Your Enterprise Software Is Becoming Invisible: What That Means for the Future of Pharma Websites

The Executive Thesis: The Interface Is Moving Up the Stack

The central change is not that enterprise software is disappearing. It is that intelligence is beginning to separate from the application interface. Today, the common model is:

Person → application → capability

The emerging model is:

Person → AI → capability → enterprise system

The employee no longer needs to know which application contains every part of a process. The AI can coordinate across systems. A marketing leader might ask: “Show me which HCP experiences need attention.” A content leader might ask: “Prepare the changes related to this labeling update.” A campaign manager might ask: “Find the approved assets for this launch.” A regulatory operations team might ask: “Identify every page containing this claim and show me what requires review.”

The AI determines which systems are needed to answer or execute the request. That changes the strategic question. Leaders no longer need to ask only:

“Which platform has the best interface?”

They also need to ask:

“Can our trusted content and business capabilities be safely used through any interface?”

That question has consequences for technology investment, operating models, governance, measurement and talent. It also changes what a website is.

The Shift Has Already Started

Enterprise software is increasingly being designed so users do not have to begin every task inside the application itself.

Adobe now allows compatible AI environments to interact with Adobe Experience Manager through its Model Context Protocol server. Users can perform certain content operations through natural-language requests while remaining subject to their existing AEM permissions. Adobe Experience League

Optimizely has introduced similar MCP access to its CMS environment, allowing compatible AI clients to interact with content models and operations. It has also begun providing analytics around AI-agent activity on websites. Optimizely

Contentful has created MCP capabilities that allow AI systems to work with content through its existing management APIs and permission structure. Contentful

Acquia has been emphasizing the growing importance of Drupal’s structured content model as organizations prepare content to serve both humans and AI systems. Acquia

Salesforce has also moved beyond simply embedding AI into CRM. Its AIforce strategy is designed to make Salesforce data, workflows, permissions and business logic available through multiple AI interfaces rather than only through the traditional Salesforce experience. Salesforce

These are vendor descriptions of specific products, not evidence that an integrated autonomous pharma workflow is already available. Availability, licensing and configuration vary. Model Context Protocol (MCP) connects AI clients to tools; the protocol itself does not confer permission or regulatory compliance. These announcements come from different vendors. Their strategic significance is similar:

The application may no longer be the primary place where work begins.

For pharma leaders, this means platform strategy must expand beyond interface quality. The more important questions become:

• Can the platform expose trusted information?
• Can it preserve permissions outside the application?
• Can it distinguish approved content from draft or expired content?
• Can it support controlled actions?
• Can it provide evidence and provenance?
• Can it participate in a broader enterprise workflow?

The shift has started at the platform level. The next question is why it matters so much in pharma.

Why This Matters More in Pharma

Enterprise AI becomes more difficult when “close enough” is not good enough. Pharmaceutical information is contextual. A sentence may be appropriate for:

• one product;
• one indication;
• one audience;
• one market;
• one point in time;
• and one approved use.

Change the context and the same sentence may become incomplete, misleading or inappropriate. Consider a request such as:

“Find every U.S. HCP page affected by the new indication language, identify the associated claims and assets, determine what may need re-review, and prepare the changes for the appropriate teams.”

Answering that request may require information from:

• the CMS;
• the DAM;
• Veeva or another regulated-content platform;
• claims and reference repositories;
• workflow systems;
• analytics;
• CRM;
• identity systems;
• and internal approval history.

No single application necessarily contains everything the AI needs. That is why pharma cannot approach AI as merely a conversational layer placed on top of existing systems. The AI must understand context. It must know which information is authoritative. It must know what is current. It must know who is asking. It must know what action is permitted. And it must know when a human decision is required. For pharma leaders, the business implication is direct:

AI readiness is not primarily a model-selection problem. It is a content, context and governance problem.

The Software Does Not Disappear. Its Role Changes.

The future is unlikely to be one in which companies stop using CMS platforms, CRM systems, DAMs, workflow tools or regulated-content platforms. These systems remain essential because they contain authoritative information and enforce business controls. What changes is how people reach those capabilities. The CMS may no longer be the place where every task begins. The CRM may no longer be the only place where customer work is initiated. The regulated-content platform may no longer be accessed only through its own interface.

