01What is an AI Agent? Defining it precisely in a marketing context
Let's pin down what we mean by "AI Agent" up front. An agent is a stand-in — something that acts on your behalf. In a marketing context, an AI Agent is an AI that, in your place, deals with websites, apps, and APIs (the doorways software uses to exchange data), and runs the whole flow: research, compare, decide, and actually carry out the transaction. It is not a "chatbot" that only returns canned answers, nor an "automation script" that repeats a fixed task. What makes it different? Four points.
Plain automation just "repeats a fixed procedure." An AI Agent "takes a goal you hand it, breaks it into small tasks, and works out the order itself." For example, it turns a loose instruction like "optimize next month's web ads" into: pick the channel → submit the ad → adjust the bid → check the results — on its own.
Earlier AI only "answered questions." An agent "opens a browser, fills in the fields, and even completes the payment." Anthropic's Computer Use (the company behind Claude) and OpenAI's Operator (the company behind ChatGPT) literally read what is on the screen and return mouse and keyboard moves.
It carries memory across conversations, remembering past purchases, price-negotiation exchanges, and preferences, and uses them in the next decision. The party a CRM (the system that stores customer information) talks to shifts from "the customer" to "the customer's agent."
When an agent makes a decision, who is responsible — the user, the service provider, or the company that built the agent (Anthropic/OpenAI/Google)? It raises a new problem the old way of drawing responsibility lines cannot settle. Section 8 covers this in detail.
Only when all four hold does it mean anything to call something an "AI Agent." Calling things that miss these an "Agent" just muddies the conversation. In this series we call an AI an Agent only when all four are present — planning + execution + memory + responsibility.
02Agentic Commerce ── Reality depicted by Claude / Operator / Mariner
In October 2024, Anthropic released Computer Use. It gave an AI (the new version of Claude 3.5 Sonnet) the ability to "look at a screenshot and return where to click." Around the same time, OpenAI announced ChatGPT Operator (for paid plans, January 2025) and Google announced Project Mariner (a browser extension you add on, December 2024). Three major companies all put "AI that drives a browser itself" on sale as real services. These are not prototype demos; pay and you can use them.
Beyond that, handing the shopping itself to an agent has begun as Agentic Commerce (commerce run by agents). Amazon's Rufus turns product search into a conversation, and Salesforce's Agentforce hands the first touch of B2B sales to an agent. Combine the Agent Toolkit that the payments company Stripe released in 2025 with Anthropic's Claude Agent SDK, and a setup where the agent calls the payment doorway (the payment API) itself and settles the bill is becoming the norm.
This trend confronts the marketing side with four new realities.
It is no longer only "people" who reach a website; more and more, "an agent reads the page's contents mechanically." Unlike a person, an agent does not care about dwell time or visual flourishes. How you make information a machine can read accurately now feeds straight into CV (conversion — getting someone to actually sign up or buy on the site).
An agent can compare 50 sites in seconds. It boils down price, performance, and review highlights and shows the person only the "top 3." A product that fails the agent's yardstick disappears before a human ever sees it.
For things like "book a hotel," "restock daily goods," or "review an insurance plan," more cases finish with the agent alone. Transactions with no human in the loop increase. So the audience marketing has to persuade becomes both the person and the agent.
An agent prefers organized data (structured data — information neatly split into fields so a machine can read it) over prose. The goal of SEO shifts from "show up in a person's search results" to "get picked as a source the agent refers to." Content an AI can read easily comes to decide your traffic over the long run.
In the pharmaceutical field, when a doctor or pharmacist asks an AI Agent (say, Claude or ChatGPT's Operator mode) to "compare how drug X works," the agent goes to look at the PMDA (Pharmaceuticals and Medical Devices Agency) package inserts, each company's product site, medical papers (PubMed), and the Japan Pharmaceutical Association's guidelines. Will the agent find you properly, and read you correctly in a machine-readable form? Building for that becomes the top priority for healthcare-professional websites from here on.
03Agent-to-Agent (A2A) marketing – a new negotiation forum
By the end of Section 2 the flow "person → agent → product" came into view. Beyond it, a world where "your agent ↔ the seller's agent" deal with each other directly — A2A (Agent-to-Agent) — has begun. The MCP (Model Context Protocol) Anthropic released in November 2024 is a shared rule (like a standard socket) that lets an agent connect safely to outside information and tools. The A2A protocol Google announced in 2025 is a shared rule for agents to exchange information and converse with each other. Think of both as agreed conventions so machines can talk to each other in the same way.
Here is a typical A2A marketing case, translated into the pharmaceutical context.
