer, on where automation stops and your judgement begins.

Zapier for health communications is a practical series. New post every week.

— Ned

Until now we’ve built everything by hand — picking triggers, choosing actions, wiring steps together. That skill still matters, but the way you start a build has changed. Zapier now lets you describe what you want in plain English and have it assemble the thing for you. It’s the difference between knowing how to lay bricks and being handed the walls. You still need to know what a good wall looks like.

Copilot: say it, and it builds it

Zapier Copilot takes a plain-English description — “when a new item lands in my Inoreader folder, summarise it and add it to my Signal table” — and builds the Zap, choosing the trigger, the steps and the connections. Zapier reports success rates around 92% for simple Zaps, which is to say: for the everyday stuff, it mostly just works.

What this changes is the starting line. Instead of staring at an empty canvas wondering which trigger you need, you describe the outcome and let Copilot produce a first draft of the automation. You then check it, correct it, and switch it on. It also builds the other pieces — a Table, an Interface, an agent — from the same kind of instruction.

The catch is the obvious one: Copilot builds what you said, not what you meant. It will happily wire up something subtly wrong with total confidence. So the old understanding isn’t obsolete — it’s what lets you look at the draft and spot that the filter is in the wrong place or the date format will break the next step. Describing is faster. Reviewing is still on you.

Canvas: drawing the system before you build it

Canvas is the companion piece, and it works the other way round. It’s a visual space for mapping a workflow or a whole system — boxes and arrows for how information should move — which you can then turn into actual Zaps. Sketch the flow first, agree it’s right, then generate the build from the diagram.

For anything with more than a couple of branches, this is genuinely useful. It forces you to think the process through before committing to it, and it gives you a picture you can show a colleague or a client without making them read a list of steps. Map, agree, build — in that order — and you make your mistakes on a diagram instead of in a live automation.

What this shift really means

Step back and the pattern is clear: the centre of gravity is moving from operating the tool to specifying the outcome. The valuable skill is no longer remembering where a setting lives; it’s being able to describe precisely what you want, and to recognise when what you got back isn’t it.

That’s a comfortable shift for people who work in health communications, because precise specification is already the job. A good brief, a clear set of requirements, an exact description of the output you need — that is the same muscle. The people who get the most from Copilot and Canvas won’t be the most technical. They’ll be the clearest.

Next: 15. Zapier: when the automation makes decisions — agents that set their own steps, and the guardrails that decide whether you can trust them.

Zapier for health communications is a practical series. New post every week.

— Ned

Everything so far has been plumbing — reliable, literal, and a bit dim. It moves things and checks conditions, but it doesn’t understand anything. This week the automation gets a brain. Zapier now lets you drop AI into the middle of a workflow, keep smarter data, and put a chatbot over your own material. Used well, it closes the gap between moving information and making sense of it. Used carelessly, it’s how a confident mistake ends up in your queue.

AI steps: judgement, inserted mid-Zap

The headline change is the AI step. Between a trigger and an action, you can now ask a model to do something with the data passing through: summarise this article in one line, classify whether it’s relevant to my themes, pull out the key claim, draft a neutral one-sentence gist, tag it by topic.

Go back to the Signal filter from a fortnight ago. That filter matches keywords — it sees words, not meaning. An AI step can read the item the way a person skim-reads it and judge relevance on substance, not vocabulary. The blunt instrument becomes a sharp one. The same trick writes you a tidy summary column, or sorts items into themes, before anything reaches your eyes.

Tables: a database that thinks

Zapier Tables is a database built for automation, living inside Zapier itself rather than bolted on via Sheets. For a Signal-style queue it’s a natural home: structured columns, and crucially an “AI Enrich” feature that auto-fills a field by running a prompt against the rest of the row. A “one-line summary” column or a “relevance score” column that populates itself as each item lands — no extra Zap required.

If you’ve outgrown a Google Sheet that’s creaking under formulas and helper Zaps, this is the upgrade.

