The pharma IT playbook is being rewritten: what the rise of agentic AI means for life sciences. Veeva's 2026 R&D and Quality Summit in Copenhagen introduced three new AI product lines alongside roadmap updates across clinical, regulatory, safety, and quality. This post covers what each includes, where each stands on the release timeline, and what organizations with Veeva need to address before August. What "agentic" Means. Unlike conventional enterprise software, agentic AI does not require user initiation at each step. An agent monitors an environment, detects defined trigger points, and executes a workflow without manual handoff. The human role shifts from task performance to output review. This distinction is most notable in functions where teams spend significant time on document processing, case intake, or correspondence management. Veeva's announcement. Veeva introduced three AI product lines at Copenhagen, each operating at a different layer. Veeva AI is already in production, with case narrative generation live in Safety and deviation and complaint summaries live in Quality. In August 2026, the 26R2 release brings Veeva AI to all customers, including a conversational interface that routes queries across Vault data within each user's existing security permissions. Veeva Falcon is a separate platform that sits outside of Vault and automates what Veeva calls "agentic labor": end-to-end workflows that currently require human initiation at every step. The first three agents cover TMF document intake and quality checking, safety case triage and intake, and health authority correspondence in regulatory. Falcon is targeting early adopter availability in November 2026. Agentic Authoring rounds out the roadmap at the furthest horizon, expected in late 2027. It will monitor incoming data and proactively draft submissible documents, integrating with Vault RIM and Microsoft Word. Prerequisites for effective AI deployment. Agents operate on what's in your system. An agent checking TMF documents can only flag what's visible to it; one drafting HA responses draws entirely from your existing submission content, document tagging, and communication records. Data quality and governance structure directly determine the accuracy and reliability of agent outputs. In a GxP environment, human review of AI-generated outputs is a regulatory expectation. That requirement shapes how organizations need to structure workflows before enabling any AI capability. What to do before August. Veeva AI reaches all customers with the August 26R2 release. These are the steps worth taking now. 1. Pick the starting point. a. What to do. List the routine, repeatable tasks your team performs and attach rough frequency, along with time estimates to each. In most organizations, outputs like case narratives, CAPA responses, and periodic reports receive the most planning attention, but document completeness checking, record updates, and metadata tagging often account for more aggregate hours across the team. Before finalizing a shortlist, spend time with the people who actually perform these tasks: they will surface edge cases and data quality issues that don't appear in reports, and identify which processes carry more nuance than the documentation suggests. b. How to prioritize. Evaluate each task on four criteria: volume, time per instance, degree of judgment required, and current data readiness. The strongest first candidates score high on the first two and low on the last two. TMF document quality checking is a consistent example: the rules are explicit, volume is high, and current coverage is typically manual and inconsistent. Tasks like drafting HA correspondence responses score high on judgment and carry significant consequences for error; those are better suited as candidates after lower-risk workflows have been established. c. What this achieves. A first use case that produces measurable outcomes and provides a replicable model for subsequent use cases. 2. Document the baseline. Record current performance data for every process you plan to automate before the August release. Relevant metrics include average processing time per unit, FTE hours per week or cycle, error and rework rates, and average cycle time from task initiation to completion. Pull what you can from Veeva reports and supplement with direct input from team leads. Capture this data before the release. Once AI-assisted workflows are running, the pre-automation baseline is no longer available and any impact comparison loses its quantitative basis. The baseline data often also reveals that high-volume tasks are distributed differently than initially assumed, which can shift prioritization before resources are committed in a specific direction. A documented foundation for evaluating AI performance and substantiating continued investment, using your organization's own data rather than vendor benchmarks. 