The department that buys drugs at a hospital runs an AI Agent that exchanges price, delivery date, and stock with each maker's agent every week. Instead of an MR (a pharmaceutical company's sales rep) visiting, the agents compare several offers in a short time and show the pharmacy the final three. The MR's job shifts to filling in "what the agents cannot pick up" — that doctor's treatment preferences, how a clinical trial is progressing.
Looking at a patient's chart, a doctor asks the AI Agent to "compare the treatment options that fit this patient." The agent reads and weighs guidelines, the latest papers, and package inserts from your company and others mechanically, then lays out the options. What a pharmaceutical company wants to convey passes once through the sieve of the agent before it reaches the doctor.
When a patient says "the side effects are hard to bear," the agent (1) checks the warnings in the package insert, (2) sends a message to the attending doctor, and (3) anonymously gathers how other patients with the same symptom coped. For a pharmaceutical company's Patient Support Program, being built to exchange data with agents (API integration) becomes an essential condition.
A pharmaceutical company's compliance agent constantly watches inquiries from the PMDA, updates to the industry's voluntary rules, and competitors' promotional materials, and automatically checks whether its own materials are out of step. The materials-review staff then put their energy into "the ethically borderline parts an agent cannot fully judge."
Once these agent-to-agent exchanges become normal, the marketer's work shifts its center of gravity from making promotional material that people read to building the information doorways and data that agents read (APIs, machine-readable structured data, MCP servers, and so on). This is not a grand prediction about the future; it is a reality Amazon, Shopify, and Salesforce are already working on.
04SGE and AI Overview ── Crustal changes that turn search into “answers”
Before we get into agents, another big change is already underway. With Google's AI Overview (the feature that puts an AI summary of the answer at the top of your search; it used to be called SGE — Search Generative Experience), widened in 2024, OpenAI's ChatGPT search (formerly SearchGPT, folded in in October 2024), Perplexity, and others, search results shifted from a "list of links" to an "answer on the spot." Even at this stage, before agents spread, the path by which a person reaches your site has already been rewritten.
Let's look at what this means, concretely.
Search "side effects of drug X" and AI Overview shows a summary on the spot, so the person never comes to the site. Visits to a pharmaceutical company's site fall structurally. The flip side: whether the AI cites you as a source for that summary now governs how much the brand is seen.
Which sites AI Overview cites is decided by how trustworthy the content is, how current it is, and how well the data is organized. A pharmaceutical company's official site is judged on whether the package insert is up to date, whether the Q&A is complete, and whether the evidence is shown openly. SEO turns into SEO for being cited.
AI Overview puts "Company A vs. B vs. C" on a single screen. The pharmaceutical company's contest moves from "get people to come to our site" to "be judged fairly in the AI's comparison table." Information quality, ease of understanding, and update frequency become the difference directly.
It can happen that AI Overview's sources leave out the pharmaceutical company's official information, pick up stale information, or cite a site carrying wrong information. That distorts a patient's judgment. Delivering official information in a form the AI can easily find becomes a corporate social responsibility.
A circumstance unique to pharma: in many cases the information AI Overview or an AI Agent cites is tested on whether it stays within the rules. Touching on an unapproved use, comparing unapproved drugs, summarizing patient information improperly — these can be created and spread by a third party (the AI) beyond the pharmaceutical company's reach. This is a new problem the old rules for promotional material never anticipated. Section 8 covers it in depth.
05Paradigm shift from “delivering to people” to “delivering to agents”
Pulling the moves in Sections 2–4 together, the basic premise of marketing changes. From the 20th century to the early 2020s, marketing was a contest to grab human attention. Advertising went after people's eyes, CRM after people's memory, content after people's time, and brands after people's emotions. Every strategy was a scramble over the limited resource of human attention.
The AI Agent era is different. Getting onto the flow an agent processes is the new point of competition. Unlike people, an agent has unlimited attention, processes in an instant, and holds no emotion. In exchange, the sources it consults are limited, its judgment is mechanical, and how it decides cannot be seen from outside (a black box). Marketers have to rebuild their strategy around this new counterpart.
You need to make both content that moves people's emotions and organized data an agent can read mechanically — from the same facts, at the same time. The two are not opposed; they support each other.
On top of the old metrics (click-through rate, conversions, dwell time), you need machine-side yardsticks too: "probability of being cited by an AI," "rank in an agent's evaluation," "number of accesses to your MCP server," and the like.
An agent does not read a brand's "mood." Even so, to be chosen from the shortlist an agent shows a person, it matters that the brand stays in the mind of the human who decides last. The brand comes to work at the seam between agent and person.