Chatbots: answers from your own material

The third piece is chatbots you can point at your own content — a set of files, webpages, or a table. Feed it your style guide, your past issues, your house rules, and you have something you (or a colleague) can ask in plain English: “have we covered this topic before?”, “what’s our line on AI-written patient materials?” It’s a way to make institutional knowledge queryable instead of buried.

The caveat that never goes away

Here is the part that matters more in our field than almost any other. The moment you let a model interpret, you inherit its failure mode: it will be confidently wrong some of the time. It will summarise a study and quietly overstate the finding. It will tag something safe that isn’t. It will, asked to “extract” a statistic, invent a plausible one.

So the rule holds, and hardens: an AI step is a drafting aid, never an authority. Use it to triage, summarise and sort — work where a wrong call is cheap and caught at the next human glance. Keep it well away from anything that ships as fact without a person checking it against the source. In regulated health communications, the verification isn’t overhead. It’s the job.

Next: 14. Zapier: just describe it — using Copilot and Canvas to build automations by describing what you want, instead of wiring every step by hand.

Zapier for health communications is a practical series. New post every week.

— Ned

The Signal pipeline from last week was deliberately simple: watch, filter, tidy, file. Real work is rarely that linear. Sooner or later you want a Zap that makes choices — do this for press releases, that for preprints — and copes with the mess that real data arrives in. That is what this instalment is about: the four building blocks that turn a toy automation into something you’d actually trust.

Filters: the bouncer on the door

A Filter does one job: it lets a Zap continue only if conditions are met, and quietly stops it otherwise. “Only continue if the source is on my approved list.” “Only continue if the headline contains one of these terms.” It is the single most useful step you will add, because most automation problems are really too much happening problems. A good filter is the difference between a queue you read and a queue you avoid.

The craft is in being neither too tight nor too loose. Too tight and you silently drop things that mattered; too loose and you’re back to noise. When in doubt for editorial work, lean loose — a few extra items to skim beats a missed story you never knew about.

Paths: when the answer is “it depends”

Filters are pass/fail. Paths let a single Zap branch: if the item is a regulatory announcement, do one thing; if it’s a journal article, do another; otherwise, do a third. Each path has its own conditions and its own actions, all inside one Zap you can see end to end.

This is where automation starts to mirror how you actually think. You don’t treat an FDA warning letter the same as a conference abstract, and now your Zap doesn’t have to either. Resist the urge to build a maze, though — two or three clear branches you understand beat ten you don’t.

Formatter: making messy data behave

Data arrives ugly. Dates in five formats, names in inconsistent case, links wrapped in tracking junk, text with line breaks where you don’t want them. Formatter is Zapier’s quiet workhorse for cleaning all of that: reformat a date, trim whitespace, change case, extract a number or an email from a block of text, split a field, find-and-replace.

It is unglamorous and it is essential. Most automations that “don’t work” aren’t broken — they’re choking on data that arrived in a shape the next step didn’t expect. A Formatter step in the right place fixes more problems than any amount of clever logic.

Schedules, delays and lookups

Three more tools worth knowing, briefly. Schedule lets a Zap run on a clock — every weekday at 7am, say — rather than only when something happens, which is how you’d build a daily digest. Delay holds an action for a set time or until a specified moment, useful when you want a pause before a follow-up. Lookup checks whether a record already exists before creating it, which is how you stop a Zap filing the same item twice.

The principle underneath

Every one of these is a small, legible piece of logic. The skill is not memorising features; it’s decomposing your own process into steps clear enough to hand over. If you can write down, plainly, “when X arrives, check Y, and if so do Z — otherwise do W,” you can build it. The tools are easy. The thinking is the work, and the thinking is yours.

Next: 13. Zapier: giving your Zaps a brain — adding AI steps, Tables and chatbots, so your automation can start to interpret, not just move.

Zapier for health communications is a practical series. New post every week.

— Ned

Last time I made the case that Zapier is just “when this happens, do that.” This week, the proof: the actual pipeline that feeds Signal, the news section of this newsletter. It is not clever. That is the point. It runs every day, it never forgets, and it turns a chaotic firehose of industry news into a single, reviewable list I can sit down with.