Instead, AI may coordinate across them. That could reduce one of the least productive parts of modern knowledge work: navigating the enterprise itself. For executives, this creates a shift in investment logic. The question is not whether the interface still matters. It does. The question is whether interface quality is enough to differentiate the platform. Increasingly, it may not be. The platforms that matter most will be those that can make their capabilities:

• discoverable;
• understandable;
• permission-aware;
• auditable;
• reusable;
• and safe to invoke through other interfaces.

The application becomes less visible. The governed capability becomes more valuable.

The CMS May Matter More Even If People Use It Less

Pharma website teams have historically invested heavily in the CMS authoring experience. How quickly can someone create a page? How easy is preview? How reusable are components? How well does the platform support global sites and localization? Those questions remain important. But AI introduces a second set of requirements. Can the system make content understandable to machines? Can it identify which content is authoritative? Can it expose relationships among products, indications, claims, references and assets?

Can an AI agent distinguish approved content from draft or expired content? Can permissions still be enforced when the request originates outside the CMS interface? Can the organization trace exactly what the AI changed? Can the AI prepare a change without publishing it? Can it recognize when an action requires human review? These capabilities determine whether a CMS can function as part of an AI-enabled enterprise. The future CMS may therefore compete less on:

“How good is the editing screen?”

and increasingly on:

“How reliably can this platform provide governed, structured content to humans, systems and AI?”

That applies across platforms. Whether a pharmaceutical company uses AEM, Sitecore, Optimizely, Drupal, WordPress VIP, Contentful, Contentstack or another architecture, the underlying challenge is similar. The CMS must become more than a page-production tool. It must become a trusted source of reusable, contextual and governed content.

Capability Is Not Authority

One of the most important principles for regulated AI is simple:

Just because an AI can perform an action does not mean it should be allowed to perform it.

An agent may technically be capable of:

• reading content;
• changing content;
• generating new language;
• finding supporting references;
• creating a review package;
• submitting something into workflow;
• publishing content;
• or initiating customer outreach.

Those are very different levels of authority. Future enterprise systems should distinguish between:

Read The AI can retrieve information.

Recommend The AI can suggest what should happen.

Prepare The AI can create drafts, packages or proposed changes.

Submit The AI can initiate a controlled workflow.

Execute The AI can carry out an approved action.

Approve Consequential decisions remain with authorized humans where required.

This is more useful than simply saying a company should “keep a human in the loop.” The real questions are:

• Which human should be involved?
• At what point?
• For which decision?
• With what evidence?
• Under what level of accountability?

For pharma leaders, this creates a practical governance requirement:

Every AI-enabled capability should have an explicit action boundary.

That boundary should define what the AI may read, recommend, prepare, submit, execute or never perform without human approval. Without that distinction, AI access can become an uncontrolled expansion of enterprise authority.

The Website Itself Could Change Next

So far, this sounds like an internal enterprise-software story. It may not stay that way. Search behavior is already changing.

AI search can answer questions without a visit to every source. In a March 2025 U.S. browsing study, Pew Research Center found that users clicked traditional result links less often when a Google AI summary appeared. This is evidence about that sample, not a forecast of pharma traffic. Pew Research Center

That creates an important future possibility:

The next visitor to a pharma website may not always be a person.

An HCP might ask an AI assistant: “Find the latest manufacturer information on Product X.” A patient might ask: “What support programs are available?” A caregiver might ask: “What is the approved dosing information?” Eventually, an authorized agent might go further: “Arrange for Medical Information to contact me.” “Request a representative.” “Find the latest approved professional information.” The person may never open the traditional website. Yet the pharmaceutical company’s digital infrastructure may still power the experience. This changes the business definition of a website.

The website is no longer only a destination. It becomes a source of trusted information and a gateway to governed capabilities. That is the point at which the website begins to become invisible.

The Future Pharma Website May Serve Two Audiences

Today, most digital experiences are designed primarily for humans. In the future, pharmaceutical digital infrastructure will increasingly need to serve both humans and machines.

The human needs:

• clear language;
• intuitive navigation;
• accessibility;
• trust;
• appropriate visual hierarchy;
• useful interaction;
• and clear safety information.

The machine needs:

• structured information;
• explicit entities;
• clear relationships;
• provenance;
• content status;
• consistent terminology;
• dates;
• permissions;
• and authoritative sources.

