Reading the 2027 Job Listings in Advance ── Verifiers, context suppliers, and go-betweens — the shape of new work in pharma review, medical affairs, and safety

We have all heard enough about AI erasing jobs. This article looks at the other side. It makes concrete predictions about the new jobs born at the boundary where AI and people divide their work — jobs that verify output and accept responsibility for it, jobs that supply context, and jobs that act as go-betweens among AIs and between AI and people — and brings them down to the working reality of the pharmaceutical industry.

01The Jobs That Multiplied After Coachmen Disappeared — How to Predict Occupations Honestly

Around 1900, the streets of New York and London were filled with horse-drawn carriages. As automobiles spread, coachmen (= the people who drove those carriages for a living) vanished within twenty to thirty years. But the total number of occupations did not shrink. Drivers, auto mechanics, gas station attendants, traffic controllers, car insurance adjusters, driving test examiners. Around the jobs that disappeared — at the "boundary" between the new technology and human society — jobs that had never existed before piled up. One American study offers this estimate: more than half of the jobs held in 2018 were in occupations that did not exist in 1940 (per published reporting and academic work).

There is an asymmetry here that is easy to miss. The process by which jobs disappear is conspicuous. Factory closures and layoffs make the news. The process by which new jobs are born is quiet, and for a while they do not even have names. The term "prompt engineer" (= someone who designs the written instructions given to an AI) appeared in job listings soon after generative AI became a topic, but the substance of that work had already been done, namelessly, by somebody before that. The speed at which technology erases jobs is easy to observe; the speed at which jobs appear at the boundary is hard to observe. Discuss AI and employment without correcting for this asymmetry, and you will drift toward pessimism.

That said, if all we can say is "new jobs have always appeared in history, so this time will be fine too," that is not a prediction — it is a prayer. So this article declares its method up front. Predictions of new occupations for 2027–2029 will be sorted honestly into three tiers of confidence.

One more promise. This article makes no numerical predictions about salaries or headcounts. Those depend on institutions, the economy, and geography, and cannot be derived from the slope of the models. What can be written is "which jobs, for what reasons, in what order" — and this article stays within that limit.

02Checking the Slope — Agentic AI Has Already Created "Commissioners, Verifiers, and Responsibility-Holders"

Let us confirm the "slope" the predictions rest on. The model lineage of Anthropic (= the company that develops the AI model "Claude") has run from Opus 4.7 to Opus 4.8, then to Fable 5, which put weight on narrative and long autonomous work, and Sonnet 5, which balanced response speed with practical performance. Through this whole line, what has grown consistently is not one-shot cleverness but "agentic capability" (= the ability to plan and keep executing multi-step work on its own, without a human giving detailed instructions at each step). Work continues for tens of minutes to hours while the human is away from the desk. This one point is the biggest change of the past two years.

When AI tilts from waiting for instructions to executing autonomously, the human side's center of gravity moves too. The work of the person who "makes" things by hand shrinks, and three roles thicken in its place. The person who commissions work from AI (defining what to do, to what quality standard, and how much to delegate). The person who verifies the deliverable (checking whether the AI's output is correct and whether anything is missing). And the person who bears responsibility for the result (signing off on the decision to release the AI's product into the world). Commission, verify, take responsibility. This triad is not, in fact, a new invention. It is exactly the structure a company uses when it sends work to an outside contractor (= a vendor). The only differences are that this contractor never tires, answers instantly, and keeps getting rapidly cheaper.

The sprouts are already visible. In software development, a working style is spreading in which AI writes most of the code and humans spend their time on code review (= reading a program someone else wrote and pointing out errors and design problems). Anthropic itself has stated publicly that a substantial share of its own code is written by Claude (per official statements and public reporting). The substance of a developer's job is quietly being rewritten from "write" to "commission, read, and decide whether to pass or stop." The job title on the listing is still "software engineer," but the content changes first and the name catches up later. That is the typical order in which occupations are born.

Will this structure spread beyond software? I believe it will, conditionally. There are three conditions.

• The deliverable takes the form of documents or data. Code, review records, reports, and contracts are easy for AI to generate and verify directly. Physical work is still outside the boundary.
• Errors are costly, and verification requires specialized knowledge. Where errors are cheap, no one hires a verifier. The more expensive the errors, the more valuable "the person who takes responsibility for the AI's output" becomes.
• Regulation or commercial custom demands that someone be answerable. In areas like pharmaceutical promotional-material review or drug safety information, where the existence of a signatory and a responsible person is fixed by the system, the seat of responsibility itself does not disappear no matter how clever AI becomes.

One of the industries that satisfies these three conditions most fully is the one most readers work in: pharma. The cost of error connects directly to patient health, the documents are vast, and laws and industry codes specify where responsibility sits. Before pharma is "an industry whose jobs AI erases," is it not a candidate for "the industry where the new occupations of commissioning, verifying, and taking responsibility are institutionalized first"? From the next section on, this slope is extrapolated to 2027–2029, and concrete occupational profiles are drawn according to the three-tier method.