Here is the whole thing, step by step.

The problem it solves

Keeping across AI in health communications means watching dozens of sources — trade press, company announcements, regulators, preprint servers, a few sharp individuals. Done by hand, that is an hour of tab-juggling every morning, and the day you skip it is the day the story breaks. I wanted the watching to happen on its own, so my time goes on the judgement: what matters, and why.

The pipeline

Trigger — a new item appears in my reading feeds. I keep my sources organised in an RSS reader (Inoreader), grouped into a single folder of things worth monitoring. Zapier watches that folder. The moment a new article lands, the Zap fires. No app to open, no button to press.

Filter — is this actually relevant? Most of what comes through is noise. A Filter step checks each item against a set of conditions — does the headline or summary mention the themes I track? — and quietly stops anything that doesn’t qualify. Nothing downstream happens unless the item clears the bar. This one step is the difference between a useful queue and an unreadable one.

Format — make it consistent. A Formatter step tidies the raw feed data into the shape I want: a clean title, the source, the link, the date in a sensible format. Feeds are messy; this makes every entry land the same way.

Action — add it to the queue. Finally, Zapier writes a new row into a Google Sheet — the Signal Queue. Title, summary, source, URL, date, each in its column, ready for scoring. That sheet is the thing I actually open.

That’s four steps: watch, filter, tidy, file. One trigger, three actions, running unattended.

What I’m left with

By the time I sit down, the morning’s reading has already sorted itself into a single list of candidates — filtered, formatted, and waiting. My job starts where the automation stops: reading each one properly, deciding what earns a place, and writing the line that makes it worth your time. The machine does the gathering. I do the choosing.

That division is deliberate, and it is the whole philosophy of this series. Automation is brilliant at collecting and hopeless at judging. Let it collect.

The honest limitations

The filter is blunt. It works on keywords, not understanding, so it lets through things that match the words but miss the point, and it occasionally drops something it shouldn’t. I’d rather it err towards letting too much through — a few extra items to skim beats a missed story. Making that filter smarter, so it judges relevance more like I would, is exactly where the AI steps come in — and that is the next stage of this series.

For now, the lesson is the one that matters most: you do not need anything sophisticated to claw back real time. A trigger, a filter, a tidy-up and a destination will do it.

Next: 12. Zapier: when one step isn’t enough — filters, paths and tidy data, and how to build a Zap that handles the messy real world.

Zapier for health communications is a practical series. New post every week.

— Ned

If you work in health communications, you already run on small, repetitive tasks that don’t need your brain but still eat your day. Saving an email attachment to the right folder. Copying a line from a form into a tracker. Pasting a link into a spreadsheet so it doesn’t get lost. Pinging a colleague when something lands. None of it is hard. All of it is friction.

Zapier is the tool that does that moving-around for you, automatically, while you get on with the work that actually needs judgement.

That’s the whole pitch. Everything else is detail.

The one idea

Strip away the marketing and Zapier rests on a single sentence: when this happens, do that.

When a new email arrives, save the attachment. When a form is submitted, add a row to a sheet. When an item appears in a feed, drop it into a tracker. The first half is the trigger — the thing that kicks it off. The second half is the action — what you want done in response. String the two together and you have a “Zap”: a small, automatic rule that runs without you.

You are not programming. You are describing a piece of your own routine to a tool that can then repeat it forever, in seconds, without getting bored, distracted, or forgetting step three on a Friday.

The four words worth knowing

You can ignore almost all of Zapier’s vocabulary, but four words make everything else make sense:

• App — any service Zapier can talk to: Gmail, Google Sheets, Outlook, Slack, Dropbox, your RSS reader, thousands more.

• Trigger — the event that starts a Zap (”new email in this label”, “new row in this sheet”).

• Action — what Zapier does in response (”create a document”, “send a message”, “add a row”).

• Zap — one trigger plus one or more actions, switched on and running.