These are different needs. But they should come from the same source of truth. That creates a future model with direct business implications:

One governed content foundation. Multiple interfaces.

The traditional website becomes one interface. An AI assistant becomes another. A search engine becomes another. An internal employee agent becomes another. An authenticated HCP agent may eventually become another. For pharma leaders, this means digital teams should stop treating every channel as a separate content problem. The strategic objective should be to create a governed foundation that can serve multiple channels without losing context, approval status or accountability. That is a more scalable model than maintaining disconnected experiences for every new interface.

From Website Platform to Digital Capability Platform

Executives may find it useful to think of the future architecture in five layers.

1. Systems of record

These contain authoritative information and workflows:

• CMS / DXP;
• DAM;
• CRM;
• Veeva and other regulated-content platforms;
• Medical Information systems;
• claims and references;
• labeling systems;
• analytics;
• workflow;
• identity and consent.

2. Context

This tells AI what the information means. For example:

Claim → evidence

Claim → indication

Claim → audience

Claim → market

Content → approval status

Asset → expiration date

Label change → affected content

HCP → permitted capability

This context layer may become one of the most valuable parts of the architecture because it allows AI to interpret information rather than merely retrieve it.

3. Governance

This determines:

• identity;
• permissions;
• risk;
• provenance;
• required review;
• human escalation;
• logging;
• and permitted actions.

4. AI orchestration

The AI determines what needs to happen and which governed capability to use.

5. Experience

The user may encounter the organization through:

• a website;
• an app;
• search;
• an AI assistant;
• an internal enterprise agent;
• an HCP assistant;
• or a future interface that does not yet exist.

The underlying enterprise does not disappear. It becomes less visible. For executives, the architecture suggests a change in ownership. The website team may no longer own the entire digital experience. Content, data, regulatory, commercial, technology, identity and governance teams will need to coordinate around shared capabilities. The future website is therefore not only a technology project. It is an operating-model project.

How SEO, GEO and Agent Readiness Work Together

Marketing leaders are beginning to think beyond traditional search. GEO builds on SEO; it does not replace it. Google says its existing SEO practices apply to AI Overviews and AI Mode, with no special AI schema or new machine-readable file required. Clear, indexable, helpful content remains the foundation. Google Search Central

SEO asks:

Can a search engine find and rank our content?

Generative Engine Optimization, or GEO, asks:

Can an AI system understand our content well enough to use and cite it accurately?

The next stage could ask:

Can an authorized AI safely interact with the capability behind the content?

That produces a useful progression:

SEO → Discovery

GEO / AEO → Understanding and citation

Agent readiness → Trusted interaction and action

For pharma marketers, the implication is important. GEO should not simply mean trying to manipulate AI-generated answers. The more durable strategy is making authoritative information easier for machines to understand correctly. That means strengthening:

• structured content;
• clear product and indication entities;
• claims and supporting evidence;
• publication and review dates;
• provenance;
• canonical sources;
• audience context;
• approval status;
• content freshness;
• semantic markup;
• and consistent terminology.

These improvements help traditional search. They also help AI systems interpret content. Eventually, they may help AI systems determine what information can be safely used. For marketing leaders, GEO is therefore not merely a communications tactic. It is an extension of content governance. The organizations that perform well in AI-mediated discovery will likely be those that have already made their information clear, current, attributable and context-rich.

The Marketing Metrics May Change Too

AI-mediated experiences create a measurement problem. Consider an HCP who asks an AI assistant a question. The assistant retrieves accurate information from the manufacturer’s digital ecosystem and cites the manufacturer as its authoritative source. The HCP gets exactly what they needed. But never visits the website. Traditional analytics might record:

0 sessions.

The business outcome might be:

successful delivery of authoritative information.

Those are not the same thing. Marketing leaders may eventually need additional measures such as:

AI visibility

How often does authoritative company content appear in relevant AI responses?

Citation visibility

Is the organization being cited as the authoritative source?

Agent retrieval success

Can AI systems find the correct information?

Structured-content coverage

How much high-value content has explicit machine-readable relationships?

Claim traceability

Can important statements be traced to current evidence and approval context?

Content freshness

Are AI systems retrieving the latest approved version?

AI-mediated outcomes

Did the digital infrastructure successfully deliver the intended result even when the person never visited the website?