An agent cannot handle closed (externally invisible) information. To get delivered through an agent, you have to publish product information, prices, stock, and regulatory information in a machine-readable form. Keeping information hidden becomes a competitive disadvantage in itself.
For the pharmaceutical industry to keep up with this change, the websites for healthcare professionals, the information sites for patients, and the in-house systems that support MRs all need to be rebuilt in a two-track structure, one for people and one for agents. This is not a mere system upgrade; it is a rethink of content strategy itself.
06Agent Experience (AX) design principles
Just as the idea of UX (User Experience — how it feels for a person to use a service) changed the Web in the 1990s, the idea of AX (Agent Experience — how usable it is seen from an agent) will change corporate sites from here on. AX is how "easy to use" your information is when an AI Agent accesses it, sizes it up, and uses it. Here are the four principles of AX design.
Keep product information, package inserts, indications, and side effects in a machine-readable form. Concretely, use Schema.org (an effort that sets shared "tagging rules" for information on the Web; its medical items include MedicalEntity, Drug, MedicalStudy), HL7 FHIR (an international common format for exchanging medical data), and your own JSON-LD (a way of embedding machine-facing notes inside the page). An agent reads this organized data before it reads prose.
MCP, which Anthropic turned into a shared rule, is the "doorway" that lets an agent connect safely to your information. In the future a pharmaceutical company will publish an "MCP server for its own product information" and be queried directly by Claude or ChatGPT agents.
An agent prefers to consult "the trustworthy original source." For a pharmaceutical company's official site to be delivered through an agent, it has to beat other media on how current its package inserts, clinical-trial data, and regulatory updates are.
An agent judges how reliable a source is. A build that makes clear who said what, when, and on what basis (update timestamps, author attribution, citations to the underlying literature) raises an agent's rating of you. This goes hand in hand with readability for people.
The four AX principles are not about throwing away UX (how it feels for people). Doing UX and AX together is the new goal. Beautiful and easy for a person to read, and at the same time accurately readable for an agent — striking both becomes the norm for corporate information from here on.
07Implications for pharmaceuticals ── Dialogue design between MR Agent and physician Agent
From here we narrow to matters specific to pharma. The job of the "MR" (Medical Representative — a pharmaceutical company's staff who bring drug information to doctors and pharmacists) has long run on "people meeting in person to convey information." A detail (a face-to-face meeting), a Lunch & Learn (a study session over lunch), a booth at a conference — all assumed meeting in person. The MR in the AI Agent era takes on four new roles.
Grasp which AI Agent (Claude, ChatGPT, Doximity GPT, OpenEvidence, and so on) a doctor uses and which sources that agent looks at. "Getting your company's information to the doctor's agent accurately" is added to the MR's goals.
Keep a system that can answer the information queries a doctor's or pharmacist's agent sends — efficacy, dosage and administration, adverse effects, interactions, the latest evidence — from accurate primary sources inside the company. What is handled here is information only: price, stock, and delivery — commerce and logistics — are not the MR's job; they belong to the pharmaceutical wholesaler (distribution) and the hospital's purchasing and pharmacy departments (an MR does not negotiate price). The MR puts energy into "the areas that can't be left to an agent — trust, on-the-spot circumstances, relationships."
Maintain package inserts, clinical-trial results, and guideline updates in three forms at once — for doctors, for patients, and for agents. The MR carries "the patterns in what doctors in the field tend to ask" back to the company and keeps improving the AX build.
An agent can calculate trust but cannot build it. For the weightiest decisions — prescribing a new drug, handling a side effect, explaining to a patient — trust between people is still needed. The MR's role does not thin out; it concentrates on the moments where the decision matters most.
On the flip side, in delivering information straight to patients (DTC — communication aimed at consumers), distribution through agents quickly takes the lead. Patient support programs, side-effect consultation desks, content that teaches how to take a medicine — building all of them on the assumption that they "reach the patient via an agent first" will become standard within five years.
08Regulation/Ethics ── Boundary of Pharmaceutical and Medical Device Law × AI Agent
As agent-run commerce spreads fast, a heavy question is put to the pharmaceutical industry's regulatory framework. Japan's Pharmaceutical and Medical Devices Act, the US FDA (Food and Drug Administration) regulations, and the EU's EMA (European Medicines Agency) guidelines were all written assuming promotion carried out by people. How current rules should treat the following acts by an agent is not yet settled.
If a doctor's agent answers, "for this patient there is room to consider drug A in an off-label way," does the pharmaceutical company bear responsibility? Did the agent reason this on its own, or did it build on information the company put out? Being able to trace which information it came from (source tracing) becomes a new must.