3. Audit data and metadata quality. Pull a report on your highest-volume document and object types in Vault and assess three things: classification consistency, required field completion rates, and accuracy of document status values. Most organizations find a meaningful percentage of records with missing metadata, inconsistent naming conventions inherited from early configuration decisions, or status fields that no longer reflect reality. Begin with the document types associated with your first planned automation target. If you're planning to use Veeva AI for TMF document processing, audit TMF metadata first. If the focus is regulatory, start with submission content plans and registration records. Agents cannot infer missing data, and they won't surface what's miscategorized. Agent outputs that accurately reflect the state of your Vault environment from the point of deployment, rather than outputs that require manual investigation before they can be acted on. 4. Review permission structures. Veeva AI only surfaces information that a user can already access, making the permission model more consequential in an AI-assisted environment than in a human-only workflow. If a regulatory user can't see certain submission documents, the AI Tab won't surface them in response to queries. If your permission sets were configured during initial implementation and haven't been revisited since, they likely reflect an outdated organizational structure. Run a review against current roles and responsibilities. Identify users whose access is too narrow for the workflows the AI will support, and tighten access wherever it is broader than it should be. Pay particular attention to any system-level roles that agent processes may inherit. A permission structure aligned with current organizational roles and designed to support AI-assisted workflows. 5. Update SOPs and assign review ownership. In a GxP environment, human review of AI-generated outputs is a regulatory expectation. Every AI output your organization acts on requires a defined review process, a named accountable role, and an audit trail documenting that review occurred. Map each workflow you plan to enable against three questions: which role reviews this output; what criteria is the output reviewed against; and how is that review documented. For case narrative generation in Safety, this means updating the case processing SOP to reflect an AI-assisted step. For deviation summaries in Quality, it means defining the QA reviewer's checklist. For regulatory, it means establishing review and approval processes for HA correspondence drafts. If your organization already has enterprise-level AI governance policies, verify whether they are specific enough to cover Veeva-based agent outputs. General policies typically do not address the process-level detail that GxP review requires. Update SOPs for the specific workflows you plan to enable before the August release. A compliant deployment path for Veeva AI, with review infrastructure in place before the capability goes live. Closing. Daelight Solutions works with life sciences organizations at exactly this stage. Daelight Solutions conduct Vault configuration and data readiness assessments to identify gaps before they affect AI performance, review permission structures and governance frameworks for AI output review, and offer dedicated health checks for Vault RIM and eTMF environments. If you want to understand where your organization stands ahead of the August release, reach out to schedule a Vault readiness assessment.

Salesforce lands 140 life sciences clients, pitches 'headless' AI to pharma. A patient support program engagement run as a remote call, with meeting controls for audio, screen share and recording alongside the approved deck. (Image: Salesforce) As Peter Gassner, CEO of Veeva Systems, the influential cloud software provider for the life sciences industry, took the stage at the Veeva Commercial Summit in Boston, Salesforce announced that more than 140 life sciences organizations are using its competing Agentforce Life Sciences platform. In the mix as Salesforce clients are several marquee industry names like Novartis, AstraZeneca, Moderna and Merck Animal Health. "The 140 represents customers across our entire portfolio," said Joe Ferraro, SVP and GM of life sciences at Salesforce, "regardless of whether they are using it for patient engagement, clinical engagement, commercial engagement or medical engagement." Salesforce is positioning Agentforce Life Sciences as a broad, AI-native platform across clinical, medical, patient and commercial engagement. So is Veeva. As Gassner put it at the Commercial Summit: "The Agentic Commercial model changes everything about how we approach commercialization." For example, he pointed to everything "from digital and non-personal promotion," how field teams operate as well as content approval and creation. 'Everybody is using agents' When asked about how Salesforce's agentic AI strategy compares to Veeva's, Ferraro said, "Everybody is using agents now. Salesforce has been pioneering AI for the last decade, whereas many legacy systems have been slow to innovate. I would say a combination of our product vision two years ago, when we started this process of [building Agentforce Life Sciences], and customer pressure really forced Veeva's hand on this." Ultimately, Ferraro differentiates Agentforce Life Sciences from Veeva's offerings in terms of experience. "We are talking apples and oranges," he said. "We have been in the AI game at Salesforce since the predictive and machine-learning days." Salesforce CEO Marc Benioff may have ramped up Salesforce's AI focus in the wake of the ChatGPT launch, but the history goes back further. At an internal all-hands in 2014, he declared Salesforce an "AI-first company", the same year it acquired the machine-learning startup RelateIQ for $390 million, and Salesforce launched Einstein, its first CRM AI, at Dreamforce in 2016. In the years since, Benioff has made autonomous agents the company's central pitch, launching the Agentforce platform at Dreamforce in September 2024, branding it "digital labor," and setting a goal to empower one billion agents by the end of 2025. That