Agent task completion

Can an authorized AI complete a permitted action successfully?

Escalation quality

Does the system recognize when a human needs to become involved? Over time, the most important digital-performance question may shift from:

“How many people came to the site?”

toward:

“How often did our digital ecosystem deliver a trusted outcome?”

That is not an argument to abandon traffic, engagement or conversion metrics. It is an argument to place them within a broader outcome model. If AI becomes a major discovery and interaction layer, website sessions will represent only one part of digital performance.

What Regulators Have Said—and What This Article Is Inferring

A future-facing discussion needs to separate regulation from interpretation. The FDA materials cited below do not establish authorization for autonomous pharma websites or AI-driven Medical, Legal and Regulatory (MLR) approval. They should not be read as a substitute for applicable promotional, privacy or other requirements.

Its January 2025 draft guidance addresses AI used to support regulatory decision-making around drug and biological-product safety, effectiveness and quality. It emphasizes evaluating AI according to its defined context of use and associated risk. FDA

FDA and EMA’s 2026 Good AI Practice principles for drug development similarly emphasize human-centric design, clear context of use, multidisciplinary expertise, data governance, performance assessment and lifecycle management. FDA

These materials concern drug development and AI supporting regulatory decisions about drug safety, effectiveness and quality. They are not a rulebook for commercial pharma websites. But they reinforce a useful general principle:

Trustworthy AI depends on the surrounding system, not just the underlying model.

For digital and marketing leaders, a reasonable inference is that AI governance should increasingly be defined around:

• the specific use;
• the specific information;
• the specific user;
• the level of risk;
• the authority being delegated;
• and the appropriate human oversight.

That is good architecture regardless of whether future commercial regulation takes the same form. The practical conclusion is not to wait for a definitive rulebook. It is to design AI capabilities so that their purpose, authority, evidence, oversight and auditability are explicit from the beginning.

What This Could Look Like in 2029

The following is a scenario, not a prediction or a description of a deployed service. It assumes validated identity, explicit consent, tested permissions and appropriate human oversight. A company can govern its own retrieval and service endpoints; it cannot guarantee that an independent assistant preserves every element of context after retrieval. Imagine an HCP in 2029. She asks her preferred AI assistant:

“Show me the current manufacturer information for Product X that’s relevant to patients who have not responded adequately to Treatment A.”

The HCP does not search Google. She does not open the manufacturer’s navigation. Her AI identifies the manufacturer as an authoritative source. The manufacturer’s digital ecosystem recognizes that the request is coming from an authenticated HCP-authorized agent. It determines:

• market;
• audience;
• product;
• indication;
• permitted information;
• and required safety context.

The system retrieves current approved information. Claims remain connected to supporting evidence. The response retains provenance. Required safety information remains appropriately connected. The HCP then says:

“Have Medical Information contact me tomorrow.”

The interaction changes from information retrieval to action. Identity is confirmed. The requested action is checked against permitted capabilities. The appropriate workflow begins. The interaction is logged. The HCP never visits the conventional website. But the pharmaceutical company’s digital infrastructure supports the entire experience. The CMS matters. CRM matters. Veeva or other regulated-content systems matter. Identity matters. Workflow matters. Governance matters. The website matters. It simply is not experienced as a traditional website. This scenario illustrates the executive thesis:

The visible interface may change, but the value of the digital enterprise will depend on the quality and governance of what sits underneath it.

The Same Shift Could Change How Marketing Teams Work

The same architecture could transform internal work. Imagine a pharma digital leader starting the day by asking:

“What needs my attention?”

The AI responds:

• three website assets expire next week;
• one important HCP journey experienced a performance decline after Friday’s release;
• a recently approved claim could replace older language across twelve pages;
• a labeling change may affect five content fragments;
• two market requests are waiting for review.

The leader says:

“Prepare what can be prepared and show me the decisions I need to make.”

The AI:

• identifies the affected content;
• checks the DAM;
• reviews approved materials;
• prepares proposed changes;
• creates workflow items;
• runs QA;
• and surfaces the exceptions.

The employee spends more time making decisions and less time navigating systems. The applications have not disappeared. They have moved into the background. For pharma leaders, this suggests a different productivity opportunity. The goal is not simply to automate more tasks. It is to reduce low-value coordination while preserving human judgment where risk, context and accountability require it.