How do today's comparative-advertising rules treat an agent building and showing a "comparison table of drugs A and B"? Even if the pharmaceutical company never approved that comparison, is a comparison that arises because the company's information was cited subject to the rules?
When an agent summarizes side-effect information and shows it to a patient, the warning may come out weaker or stronger. A rewording beyond the pharmaceutical company's control spreads widely. How do you design an obligation that keeps the agent able to follow a reliable link back to the original source?
When a patient is harmed by an agent's wrong recommendation, how to split responsibility among the company that built the agent (such as Anthropic), the side using the agent (the medical institution), and the side that put out the information (the pharmaceutical company) is unsettled. A pharmaceutical company has to pay the utmost attention to a build where its information cannot be misread.
The regulators (the PMDA, the Japan Pharmaceutical Association, the industry's voluntary standards) are also working on a framework to meet these points. Compliance and regulatory departments at pharmaceutical companies will make setting up internal rules for distribution via agents, on top of the conventional review of promotional material, a priority for 2026–2027.
094 Steps to Start Now - Implementation for Pharma Marketers
Now let's bring the discussion down to actual steps. The shift into the AI Agent era takes several years, but there are four concrete steps a pharmaceutical marketer can start next week.
For your company's healthcare-professional site, have Claude or ChatGPT "compare how drug X works," and watch how your information does or does not get cited. Analyze why it isn't cited. Make it a once-a-week habit.
Put package inserts, indications, side effects, and drug-price information into a machine-readable form with Schema.org / JSON-LD (tag it). What shows on screen stays the same; you only add the machine-facing notes behind it. Team up with your SEO staff and it can be done within six months.
Build an MCP server that returns your product information as an in-house prototype (a POC — Proof of Concept, a small experiment to test whether it works). With Anthropic's MCP development kit you can get a working version in about a month. It becomes material for your MR, regulatory, and IT people to talk over.
Within three months, draft three things: "notation rules for when company information goes out via agents," "a policy for how to answer inquiries from agents," and "employee rules for using agent output outside the company." Starting the discussion matters more than how finished it is.
The priority order of the four is STEP 1 → 2 → 4 → 3. Without STEP 1 (the audit) you can't see where you stand, and STEP 2 (organizing data) is easy to push ahead on its own, technically. STEP 4 (making rules) needs internal agreement, so it takes time. STEP 3 (the MCP server) is more meaningful once you take it up after the STEP 1–4 discussion.
The marketing of the AI Agent era this issue covered came down to one question: how does a company face the fundamental change that the audience is no longer only people? What Anthropic Computer Use, ChatGPT Operator, Project Mariner, Rufus, and Agentforce have shown is the reality that agents which drive a browser in a person's place, compare products, and even reach a decision are already running as real services. As MCP (the shared doorway standard that connects agents), A2A (agents dealing with each other), AI Overview (the AI summary shown above search), ChatGPT search, and Schema.org's medical tags combine, the way a pharmaceutical company's information gets "read" shifts its center of gravity from web pages toward APIs, organized data, and forms a machine can trust. Information design, APIs, organizing data, regulatory response — what this issue covered was rebuilding the side of marketing you can put together with logic.
Next time (Part 3) steps into the other big area logic can't fully capture — the side of feel and atmosphere. Vibe Marketing ── designing a brand's "atmosphere" with AI. The word "Vibe" (the air, the mood) that spread fast on social media from 2024 to 2026 comes from a realization: a brand's air, tone, and emotional texture can be treated as something AI can design and reproduce, not just left to human instinct. Now that generative AI can turn out images, video, music, and text in bulk, a brand's decisive ground shifts from "what to make" toward "what kind of atmosphere to give off." We'll dig into this theme in depth in the context of pharmaceutical brands.
- As a result of the commercialization of AI Agents (Claude/Operator/Mariner) as "proxy buyers," marketing targets have come to include not only people but also agents. When Anthropic Computer Use (October 2024), ChatGPT Operator (January 2025), and Project Mariner (December 2024) were commercialized from the three major platforms, the trend was already irreversible.
- Agent reads the authority of structured data, MCP servers, and primary information. Agent Experience (AX), which designs these, is a new design area alongside UX. Schema.org's medical schemas (MedicalEntity, Drug, MedicalStudy, etc.), HL7 FHIR, and the MCP protocol published by Anthropic in November 2024 will be the core of pharmaceutical companies' "Agent-readable" designs.
- Pharmaceutical marketers can get started next week with four steps: AX audit → structured data → internal guidelines → MCP server POC. Regulatory compliance and ethics will be promoted in parallel with regulatory development. Starting a discussion is more important than completeness.