That’s the entire mental model. If you understand “when this app does X, make that app do Y,” you understand Zapier.

Why this matters for a small operation

Big teams buy automation to save headcount. For a one-person publication or a lean health communications team, the value is different and, frankly, better: it gives you your attention back. The hour you’d otherwise spend filing, copying and chasing is the hour you don’t have for the thing only you can do — the editorial call, the nuance, the sentence that has to be exactly right because a regulator might read it.

Automation is at its best on work that is dull, rules-based and repeated. It is at its worst — and we will come back to this hard, later in the series — on work that needs taste, context or accountability. Knowing which is which is the whole skill.

What Zapier is not (yet)

A point worth planting early, because it gets muddled: in its simplest form, Zapier is not making decisions. It is not reading your content and judging it. It is plumbing — reliable, literal, slightly dim plumbing that does exactly what you told it and nothing more. That is a feature, not a flaw: predictable is precisely what you want from something running unattended.

The clever, decision-making, “agentic” end of Zapier is real, and it is where this series is heading. But you earn the right to it by understanding the plumbing first.

Build one in five minutes

The fastest way to get it is to make one trivial Zap and watch it work. Try this:

• Sign up for the free plan and click Create.

• For the trigger, choose an app you already use — say Gmail — and the event “New Attachment”.

• Connect your account and pick the label or search term to watch.

• For the action, choose Google Drive and “Upload File”, pointing it at a folder.

• Run the test. Zapier grabs a recent attachment and files it for you.

• Switch it on.

That’s it. A small, dull job now does itself. Nothing here will change your life — but it is the same shape as everything that follows, including the workflow I actually rely on every week.

Next: 11. Zapier: how I actually build Signal — the real pipeline that fills the queue behind this newsletter, start to finish.

Zapier for health communications is a practical series. New post every week.

— Ned

🥜🥜🥜🥜🥜 CRACKING

What it is: A copy-paste prompt template, posted on X by @MedCommsAI.

Why it’s worth your time: It forces the output into three sections — Clinical Context, Decision Support, Educational Points — which is how clinicians actually consult reference material, so you get something scannable and decision-oriented rather than an essay. Tested on turning a dense treatment guideline into a one-page HCP reference, it produced a usable first draft with the right architecture and tone in seconds. The structure is therapeutic-area-agnostic, so the same prompt works for any condition.

How to use it:

• Paste the template and swap [THERAPEUTIC AREA/CONDITION] for your topic.

• Keep the three section headers exactly — Clinical Context, Decision Support, Educational Points. They are what make the output usable.

• Close with the format instruction: “Concise, scannable layout. Professional health communications tone.”

Watch out for: It generates structure, not evidence — every clinical claim it produces needs a writer and a source check before it goes anywhere near a physician or MLR.

Workflow:

TRIGGER → New treatment guideline or label update lands from medical affairs

STEP 1 — Source gather (medical writer): pull the guideline, label and pivotal data into one place

STEP 2 — Draft scaffold (featured prompt): generate the three-section reference structure in Claude or ChatGPT

STEP 3 — Evidence fill (medical writer): replace each placeholder with sourced, referenced claims

STEP 4 — MLR review (human / Veeva Vault): claims, fair balance and references checked and signed off

OUTPUT → Approved one-page physician reference tool in the house template

An MLR pre-check that lives inside Veeva Vault

🥜🥜🥜🥜 GOOD NUT

What it is: A tool — two AI agents, Quick Check and Content, now built into Veeva PromoMats (announced December 2025).

Why it’s worth your time: The Quick Check Agent screens content against editorial, brand, market, channel and compliance rules before it reaches MLR, so reviewers stop burning hours on the errors a checklist could have caught; the Content Agent summarises documents and reads the visuals to orient reviewers fast. The part that matters in a regulated environment is that it runs inside Vault — your promotional copy never leaves the system, which is the usual reason teams won’t put it near a chatbot. Moderna is among the first users and is openly aiming at “nearly touch-free” routine review.

