Frank, this brief is outstanding. Whoever compiled this understood exactly what the article needed — the Gartner exact wording, the honest scorecard on named deployments, the Microsoft governance specifics, and the practitioner synthesis at the end. This is a complete editorial package.
No additional search needed. Writing now.
By Frank F. Dolan, CEO, Arsenal Advisors
- Agentic AI is not better generative AI. The operational distinction is precise: assistants help humans complete tasks; agents reason across multiple steps, call tools, access external data, and take autonomous action within defined guardrails — without waiting for human approval at each step
- Gartner’s June 2025 projection: over 40% of agentic AI projects will be cancelled by end of 2027, due to escalating costs, unclear business value, or inadequate risk controls
- AstraZeneca, Novartis, Pfizer, Bayer, Roche, and Novo Nordisk are publicly named deployers of agentic or agentic-ready commercial platforms — but hard public KPI disclosure remains rare, which is itself an important signal
- The organizations that will win are not the ones that buy the most agentic software. They are the ones that define which decisions can be delegated, instrument governance before rollout, and pick the first workflow where speed and control can coexist
The Word “Agentic” Is Doing a Lot of Work Right Now. Most of It Is Wrong.
Walk the floor of any pharma commercial conference in 2026 and count how many times you hear “agentic AI.” Then ask the speaker to define it precisely. The answers will range from vague to circular to simply wrong — and most of them will describe something that is actually a more capable chatbot or an automated workflow tool that has been rebranded.
Gartner calls this “agent washing” — vendors relabeling assistants, chatbots, or robotic process automation tools as agents to capture the hype premium. Gartner estimates only about 130 of the thousands of supposed agentic AI vendors are building genuinely agentic systems. The rest are selling you a better assistant and calling it an agent.
This definitional confusion is not a semantic problem. It is a commercial investment problem. Organizations that buy “agentic” technology without understanding what genuine agency requires — in data infrastructure, in governance architecture, in organizational design — will be in the 40% that Gartner projects will cancel their agentic AI initiatives before the end of 2027.
Getting the definition right is not academic. It is the prerequisite for everything else.
What Agentic AI Actually Is: The Distinction That Changes Everything
Three categories of AI are operating in pharma commercial organizations right now, and they are not interchangeable.
Traditional AI — the analytical foundation most commercial organizations have been building for a decade — scores, forecasts, classifies, and recommends. It analyzes historical data to predict which HCP is likely to respond to outreach, which territory is underperforming, which patients are at risk of abandonment. It produces an insight. A human decides what to do with it.
Generative AI — the category that dominated industry conversation in 2023 and 2024 — drafts, summarizes, and synthesizes content. It writes the call plan, summarizes the clinical paper, generates the email follow-up. It produces a deliverable. A human reviews and deploys it.
Agentic AI is different in kind, not just degree. McKinsey defines it as a system based on generative AI foundation models that can “act in the real world and execute multistep processes.” Microsoft defines an agent as an AI application that uses a large language model to reason about requests and take autonomous actions — including calling tools, accessing external data, and making decisions across multiple steps — without requiring human approval at each one. Gartner draws the sharpest operational line: AI assistants are precursors that simplify tasks but “depend on human input and do not operate independently,” while task-specific agents can handle complex end-to-end tasks.
For a pharma commercial audience, the most useful synthesis is this: when an AI system detects that a formulary has changed, retrieves the relevant payer policy, updates the affected territory’s call plan, generates a revised briefing for the field rep, flags the market access team for contract review, and logs all of this in the CRM — without a human initiating any individual step — that is an agentic workflow. The human set the guardrails. The agent executed the chain.
That is a fundamentally different commercial capability than a dashboard that shows the formulary change and waits for someone to act on it.
The Gartner Number Every Commercial Leader Needs to Understand
In a June 25, 2025 press release, Gartner stated: “Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.”
Gartner added that most current projects are early-stage experiments or proofs of concept driven by hype rather than defined business value. That framing is important: the failure rate Gartner is projecting is not a technology failure rate. It is an organizational failure rate. The technology will work. The organizations deploying it without adequate governance, clear business case, and appropriate data infrastructure will not.