03The First Lineage: Verification and Responsibility — The Job of "AI Output Auditor"

Picture a promotional-material review department. An era in which generative AI drafts of educational materials arrive at the review team by the dozen each day has already begun at some companies. The speed of drafting has gone up, but the step that decides whether something can go out has not disappeared. On the contrary: the more output there is, the more the pace of the whole operation is set by the person who stamps "approved." Here is the sprout of a new occupation. Call it, provisionally, the "AI output auditor": the job of examining AI-produced text, figures, and data against the law (in pharma, Japan's Pharmaceuticals and Medical Devices Act) and internal rules, and taking responsibility for the approval signature.

You may think, "isn't that the same as today's reviewer?" Half right. Today's reviewers are unmistakably the sprout. But on the trend line, the content of the work changes. From reading human-written materials one by one, to auditing AI-drafted materials in bulk. Beyond catching individual errors, the auditor comes to grasp tendencies — "under what conditions does this AI make what kinds of mistakes?" — and to design the checking net itself. On top of that, a role emerges as designer of the audit trail (= the record of when, by whom, and on what grounds something was approved). In a regulated industry, designing records that can explain, in an inspection or a lawsuit, "how the AI's output was verified before release" should be worth as much as the signature itself.

The view that responsibility stays with humans to the end has grounds. First, the architecture of current regulation. Both the advertising rules of the Pharmaceuticals and Medical Devices Act and the industry's guidelines on sales information activities assume the responsible party is a company and the people inside it — not an AI. AI has no legal personhood (= the status of holding rights and duties under law) and cannot be penalized. Second, the framework taken by the EU's AI Act (the AI regulation enacted in 2024) — mandatory human oversight for high-risk uses — has become a reference point for regulators worldwide, and medicine and pharmaceuticals will very likely stay on the high-risk side. Third, an economic reason. Neither insurance nor contracts can function without a party that accepts responsibility. A deliverable no one answers for will not enter commerce, however accurate it is.

• Sprout (now): material reviewers and regulatory affairs staff examine human-written deliverables against the rules
• Trend line (scenario for 2027–2029): bulk auditing of AI drafts, mapping the AI's error tendencies, and designing the audit trail become the main duties
• Uncertain point: how far auditing itself gets handed to AI. First-pass checks may go to AI, with humans moving up to "auditing the audit"

Let me be honest about the uncertainty. A substantial share of audit work can itself be automated. Detecting typos and prohibited expressions is an area where a growing number of reports find AI misses less than people do. So the auditor of 2029 is not a person with a red pen, but a person who designs the AI's first-pass audit, decides the exceptions, and signs at the end — that is the most plausible extrapolation today. It would fail if regulators changed the system to allow conditional automatic approval by AI; but in the pharmaceutical domain, no sign of that happening by 2029 is visible in the public record.

04The Second Lineage: Context Supply — Translating In-House Knowledge into a Form AI Can Read

Examine AI failure cases and an unexpected common thread appears. More failures come from the AI not being handed the necessary background information than from the model lacking capability. For example: "this expression was sent back in review three years ago," or "for this indication, an internal rule says to avoid this phrasing." Knowledge like this usually sleeps inside veterans' heads and in mountains of unsearchable meeting minutes. However clever the latest model, it cannot obey a prohibition it was never given. What separates successful AI adoption from failure is, as much as model performance, the work of putting in-house knowledge into a form AI can consult.

Jobs specializing in this work are already starting to get names: context engineer (= someone who designs the background information handed to an AI) and knowledge curator (= someone who selects and organizes scattered in-house knowledge). In contrast to the prompt engineer (= someone who refines the instructions given to an AI), who was briefly celebrated and then faded, these roles have staying power. The reason is simple: clever instructions become less necessary as models get smarter, but company-specific context can only ever be supplied from outside, no matter how smart the model gets. Look at the past year's model lineage (Opus 4.7 to 4.8, Sonnet 5, Fable 5): what evolved was reasoning and the handling of long context — not the power to guess "your company's unwritten rules."

Brought down to pharma, the outline of this job can be drawn quite concretely. Past review comments; the reasoning behind answers to inquiries (= questions) from health professionals; expressions that were sent back, and why. A dedicated role translates these into a structure AI can search and consult — tagging, writing out the reasons behind decisions, resolving contradictions. Read the source documents, put the logic behind each decision into words, and shelve it all in an orderly system. Trace this job's lineage and you arrive at the librarian, or the archivist of regulatory documents. Rather than a wholly new occupation, it is closer to the truth to say that the librarian's craft is being revalued in the AI era.

Concretely, the work comes down to three pillars. One is putting tacit knowledge into words: turning a veteran's "somehow this feels wrong" into explicit rules with reasons and conditions attached — converting knowledge that would leave with a retirement into an asset the AI can consult. Another is structuring past decisions: organizing prior review comments and inquiry responses into "situation, decision, reason" form, which lets the AI pull up similar cases and makes answers more consistent. And the unglamorous but indispensable third is managing prohibitions and contradictions: reconciling internal rules that differ between departments, setting priorities, and unifying them — because an AI handed contradictory context returns contradictory output.