How to use it:

• If you’re already on Veeva PromoMats, ask your admin about early access to Veeva AI.

• Pilot it on one high-volume, low-risk asset type first — HCP emails or banners — before trusting it anywhere near a launch.

Watch out for: It’s Veeva-Vault-only with no free tier, still early access (single-digit customers), and the real test is whether its flags are consistent and auditable enough for a regulator — treat it as a first pass, never the reviewer of record.

A free library of 90 medical-writing prompts worth raiding

🥜🥜🥜🥜 GOOD NUT

What it is: A resource — an ungated library of 90 copy-paste prompt templates from aingens, organised by research and writing task.

Why it’s worth your time: Unlike most “prompt pack” lead magnets, these are free, fully visible and well-built, with placeholders for population, condition, endpoint, audience and reading level, and sensible defaults baked in (PICO, GRADE, effect sizes, safety language). The abstract, medical affairs brief, patient handout and FAQ templates are strong scaffolds you can drop straight into any model. They earn their keep as structure even if you never touch the vendor’s own tool.

How to use it:

• Grab the template that matches your task — abstract, brief, patient handout, FAQ.

• Fill every bracketed placeholder before you run it; don’t leave a single one in.

• Add your own constraints on top: word count, region, reading level, output format.

Watch out for: Several prompts ask the model to “extract” study statistics and citations, which a generic model will cheerfully fabricate — use them for structure, and verify every number and reference before anything goes near MLR.

Also worth a look

Two more from the queue — worth a glance, but they didn’t clear the bar for a full slot.

Indegene’s NEXT MLR Review Automation

🥜🥜🥜 DECENT

Another agentic pre-MLR platform, this one promising claims and label alignment, local ABPI and PAAB checks, and integration with Veeva and Workfront. Squarely aimed at regulated content, but it’s a demo-gated enterprise product with no free tier and nothing you can test from the blog. File it under “proof the category is maturing,” not “try this tonight.”

Copilot Researcher Agent for Medical Affairs

🥜🥜🥜 DECENT

A set of sensible workflows — KOL prep, coaching reports, SWOT drafts — for the Researcher agent most teams already have in Microsoft Copilot and never use. The catch: the actual prompts are gated behind an email signup, and this is more Medical Affairs operations than content production — useful adjacent reading rather than a core comms tool.

That clears the strong end of the queue — back next week once the latest crop has been put through its paces.

Follow the conversation: #IrreplaceablesHealth

Below are six patterns for wiring several agents together, each with the shape it makes on the page. None of them is new — they are borrowed from distributed computing and have been quietly running your search engine and your bank for years. What has changed is that you can now assemble them yourself, in an afternoon, without writing the plumbing underneath. Here is what each one does, and where it earns its keep in health communications.

1. Classify-and-act. A router. One agent reads the incoming task and sends it down the right path — triage an inbound medical-information query to the correct specialist, or sort a piece of content by whether it needs MLR review at all. Cheap, and it stops you running an expensive workflow on something that needed a one-line answer.

2. Fan-out-and-synthesise. Several agents work the same problem in parallel, and a final agent merges their output. Useful when a task has genuinely separate facets — summarise a data readout for efficacy, safety and access at the same time, then stitch the threads into one briefing.

3. Adversarial verification. A worker produces; independent verifiers check. This is the pattern that matters most in regulated work, and the one teams skip first. A drafting agent writes, a separate agent checks every claim against source, and a third checks against the MLR rulebook. Never let the model that wrote the copy be the one that signs it off.

4. Generate-and-filter. Many cheap generators throw out ideas; a filter scores them against a rubric and removes the duplicates; the best survive. Headline options, discussion questions, plain-language summaries — generate forty, keep four.

5. Tournament. When “best” is a matter of judgement rather than a checklist, judge the candidates pairwise and let a winner emerge. Slower than a filter, better when quality is subjective — choosing the sharpest framing of an executive summary.