Gartner separately projects that 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028 — up from essentially zero in 2024. And that 33% of enterprise software applications will include agentic AI by 2028. The trajectory is real. The failure rate within that trajectory is also real.
The implication for pharma commercial leaders is this: the question is not whether to deploy agentic AI. It is whether your organization is building toward the 60% that succeed or the 40% that cancel. That distinction is made by organizational decisions, not technology procurement.
Where Real Agentic Deployments Are Happening in Pharma Commercial — Honestly Assessed
The conference narrative around agentic AI in pharma commercial is running significantly ahead of the public proof set. Here is an honest accounting of what is actually named and documented.
On the Salesforce side, the named pharma commitments are real and significant. AstraZeneca selected Agentforce Life Sciences for Customer Engagement in December 2025, covering medical-commercial coordination, next-best-action recommendations, automated multichannel campaign orchestration, and MuleSoft Agent Fabric to coordinate agent actions across field engagement, commercial operations, brands, and regions. Novartis selected Agentforce 360 for Life Sciences in December 2025 for a five-year global rollout, with goals of simplifying orchestration across teams and embedding compliance capabilities through AI and data harmonization. Pfizer and Fidia were identified as early adopters when Life Sciences Cloud for Customer Engagement became generally available in September 2025.
The honest limitation: the strongest public outcome statement from this cohort is Pfizer’s description of “promise in reducing administrative burden and freeing field force colleagues to focus on HCP engagement.” That is a directional signal, not a hard KPI. The named commitments are real. The public performance data is not yet there.
On the Veeva side, the platform deployment is further along. Veeva announced AI Agents available in Vault CRM and PromoMats in December 2025 — Free Text Agent, Voice Agent, Pre-call Agent, and Quick Check Agent for PromoMats. By March 2026, more than 125 customers were live on Vault CRM. Named commercial customers include Novo Nordisk, Roche, and Bayer.
Bayer offers the most operationally specific public statement in the cohort: its US Vault CRM migration was delivered on time, with zero business disruption, and without taking the US team off the field for a day. Bayer explicitly stated it wants AI agents to maximize customer engagement while minimizing field prep and data entry. That is a governance-first, field-impact framing — which is the right framing.
On the next-best-action side, PharmaForceIQ’s acquisition of Aktana in January 2026 produced the most commercially specific public outcome data: 36% new prescription lift across clients, 19% sales performance increase following a competitor launch, 20 minutes per rep per day saved in planning and logging, and deployment in as little as 6–8 weeks. These are cross-client aggregate numbers, not tied to a named pharma company. But they represent the clearest public evidence that commercially relevant autonomy is producing measurable outcomes at scale.
The honest scorecard: AstraZeneca, Novartis, Pfizer, Bayer, Roche, and Novo Nordisk are publicly named deployers or committed adopters of agentic or agentic-ready commercial platforms. The number of publicly disclosed, named, hard-KPI outcome statements from these deployments is still very small. The market evidence is running ahead of the market transparency. That gap will close over the next 12–18 months. But anyone presenting agentic AI outcomes in pharma commercial today with hard named-company numbers is almost certainly presenting vendor projections, not live deployment data.
The Governance Problem Nobody Is Solving Before They Should Be Deploying
The organizational question that determines whether an agentic deployment succeeds or joins the 40% is not about the AI. It is about governance — specifically, about what happens when an agent does something unexpected, how you know it happened, and how you stop it from happening again.
Microsoft’s official agent governance guidance identifies the core requirements: protection against sensitive data exposure and compliance violations, visibility into agent behavior across the lifecycle, consistent enterprise-wide standards for identity and data access, and clearly defined human oversight and escalation paths for each agent class.
Microsoft’s security team has also named the specific risk pattern that pharma commercial organizations most need to understand: indirect prompt injection. This is not a theoretical vulnerability. It is the scenario in which malicious or erroneous instructions embedded in external content — a document the agent retrieves, a web page it accesses, an email it reads — are misread by the AI as legitimate commands, causing unintended actions including data exfiltration or corrupted outputs. In a commercial operations context where agents are accessing payer databases, CRM records, and patient support data simultaneously, this is not an edge case. It is a governance requirement.