focus has not died down. On the company's first-quarter fiscal 2027 earnings call on May 27, Benioff told investors Salesforce had processed 28.6 trillion AI tokens, up 152% quarter over quarter, and called agentic AI "the biggest growth opportunity" for the company since it brought CRM to the cloud, with Agentforce annual recurring revenue at $1.2 billion. "We have thousands of engineers working on AI at Salesforce right now," Ferraro said. "There is nowhere near that scale for Veeva, so we are able to move a lot faster from an agentic standpoint." For its own part, Veeva announced Falcon on May 27, an agentic platform and set of standard agents for drug development that run in Veeva Development Cloud across its clinical, regulatory and safety applications, with early-adopter availability set for November 2026. Gassner called it the company's "first offering in agentic labor." Vault CRM's Free Text, Voice, and Pre-call agents went live in December 2025. Veeva's public roadmap listed Safety and Quality agents for April 2026, and its 26R1 release introduced a Narrative Agent for Veeva Safety. Veeva has been buying as well as building. In March 2026 it acquired Ostro, an AI brand-engagement platform for life sciences, for about $100 million. The acquisition adds conversational AI that answers patient and HCP questions from Medical, Legal, and Regulatory (MLR)-approved content rather than generating novel responses. At its Commercial Summit event, Gassner previewed a Vault AI Tab coming in August, coming in August, a conversational layer that answers questions against a customer's Vault data within existing security permissions. The Agentforce Life Sciences home screen, where a user queries the assistant with prompts such as "HCPs to meet with next" or "Show me declining prescribers." (Image: Salesforce) The bet on 'headless' The larger agentic strategy is what Salesforce calls "headless." Ferraro described it as "popping the top off the platform, exposing all the APIs, and exposing all the data models and workflows as MCP tools," referring to the Model Context Protocol standard that Anthropic introduced in 2024. MCP has since been donated to the Agentic AI Foundation, a Linux Foundation directed fund, in December 2025. Rather than requiring users to work inside Salesforce's own interfaces, the company exposes its productized workflows as MCP servers that surface inside whatever AI tool a customer already uses. "I was with a customer yesterday who has made a significant investment in Anthropic, and they are going to use our MCP service for Agentforce Life Sciences to bring all of those compliant workflows directly inside Claude," Ferraro said. Meanwhile, Veeva's Vault AI Agents are built to run inside Veeva's own applications, working in-context against its data, documents, and workflows, a depth-inside-the-app posture rather than a surface-it-anywhere one. Publicly visible examples of connectors that expose Veeva data to outside models today include third-party offerings such as CData. The two also split on the model layer. Salesforce supports a range of Salesforce-managed and customer-supplied models for Agentforce. Veeva fixes the model layer for its standard agents, Vault CRM's included, running them on Anthropic and Amazon models hosted on Amazon Bedrock, and opens model choice only for custom agents, where customers can use Veeva-hosted models or bring their own on Bedrock or Microsoft Azure AI Foundry. Agents help speed onboarding. Salesforce has noted that some of its smaller life science customers have completed onboarding in about a month. Meanwhile, companies like Chiesi and CSL are using Salesforce to unify customer data and coordinate engagement across commercial and medical operations. In terms of the rapid onboarding, agents deserve part of the credit, Ferraro said, reducing the need for large teams of consultants to oversee the process. "You get one forward-deployed engineer, and they can do the work of five or 10 previously, because they have Claude Code, or they have Codex, and they can point those tools at the Salesforce platform," he said. For instance, Salesforce built an MCP server inside Agentforce Life Sciences that configures the product itself. A request like adding a quick action to the iPad app that launches a prompt in the UI, a task a human would handle by navigating through nested configuration screens, gets executed by the agent automatically. A rep's in-visit view in Salesforce's life sciences customer-engagement app with sample content. (Image: Salesforce) Large models are probabilistic by design, trained to predict likely next tokens from context. That can help summarize a long PDF in seconds, but life sciences workflows demand predictability: how often a rep can message a physician, what the message says, and whether consent exists are all regulated. For that reason determinism is the core theme hanging over agentic software in regulated industries. On compliance, Salesforce and Veeva land in a similar place. Salesforce uses a feature it calls Agent Script to weave deterministic checks, approved templates and consent status into the probabilistic layer. Meanwhile, Veeva runs its standard agents in-context inside its own applications against its data and workflows. As a result, the model can synthesize and surface, but the regulated action runs through a fixed set of tools backed by deterministic workflows and confined actions. "[The model] cannot go off script, because those tools are unavailable," Ferraro said. "The tools are outside its toolbox."