Writing speculation as speculation: whether this becomes a formal post in 2027–2029 depends on where each company sends "the bill for AI failures." At companies that keep treating context-starved failures as user error on the front line, this job will not be born. At companies that analyze failures and trace the cause to inadequate context supply, investment in a dedicated role becomes justifiable. My reading is that the shift from the former to the latter happens at the moment AI use turns from trial into operational infrastructure — and given the slope of model evolution, I expect that moment no later than 2027.

05The Third Lineage: Bridging — Interpreters Between AIs, Between AI and People, Between AI and Regulation

On development floors in 2026, it is becoming normal for one job to be split among several AI agents (= AI programs that carry out tasks autonomously): a researcher role, a drafter role, an inspector role. Anthropic's model lineage — Opus 4.7 to 4.8, then Fable 5 — has likewise evolved less toward single responses and more toward "the power to carry a long process forward without falling apart midway." As processes get longer and the AIs involved multiply, an unavoidable question appears: who designs the "working agreements" among the AIs?

In a human organization, the division of duties between departments (= the lines that fix who is responsible for what, and how far) is set by rules and custom. For multi-agent systems (= arrangements in which several AIs cooperate), a human has to write the equivalent of those rules. When the researcher AI could not confirm something, how does the drafter treat it? When the inspector raises a doubt, is the work sent back, or escalated to a human? Get this design wrong, and the AIs — each behaving correctly on its own — will calmly produce a deliverable that is wrong as a whole. Here a new occupation is born: the designer of inter-agent divisions of duty. The signs already exist. As of 2026, the design and supervision of multi-agent operations is starting to appear in software-company job listings under names like "AI orchestration (= directing multiple AIs) lead" (per public reporting).

Further along this lineage, a heavier job comes into view: the accountability officer, who translates an AI's decision process into a form regulators and auditors can understand. The AI's internal processing itself is unreadable to humans. But a record of the process — "based on which inputs, applying which rules, with human confirmation inserted at which points" — can be kept in a form that withstands audit. As the EU's AI Act (= the European AI regulation enacted in 2024, which imposes record-keeping and explanation duties on high-risk uses) is applied in stages, this translation work becomes a legal requirement. Knowing the technology alone is not enough; knowing the regulation alone is not enough. It takes an interpreter who speaks both languages.

Pharmaceutical safety is where I expect this occupational profile to take concrete shape first. In pharmacovigilance (= drug safety monitoring: collecting and evaluating adverse-event information; "PV" below), the adoption of AI for signal detection (= statistically picking out, from masses of adverse-event reports, hints that "this drug and this symptom appear together too often") is already under way. But judging whether a hint the AI picked up amounts to a "safety concern" is the evaluating physician's job, and regulation keeps it that way. The problem is the space in between. On what grounds did the AI pick up a hint, and what did it not pick up? How were thresholds (= the cutoff values for detection) set, and how is their change history kept? An intermediary who explains this to the evaluating physician and reduces it to records that survive a regulatory inspection — call it, provisionally, the signal-detection intermediary — spans three fields: statistics, AI behavior, and PV regulation. It is not an extension of any existing role. The boundary itself is the job.

06Will Qualifications and Training Catch Up? — What Happens by 2029, and What Does Not

No qualification exists for any of the new occupations described above. And looking back at history, that is not an abnormal state. The certification system for medical representatives, too, came after the work existed on the ground: industry custom built the training framework first, and certification was granted its form afterward. Occupations create qualifications; qualifications do not create occupations. It is natural to expect AI verification and context design to follow the same order.

So what happens by 2029? First comes in-house training. In 2026–2027, companies feel their way toward what to teach the people who review AI output. Next, private organizations and training firms extract the common core of that groping and package it as private certification. This is the same path by which vendor certifications in cloud technology (= credentials issued by the providing company itself) became de facto hiring criteria in IT before any national license did. Once industry bodies begin touching on "confirmation arrangements when using AI" in guideline-level documents, demand for certification will take off quickly.

Let me also write down what does not happen. My scenario is that national licensing does not arrive by 2029. A national license can only be designed once the definition of the job has stabilized, but the boundary between AI and human work moves with every model generation. When the generation after Fable 5 arrives, part of what is "for humans to verify" today will move to the AI side, and new verification problems will appear. You cannot aim a national license at a moving target. To state the falsifying condition too: a major AI-caused medical accident that sends legislation racing ahead. Even then, it would come first not as a license but as operational regulation (= rules requiring human confirmation for specified tasks).

For the reader, the practical conclusion comes down to this: do not wait for a qualification. Whichever of these new roles you move toward, the shared question will be whether you can show, in words, "why did you make that judgment?" In material review, in medical inquiry responses, in safety evaluation — start today the habit of recording the grounds for the judgments you are already making: what you looked at, what you compared it against, where you hesitated, why you chose as you did. That record becomes teaching material for the job of giving AI context, a yardstick for the job of verifying AI output, and a manuscript for the job of explaining to regulators. Putting your decision process into words is the one shared preparation you can begin before any qualification system catches up — and it is something AI cannot write for you. It exists only inside your head.