AI agents will redraw your org chart long before they move your P&L

Dushyant SapreFounder & CEO, SwishX·9 min read
Contents
  1. Introduction
  2. The returns are thin so far, and the surveys say so
  3. Agents collapse production and inflate everything around it
  4. Programmatic advertising already ran a version of this experiment
  5. The scarce people will be the ones who own the source of truth
  6. Headcount is the wrong argument to have first
  7. Four moves for the next planning cycle

Agentic AI barely shows up in life sciences P&Ls yet, and the surveys say so plainly. The bigger change is already underway in who does the work, and it rewards the companies that redesign roles before the returns arrive.

AI agents are going to change the shape of life sciences commercial organizations before they change the economics. The org chart moves first. The P&L follows later, and only for the companies that redesign the work instead of bolting agents onto the old one.

That is not how the debate usually runs. Most boardroom conversations about agentic AI in pharma end up at a single number: how many roles it removes. I think that is the wrong question, asked too early. The more useful one is which roles become scarce once producing a draft costs almost nothing, and whether you have the right people sitting in them.

The returns are thin so far, and the surveys say so

Start with the evidence, because the hype is loud. Deloitte's 2026 life sciences outlook surveyed 280 C-suite executives in August and September 2025. Only 22% said they had successfully scaled AI, and 9% reported significant returns. Agentic AI appeared as a new category that year, cited by 30% as a trend shaping their 2026 strategy.

The cross-industry picture is similar. McKinsey's 2025 State of AI survey found 23% of organizations scaling an agentic system somewhere in the business, but no more than 10% scaling agents in any single function. Only 39% attributed any EBIT impact to AI, and most of those put it below 5% of EBIT. BCG's September 2025 study of 1,250 executives classed 5% of companies as built for AI and found 60% reporting minimal revenue and cost gains.

Gartner went further in June 2025, predicting that over 40% of agentic AI projects will be canceled by the end of 2027. It also estimated that only about 130 of the thousands of vendors selling agentic products offer the real thing. Johnson & Johnson's experience points the same way: after overseeing as many as 900 AI experiments, it found that 10% to 15% of use cases drove about 80% of the value, and refocused on those.

So if you are waiting for the P&L to tell you whether agents matter, you will wait a while. That is fine as a financial judgment. As an organizational one, it is a trap.

Agents collapse production and inflate everything around it

A simple way to see what is happening is to split any piece of commercial work into four jobs. Someone specifies it: the objective, the audience, which claims can be made and on what evidence. Someone produces it. Someone checks it against the label, the references and the brand rules. And someone governs it, deciding who may approve what, keeping the record and answering for it later.

Life sciences commercial teams were built around the second job. Brand teams, agencies, medical writers and production studios exist mostly to make things. Agents are very good at exactly that job, and not yet trustworthy at the other three without people.

Agents raise the stakes over the chat assistants of the last two years. An assistant drafts when asked. An agent takes a goal, breaks it into steps, calls tools and hands its output to the next step, sometimes to another agent. So the first draft is no longer the only thing made without a person watching: so is the routing, the first compliance pass and the resizing for each channel. Every one of those steps needs a record of what it was told, which source it used and who signed off, or the governance job becomes impossible to do after the fact.

The clearest public example I know comes from the clinical side. Anthropic's case study on Novo Nordisk reports that the time spent producing clinical study documentation fell from more than 10 weeks to 10 minutes, and that writing time on clinical study reports dropped by 90%. The detail I find more telling is the stated goal: to get documentation into human hands for review and approval sooner. Review cycles were cut in half, not removed. The drafting collapsed, and the checking became the work.

I expect the same pattern across commercial content, from detail aids to approved email to field coaching. When production approaches zero cost, volume rises, and every extra asset lands on the same few people: the ones who decide what is true, and the ones who confirm it. Agents do not remove the bottleneck. They push it upstream into specification and downstream into review, and make it impossible to ignore.

Programmatic advertising already ran a version of this experiment

Having spent years in ad tech, I recognize the shape. Programmatic buying automated the transaction in display advertising: the insertion orders, the negotiations by phone, the manual trafficking of ad tags. It did not stay at the margin. eMarketer forecasts that programmatic will account for 90% of display advertising budgets worldwide by 2026.

What changed most was not the number of people in media. It was which people mattered. Value moved to whoever owned the inputs and the rules: the audience definitions and the data behind them, the frequency caps and brand safety policies, the judgment about which supply could be trusted. And a whole layer grew up to check the machine. DoubleVerify, one company paid to verify where ads ran and whether a person could see them, reported $748.3 million in revenue for 2025.

In my view the people who came out ahead were rarely the fastest at executing trades. They were the ones who could write the rules the machine executed and read its output with suspicion: planners who understood the data, analysts who could tell when an optimization was gaming its own metric, and client leads who could explain why the dashboard and the business result disagreed. The roles that shrank were the ones that existed to move information between systems by hand.