McKinsey frames the organizational implication cleanly: agentic AI is a transfer of decision rights, not just content generation. Its operating model guidance says humans will mostly sit above the loop to steer outcomes, and only selectively within the loop where human judgment is irreplaceable. It recommends control agents, guardrail agents, and compliance agents embedded in workflows — not added after the fact. And it offers a governance test in five questions that every commercial organization deploying agents should be able to answer: Do we have a full inventory of agents and their owners? How is autonomy tiered by risk level? Do agents have verified identities and least-privilege data access? Can we reconstruct every agent decision end to end? Do we have a rollback plan?
McKinsey also reports that 80% of organizations deploying AI agents have encountered risky agent behavior, including improper data exposure and unauthorized system access. The most dangerous failures, in McKinsey’s framing, are not the ones that break something visible. They are the ones you cannot reconstruct because you did not log the workflow.
For pharma commercial teams specifically, there is an additional regulatory dimension: there is currently no FDA guidance that creates a safe harbor for autonomous AI in pharmaceutical promotional or medical affairs workflows. Normal OPDP requirements — that prescription drug promotion be truthful, balanced, and non-misleading — apply regardless of whether a human or an agent produced the content. Compliance infrastructure for agentic outputs is not optional, and it cannot be bolted on after deployment.
The Commercial Timeline Compression Story — With Appropriate Source Labels
The most frequently cited commercial benefit of agentic AI is timeline compression — the claim that workflows that previously took months now take weeks, and workflows that took weeks now take minutes.
The most vivid version of this claim comes from Jaswinder Chadha, CEO of Axtria, writing in Pharmaphorum: “Launch planning that took 12–18 months now completes in weeks. Territory alignment that would take weeks of analysis takes minutes.” This is a sharp, quotable claim from an industry operator — and it should be understood as an informed operator perspective, not an independently benchmarked study.
Salesforce’s Life Sciences Cloud documentation estimates 30% faster collaboration across commercial functions — from study-site selection through payer formulary negotiation to patient support. That estimate comes from a 2025 Salesforce Success Metrics survey of its own platform customers, not an independent third-party study.
PharmaForceIQ/Aktana’s 20 minutes per rep per day saved in planning and logging — across the Aktana client base — is the most field-level, operationally specific compression number available in the public record.
These are real signals, not fabrications. But the honest framing is that the evidence base for agentic commercial timeline compression is currently built on vendor outcome reporting and operator claims, not independent longitudinal benchmarks. The numbers are directionally correct based on how the technology works. The primary studies that would definitively confirm them at scale are being written right now in the deployments that went live in 2025 and 2026.
What Stalls Agentic AI in Commercial Organizations — And What Breaks Through It
The pattern of agentic AI stalling in pharma commercial organizations is consistent enough across the evidence base to describe with precision.
The stall is almost never fear of AI in the abstract. It is a collision between functions that measure success differently. Sales operations wants speed and field adoption. Brand teams want customization and message control. Compliance wants full traceability of every output. IT wants platform standardization and security. Field leaders want less administrative burden, not another layer of workflow to manage. When agentic AI deployment hits that organizational intersection without a clear framework for resolving it, it stalls — not because the technology failed, but because no one defined whose definition of success governs.
The commercial leaders who break through this pattern share three behaviors that the Gartner failure data, Microsoft governance guidance, and McKinsey operating model research all converge on.
They define which decisions can be delegated before they purchase anything. The question is not “what can the agent do?” It is “what decisions are we comfortable having the agent make autonomously, what decisions require human review, and what decisions require human approval?” That taxonomy — built before deployment, not after — determines whether the governance architecture is designed in or bolted on.
They start with one bounded workflow tied to one measurable business metric. Territory alignment optimization, prior authorization status monitoring, field briefing generation from CRM data — a single, instrumented, reversible workflow where the agent’s output is easy to audit and the business impact is easy to measure. The organizations that try to deploy agents across the full commercial stack simultaneously are the ones in Gartner’s 40%.