Marketing Expertise May Become More Valuable, Not Less

AI can make execution cheaper. That does not automatically make domain expertise less important. In fact, it may make judgment more valuable. Future marketing leaders may spend less time asking:

• “Which screen contains the report?”
• “Where is that asset stored?”
• “How do I create this workflow?”

And more time asking:

• “What outcome are we actually trying to achieve?”
• “Which source should the AI trust?”
• “Is this appropriate for this audience?”
• “What is the risk?”
• “Which exception matters?”
• “Should this process exist at all?”
• “When should automation stop?”
• “What did we learn from this interaction?”

That is a higher-value use of experienced talent. As interfaces become easier, institutional understanding becomes more valuable. The future role of the marketing leader may therefore become less about operating software and more about defining:

• objectives;
• context;
• acceptable risk;
• decision rights;
• exceptions;
• and outcomes.

AI can coordinate work. It cannot remove the need for accountable judgment.

Seven Priorities for Pharma Leaders

The good news is that companies do not need certainty about 2029 to prepare. Most of the foundational work creates value today.

1. Structure information beyond the page

Do not let critical relationships exist only because people can visually interpret them. Create explicit connections among:

• product;
• indication;
• claim;
• evidence;
• audience;
• geography;
• safety information;
• asset;
• approval state;
• and expiration.

This improves reuse, governance, search and future AI readiness.

2. Map business capabilities, not just applications

Instead of saying: “AEM manages the website.” “Veeva handles review.” “CRM manages customers.” Map what the enterprise can actually do. For example:

• retrieve approved content;
• validate references;
• request review;
• authenticate an HCP;
• identify affected pages;
• arrange Medical Information follow-up;
• publish;
• personalize;
• measure outcomes.

Then determine which systems own those capabilities and what authority is required.

3. Strengthen provenance

Important digital content should increasingly answer:

• Where did this come from?
• Who owns it?
• What supports it?
• Is it approved?
• When was it reviewed?
• Is it still current?
• What depends on it?

That is useful now and essential in a future where AI systems assemble information dynamically.

4. Make approval state explicit

AI systems should not have to infer whether something is:

• draft;
• approved;
• market-specific;
• audience-specific;
• expired;
• superseded;
• withdrawn;
• or pending review.

Make those states machine-readable.

5. Separate permission from capability—and make governance portable

Define clearly which actions AI may:

read, recommend, prepare, submit, execute, or never perform without human approval.

This should become part of AI product design from the beginning. Permissions, provenance, logging and approval rules should increasingly remain intact regardless of whether a capability is accessed from:

• the CMS;
• an internal assistant;
• ChatGPT;
• Claude;
• another enterprise AI;
• or a future HCP agent.

The governance should travel with the capability.

6. Build evaluation before autonomy

Test not only whether AI produces good answers. Test whether it:

• uses authoritative sources;
• preserves required context;
• respects audience boundaries;
• identifies expired content;
• maintains safety information;
• follows permissions;
• escalates correctly;
• and refuses inappropriate actions.

7. Prepare analytics for AI-mediated journeys

Start learning how AI crawlers, AI referrals, citations and machine-mediated interactions appear in your digital ecosystem. Distinguish observed referrals and authenticated completions from sampled answer visibility. Bot visits do not prove human reach, citations or successful outcomes; collect only appropriate, consented data and report measurement gaps. The goal is not to chase every new bot. It is to avoid measuring tomorrow’s customer journey with yesterday’s metrics. These actions are not speculative bets on one vendor or one AI model. They are investments in organizational clarity.

That is why they create value even before the interface changes.

The Contrarian View: Enterprise Platforms May Become More Valuable

It is tempting to conclude that AI makes CMS, CRM and enterprise software less important. The opposite may happen. Their interfaces may become less differentiated. But the value moves underneath. The strongest enterprise platforms will increasingly be the ones that reliably know:

• what is true;
• what is current;
• what something means;
• where it came from;
• who may use it;
• what they may do with it;
• what evidence supports it;
• and when a human must decide.

Those capabilities become more valuable as AI takes on more work. This could be particularly important for pharma. The industry’s regulatory and governance requirements have created enormous stores of:

• approved content;
• claims;
• references;
• labels;
• review history;
• audit trails;
• workflow decisions;
• and controlled processes.