The analogy has limits, and they matter. Media buying optimizes toward a number, and a bad placement costs money and some embarrassment. A promotional claim is closer to binary: it is on label and supported, or it is not, and the consequence is regulatory. So life sciences cannot bolt checking on afterwards the way advertising did with verification. It has to sit inside the work from the start. Programmatic shows the direction of travel. It does not show the safety margin.

The scarce people will be the ones who own the source of truth

If production is cheap, the scarce asset is whatever every agent writes from: the approved claims, the evidence behind each one, the brand rules, the safety language, the audience definitions. In most companies that knowledge is spread across a claims matrix in a spreadsheet, a messaging house in a slide deck, reference PDFs in a DAM and the memory of two or three senior reviewers. It holds together because people fill the gaps. Agents do not fill gaps. They guess, or they stop.

That turns ownership of the source of truth into a real job. Someone has to keep the claims library current on the day the label changes, retire claims whose evidence has aged out, and say which reference supports which sentence in which market. I would argue that person becomes one of the most important people on a brand. Most organizations have not named them, and many have handed the work to whoever has spare evenings.

For a global company the source of truth is not one thing. The same molecule can carry a different label in the US, the EU and Japan, with different indications and different safety language, and each market's code sets its own limits on promotion. An agent working from a global master will produce something fluent and wrong in at least one market, which is why a perfect translation can still be an off-label claim. So either the owner role multiplies by market, or the source carries market rules explicitly. In both cases the work sits with people who understand the label, not with the people who used to build the slides.

The early moves by large companies point the same way. Pfizer's internal generative AI platform, Charlie, built with Publicis, was reported in 2024 to sort content into red, yellow and green risk tiers to flag what needs medical review. Veeva made a Quick Check agent available in PromoMats in December 2025 that reviews content against internal guidelines before it reaches MLR. Both are investments in triage and checking, not drafting. Moderna went further on structure and put talent and digital technology under one executive, a Chief People and Digital Technology Officer, before anyone could show the return on that choice.

Reviewers change too. A medical or regulatory reviewer who reads finished assets line by line is the most expensive possible use of that expertise. The better use is upstream: approving claims and rules once, so every asset built from them inherits the decision, and spending review time on what is genuinely new. MLR review does not go away. It shifts from inspecting outputs to certifying inputs.

Headcount is the wrong argument to have first

The case for leading with headcount deserves a fair hearing. Commercial cost is large, finance teams are under pressure, and in McKinsey's survey 32% of respondents expected AI to reduce their overall workforce in the coming year, against 13% who expected growth. If an agent drafts in minutes what an agency took weeks to deliver, someone will reasonably ask why the budget has not moved.

My answer is that demand for content in this industry is elastic and the supply of approval is not. Every new indication, market, segment and channel multiplies the number of assets a brand could use. Today that demand is rationed twice, by production cost and by review time. Remove the production constraint and all of the rationing falls on review. Cut reviewers or brand owners to pay for agents, and you get more drafts and fewer approved assets.

The second-order effects are where the money eventually shows up. Agency economics change first, because a model priced on hours of production is exposed when production stops taking hours. Field roles change next: Veeva's pre-call agent already assembles insight before a rep's visit, which moves the rep's value toward judgment in the room. Global companies feel it most, because each market has its own label and its own code, so the checking load grows with every country even when the draft is free.

None of this appears cleanly in a quarter. It appears as cycle time, as the share of assets reused rather than rebuilt, and as the number of people a brand needs to clear a launch. Those are org design numbers long before they are P&L numbers.

Four moves for the next planning cycle

First, name the owners. For each brand, one accountable person for claims and evidence and one for brand rules, with the authority to say no and the hours to do the job properly. If the source of truth is maintained on goodwill, your agents will write from stale facts at high speed.

Second, measure the checking step instead of the drafting step. Track time from first draft to approval, first-pass approval rate, and how often reviewers send work back for problems the source should have prevented. Draft speed is now a vanity metric. It is also why AI that can't cite its sources stalls: it is fast at the one step that was never the constraint.

Third, move reviewers upstream before you add more agents downstream, and rewrite agency contracts around approved output rather than hours. Fourth, stay skeptical. Ask every vendor selling agents which of the four jobs its product does and what a human still signs. If the answer is all four and nothing, you are looking at agent washing.

From the outside, I would watch three signals. Whether large companies start hiring for roles with names like claims owner or content governance lead. Whether agency holding companies change how they price production. And whether reviewers start appearing in brand planning meetings rather than only at the end of the queue. Each of those is the org chart moving ahead of the numbers.

I am less sure about timing. It could take two budget cycles or five before the economics are obvious to an analyst. I am fairly sure about the order. The winners will not be the companies with the most agents. They will be the ones that can say, for every sentence they publish, who decided it was true.

Filed underPerspectives

Keep reading

All articles

Want to see Magic Studio on your brand?

Video, visual aids, ads, approved email and web, made from your approved claims and ready for MLR review.