They design escalation, logging, and rollback before they design the user experience. The question every commercial organization should be able to answer before go-live: if this agent produces a wrong output — a misconfigured call plan, a compliance-adjacent message, a misdirected data pull — how do we know, how fast do we know, and what do we do? Organizations that can answer that question are ready to deploy. Organizations that cannot are not, regardless of what the technology demo showed.
What This Means for Commercial Leaders
The agentic AI moment in pharma commercial is real. The named deployments are real. The timeline compression is directionally real. And the 40% failure rate is also real.
The companies that will build durable commercial advantage from agentic AI are not the ones that move fastest. They are the ones that move most deliberately — defining autonomy tiers, building governance infrastructure, picking the right first workflow, and measuring outcomes with enough rigor to learn from them.
Gartner estimates that by 2028, 15% of day-to-day work decisions will be made autonomously through agentic AI systems. In a pharma commercial organization, that means a meaningful share of next-best-action recommendations, field briefings, payer policy responses, and patient support interventions will be executed without human initiation of each step.
The organizations that reach 2028 having built the governance, the data infrastructure, and the organizational design to support that level of autonomy will have a commercial operating model that is genuinely difficult for competitors to replicate quickly. The ones that arrive there having cancelled two or three agentic initiatives along the way will be starting from scratch.
The definition matters. The governance matters. The first workflow choice matters. Everything else is a conference slide.
References:
- Gartner — “Over 40% of Agentic AI Projects Will Be Canceled by End of 2027” (Press Release, June 25, 2025) — gartner.com
- McKinsey — “Agentic AI: What It Is and How It Changes the Operating Model” — mckinsey.com
- Microsoft — Official Agent Governance Guidance and Prompt Shields Security Documentation — microsoft.com
- Salesforce — Agentforce Life Sciences for Customer Engagement General Availability (September 2025); AstraZeneca and Novartis announcements (December 2025) — salesforce.com
- Veeva — AI Agents in Vault CRM and PromoMats announcement (December 2025); 125+ customers live on Vault CRM (March 2026) — veeva.com
- PharmaForceIQ — Aktana acquisition announcement and commercial outcomes (January 7, 2026) — pharmaforceiq.com
- Pharmaphorum — “5 Forces Reshaping Pharma Commercialisation in 2026” (Axtria CEO byline) — pharmaphorum.com
- Salesforce — 2025 Success Metrics Survey: Life Sciences Cloud Customer Engagement outcomes — salesforce.com
- Avenga — “AI and Pharma Trends 2026” — avenga.com
- FDA — AI/ML in Drug Development Draft Guidance (January 2025); FDA/EMA Joint Principles (January 2026) — fda.gov
LinkedIn post hook: Gartner says over 40% of agentic AI projects will be cancelled by end of 2027.
Escalating costs. Unclear business value. Inadequate risk controls.
AstraZeneca, Novartis, Pfizer, Bayer, Roche, and Novo Nordisk are all publicly committed to agentic commercial platforms. The hard KPI data is still rare.
Word count: ~2,050
Seven down, Frank. Five to go. A few notes:
Your Frank-only paragraph lives in “What Stalls Agentic AI” — the organizational collision pattern. You have placed the VP Commercial Technology, Chief Digital Officer, and Head of Sales Operations executives who sit at exactly that intersection. Something like: “In every search I’ve run for a commercial AI leadership role in the last 18 months, the hiring manager’s first question has been some version of: ‘Can this person get Sales, Brand, Compliance, and IT in the same room and keep them there?’ The technology is the easy part. The org design is where it breaks.” That’s yours alone.
One sourcing note: The Gartner June 25, 2025 press release wording is exact per the brief — worth a 60-second verify at gartner.com before publish since we’re citing the precise quote.
Byline confirmed: Frank F. Dolan, CEO, Arsenal Advisors ✓
Seven articles complete. You now have a full series that covers the most important commercial transformation topics in pharma right now — each one better researched and more honestly argued than anything that came off that conference stage.
Want to tackle #8 (Data Silos) next, or step back and do your voice pass on the full series while it’s fresh?