Today, those controls can feel like friction. In an agentic future, they may become a strategic advantage—if they can be converted into machine-usable institutional context. That is the contrarian conclusion:

The companies that have invested most seriously in trusted content and governance may be better positioned for AI than the companies that have simply adopted the most AI tools.

The Pharma AI Futurist View

The important story is not that Adobe added AI to AEM. Or that Optimizely added MCP. Or that Contentful made content accessible to agents. Or that Salesforce wants its capabilities available through different AI interfaces. Those are signals. The larger change is this:

Intelligence is beginning to separate from the application interface.

Today, people open software and tell it exactly what to do. Tomorrow, people may increasingly state the outcome and let AI coordinate the enterprise capabilities required to achieve it. After that, customer agents may begin doing the same. When that happens, the pharma website changes. It stops being only a collection of pages. It becomes one interface into a larger governed system of:

trusted knowledge, evidence, identity, permissions and capabilities.

Humans may access that system through a website. Search engines through structured content. Generative engines through authoritative retrieval. Employees through enterprise AI. And eventually, authorized agents through controlled capabilities. Different interfaces. One governed foundation. That is why the question for pharmaceutical leaders should not simply be:

Which AI should we buy?

Or:

Which CMS should we standardize on?

A more important question is:

If our future customer—or our customer’s AI agent—never opens the interface we spent years optimizing, is the information and capability underneath it structured well enough to be understood, trusted and safely used anyway?

The future pharma website may not belong to the company with the best chatbot. It may belong to the company that has done the hard work of making its organization understandable to intelligent machines. And that work starts long before the interface disappears.

Frequently Asked Questions

What is the future of pharmaceutical websites with AI?

Pharma websites are likely to remain important human experiences, but AI may increasingly mediate how people discover and use pharmaceutical information. Over time, websites could become one interface over a broader governed system of structured content, evidence, identity, permissions and services.

Will AI replace pharma websites?

Probably not. A more plausible future is that websites continue serving humans while AI systems increasingly retrieve authoritative information and, where appropriate, invoke permitted services on users’ behalf.

Will AI replace AEM, Sitecore, Optimizely or other CMS platforms?

There is currently little evidence that enterprise CMS platforms will simply disappear. Their visible interfaces may become less central while their content models, APIs, workflows, permissions and governance become more important.

Which CMS platforms could support this future?

The concept is platform-independent. Adobe Experience Manager, Sitecore, Optimizely, Drupal/Acquia, WordPress VIP, Contentful, Contentstack and other enterprise or composable CMS platforms could participate if they can expose sufficiently structured and governed content.

What is an agentic pharma website?

An agentic pharma website is a future architectural concept in which pharmaceutical digital infrastructure serves both humans and authorized AI agents. Agent access would need to be constrained by identity, purpose, permissions, content status, risk and applicable regulatory requirements.

What is GEO for pharmaceutical websites?

Generative Engine Optimization focuses on making authoritative content easier for AI systems to understand and accurately attribute. For pharma, this should emphasize structured content, clear entities, provenance, current information, claim-to-evidence relationships and authoritative sources.

Does FDA permit autonomous AI to approve promotional content?

The FDA materials cited here do not establish such permission. Current FDA AI guidance discussed in this article focuses on drug-development and regulatory-decision-making contexts rather than autonomous commercial MLR approval.

What should pharma leaders do now?

Prioritize structured content, provenance, machine-readable approval status, clear content relationships, identity and permissions, AI evaluations, risk-based action boundaries, analytics for AI-mediated journeys and explicit human escalation.

Related Reading

Your Next Pharma Website Visitor May Be an AI Agent explores external customer journeys. The Hidden Workforce Behind Pharma AI explains how domain expertise becomes reusable AI infrastructure.

Sources and Further Reading

Sources checked September 22, 2026. Vendor announcements describe their own capabilities; the architecture, seven priorities and future scenarios are the author’s analysis.

• Adobe Experience League: AEM MCP Server
• Optimizely: 2026 Optimizely CMS 13 release notes
• Contentful: MCP use cases
• Acquia: The CMS That Feeds AI
• Salesforce: Salesforce unveils AIforce
• Pew Research Center: Google users and AI summaries
• Google Search Central: AI features and your website
• FDA: AI regulatory decision-making draft guidance
• FDA: Good AI Practice principles for drug development