6. Loop until done. An agent works, asks “is there anything new here?”, and spawns another pass if the answer is yes. A literature scan that keeps searching until it stops finding anything, then stops. The discipline is entirely in the exit condition; without a good one, it runs forever.

Where Cowork comes in

A year ago, wiring these together meant code — queues, retries, state you had to hold somewhere. Cowork collapses most of that. You describe the flow in plain language and it does the assembly: spawns the subagents, fans the work out, runs the verifier, loops until the exit condition trips, and reaches into your actual tools — the document store, the spreadsheet, the publishing platform — through connectors. The single advisory board agent from the last piece stops being the whole system and becomes one node in a larger graph. You compose the graph by describing it, not by building it.

That is the real shift, and it is easy to undersell. The patterns were always available to anyone with an engineering team. What is new is that the person who actually understands the regulatory and editorial constraints — you — can now assemble them directly, without translating every requirement through a developer first.

A worked example: the advisory board pack

Stitch four of the six together and you have a workflow that does real work.

Classify the incoming materials by therapy area and route them (1). Fan out across the pre-reads, advisor profiles and prior transcripts, and synthesise a board briefing (2). Generate candidate discussion questions and filter them against a rubric — non-leading, on-objective, compliant — discarding the rest (4). Then, and this part is non-negotiable, run the whole pack past a verification agent that checks every factual claim and flags anything MLR would catch (3). After the meeting, a loop-until-done pass mines the transcript for themes, spawning follow-ups until no new insight appears (6).

What used to be a fortnight of associate time becomes a reviewed first draft by the morning after the meeting. Not a finished pack — a first draft, with the checking already done and the gaps already marked.

The patterns are old. The ability to assemble them without an engineering team is the actual change. And, as ever, adversarial verification is the step you cannot afford to leave out — in this field, the workflow that generates is only as trustworthy as the one that checks it.

— Ned

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.

Alnylam bets $2bn on biology-native AI to redesign RNAi discovery

Alnylam has signed a $2bn partnership with Inceptive — the startup founded by one of the architects of the transformer — to build foundation models aimed squarely at drug design rather than back-office efficiency. For health communications, the signal isn’t the chemistry; it’s that the upstream science your clients talk about is increasingly machine-shaped, and the story of how a molecule was discovered is about to get harder to tell both accurately and simply.

Agentic AI is stuck in pilot purgatory, Forrester finds

Three-quarters of enterprises say they’ve adopted agentic AI; Forrester’s data says almost none have it in production. That gap is the one most health communications directors already feel in their bones — the vendor demo that evaporates on contact with governance, permissions and real workflows. Useful cover the next time a client asks why the agent that dazzled in the pitch still isn’t drafting their submissions.

LLMs invent confident pharmacology for drugs that don’t exist

In a clinician-relevant study, models generated plausible clinical content for entirely fictitious drugs, prompted by nothing more than familiar prefixes and suffixes. For anyone producing AI-assisted medical or patient-facing copy, this is the failure mode worth losing sleep over: the model isn’t reasoning about the molecule, it’s pattern-matching the name — and it does so with complete confidence. Verification against source isn’t a nicety here; it’s the only thing between you and a fabricated mechanism of action.

Multi-agent AI “consensus” buries the facts it should surface

The fashionable fix for one unreliable model — put several in a room and let them agree — turns out to lose critical information in the deliberation rather than catch it. If your evidence synthesis or advisory-board workflows are drifting toward multi-agent review on the assumption that agreement equals accuracy, here is the counter-evidence: a confident shared answer can be the most dangerous output of all. Build the disagreement in deliberately, or you automate away the one signal worth having.

UK publishers can now opt out of Google’s AI summaries

A CMA ruling lets publishers block Google from ingesting their content into AI Overviews, and forces Google to cite sources with links. For medical publishers, journal owners and anyone whose clinical content has quietly been feeding AI answers, this is the first real lever to control how that material is used and attributed. Worth raising with any client who owns content and hasn’t yet thought about where it ends up.

That’s it for this edition. Back Wednesday.

— Ned

Follow the conversation: #IrreplaceablesHealth