A Field Guide to Medical Communications, 2026–2030
Updated 2026-07-13 · markdown version
This is a map of a war that has already started, drawn by a participant. Treat it accordingly: it is Mantir's forecast — scenario, not reportage — anchored to real events where they exist and honest speculation where they don't. The war is over a single question: when AI enters medical communications, does it replace the writer or arm the writer? Every faction below is an answer to that question with a budget attached.
The factions
The Sponsors (pharma commercial & procurement). Command the money and feel the squeeze — pipelines demand more materials, agencies bill by the deliverable, and every earnings call rewards an "AI efficiency" story. Their move: buy one-button authoring and bet the agency line item shrinks. Real signal: strategic pharma money is already inside the tooling — Eli Lilly participated in Revisto's November 2024 seed round. Their weakness: they buy at portfolio distance from the work, so they systematically underprice the substantiation gap.
The Button-Makers (sponsor-side AI vendors). Well-funded, fluent, demo-beautiful. Their pitch is speed: "90% faster" (vendor claims all, unaudited). Their products genuinely compress the typing. Their existential problem: everything they generate still has to cross the MLR wall, and a wall doesn't care how fast you approached it.
The Guilds (medcomms agencies). Scientific, editorial, and client-services pillars; decades of accumulated judgment about what survives review and what a client meant by that email. Under visible agency compression — sponsors asking why the deck costs what it costs. Their fork in the road: adopt writer-side AI and sell judgment-at-machine-speed, or race the buttons on price and lose.
The Free Companies (freelance medical writers). The most exposed and the most agile — no procurement department, no validated stack, free to adopt any tool tomorrow. (Median rate per Upwork's own data: $36/hour; specialists far above it.) In every prior automation wave, independents who armed themselves early captured outsized share as buyers learned the difference between generated and good.
The Wardens (MLR committees, final medical signatories, OPDP/PMCPA). The terrain-owners. They don't compete in the market; they are the ground it's fought on. Their incentives are asymmetric — a reviewer gets no credit for approving fast and full blame for approving wrong — which makes them structurally immune to fluency and obsessed with provenance. Nothing about generative AI changes their incentives; it only changes the volume of unverified content arriving at their gate.
The Armorers (writer-side toolmakers — where Mantir stands). Build for the writer instead of over them: evidence-linked authoring, live substantiation, pre-review agents. The bet: in a regulated industry, the tool that owns verification outlasts the tool that owns generation — the verification asymmetry guarantees the bottleneck lands there.
The terrain
Three features of the battlefield decide more than any tool's model quality:
- The MLR wall. Every deliverable crosses it; nothing ships around it. Content that arrives substantiated crosses in one round; content that arrives fluent-but-unverified queues, bounces, and re-queues. The wall converts "generated fast" into "parked longer."
- The provenance ledger. Claims, evidence, versions, annotations. Whoever maintains it — spreadsheet, vault, or writer-side library — holds the industry's actual source of truth. Buttons don't want to keep ledgers; ledgers are where their outputs get falsified.
- The trust gradient. Reviewers extend goodwill to writers whose packs verify cleanly, and goodwill is measured in rounds. It accrues to people, not tools — which is why tools that make their humans look immaculate compound, and tools that make humans into rubber stamps (human-as-the-loop) burn the gradient down.
How it plausibly unfolds
2026 — The demo years (now). Buttons proliferate; pilots everywhere; slideware victories. Early writer-side tools reach working writers. Review queues quietly lengthen where generated volume rises. You are here.
2027 — The wall bill arrives. The first honest post-mortems: generated deliverables didn't reduce total cycle time, they moved the cost from writing to review. Findings-per-submission becomes a tracked metric. Sponsors start asking vendors the awkward question: show me rounds, not drafts.
2028 — The great re-sort. Procurement splits the category it once lumped: generation tools (cheap, commoditizing) versus verification tools (sticky, workflow-owning). Agencies that armed their writers publish cycle-time numbers agencies that didn't can't match. Freelancers with substantiation tooling win work that used to require an agency's back office.
2029 — Judgment premium. The market reprices what it's actually short of: accountable judgment. Signatory and reviewer capacity — never automatable, structurally scarce — sets the industry's clock speed. Tools are evaluated by one number: verified-claims-per-reviewer-hour.
2030 — The settlement. Nobody's flag in the mud. Buttons survive as drafting utilities inside co-creation workflows — demoted from "replaces the writer" to "types the boring parts." The writer survives promoted: fewer, better-armed, owning the ledger and the judgment. The agencies that survive look less like deliverable factories and more like judgment houses with very good armories.
What would falsify this map
Forecasts you can't lose are propaganda, so: this map is wrong if regulators materially relax substantiation requirements (the wall shortens); if models learn to verify against sources as reliably as they generate (the asymmetry closes from the other side); or if sponsors prove willing to absorb regulatory risk at scale for speed (the trust gradient stops mattering). We watch all three. So should you.