Lead Author: Zachary Golden, Senior Manager & NextGen Therapies AI & Innovation Lead, Deloitte Consulting LLP

Supporting Authors: 

• Halle Stanley, Manager & NextGen Therapies AI Delivery Transformation Lead, Deloitte Consulting LLP
• Rajesh Singh, Managing Director & NextGen Therapies Technology Leader, Deloitte Consulting LLP
• Rohit Sethi, Managing Director & NextGen Therapies Digital Value Chain Leader, Deloitte Consulting LLP
   

Introduction

Advanced Therapy (AT) manufacturers are absorbing pipeline growth through headcount, and that model has a ceiling. More than 1,000 Advanced Therapies are in development today,[1] and the number of U.S. patients receiving them is projected to climb from roughly 12,000 in 2020 to more than 340,000 by 2030.[2] Independent forecasts put cumulative U.S. product-indication approvals at roughly 60 through 2030.[1]

More therapies, more geographies, more treatment sites, more patients, and the same overloaded functional teams are expected to absorb each additional layer of complexity. These are also among the most expensive therapies ever commercialized, with single-treatment prices among the highest in medicine. The margins these programs operate under leave no room to fund the gap with people. Hiring has not kept pace, and at the economics of these programs it never made sense to try.

What makes this urgent is not the volume, but the nature of the work. An Advanced Therapy Medicinal Product (ATMP) is often not a unit on a shelf. In autologous programs the product is living, tied to one identified patient, and bound by a vein-to-vein clock with no buffer stock and no second batch. Every order requires a bespoke orchestration across apheresis, manufacturing, release testing, cryogenic logistics, treatment site, and treating physician, each step held to chain of identity and audited under GxP. Complexity does not scale with volume the way it does in traditional pharma. It scales with patients, programs, geographies, and the exceptions that surface at every node. The work volume and complexity compound and headcount oftentimes will grow linearly. That is the gap, and it is why AT manufacturers hit a bandwidth ceiling faster and harder than traditional biopharma. Practitioners who can hold that orchestration together are expensive, take years to develop, and do not exist in the numbers the pipeline now demands.
 

The Way Humans Interact With Enterprise Systems Is Fundamentally Changing

For thirty years, enterprise software has operated on a single assumption: humans interact directly with systems of record to get work done. A supply chain planner logs into a bespoke Slot Management system to check slot availability. A commercial operations lead opens Order Orchestration to review an order status tied to a specific patient. A quality analyst pulls a single-patient lot deviation report from QMS. Systems holds the data and the human is the interface to it. Those two things have always been bundled together.

That bundle is coming apart with the introduction of digital staff into the digital ecosystem.

The industry calls them AI agents, but we call them digital staff, because how you name them determines how you manage them. For now: digital staff take on system-level work that today runs through a human’s hands. They read from those systems of record, synthesize signals across them, act within defined boundaries, and surface decisions to humans for review or approval. Judgment stays with the human while digital staff take the system work off their hands and the underlying system foundation stays intact.

This is not a sudden switch. It is a directional shift that will play out over the coming years. In the near term, humans still interact directly with systems for many tasks. Over time, the proportion of human-to-system interaction decreases, and the proportion of human-to-digital-staff interaction increases.

The platforms that run AT commercial and supply chain operations are already rebuilding themselves for this. The clearest evidence came in April 2026, when Salesforce launched Headless 360, a rebuild that makes every function of its platform directly callable by software agents through APIs, protocol-based tools, and command-line commands, with no human navigating a screen to get the work done. Its rationale was blunt: "We made a decision two and a half years ago: Rebuild Salesforce for agents."[3] And Salesforce is not alone. One month later, ServiceNow opened its own system of action to any AI agent through a generally available MCP server, exposing the workflows and approval chains it had built over two decades so agents could execute governed work without touching the interface.[4] For the manufacturers who run on these platforms, their own vendors are clearing the path, and the ones who move first will set the pace for everyone else.
 

A New Layer Is Emerging Between Humans and Systems

The operating model of the future has three components and one binding layer.

1. Systems of Record stay unchanged. The Order Orchestration, Slot Management, ERP, CRM, QMS, and other commercial and supply chain platforms manufacturers have invested in will remain the authoritative source of truth. That is where the data lives and transactions get written.
2. Human Decision-makers move up, not out. They remain the judgment owners, relationship holders, and accountability anchors; in a such a regulated industry that last role is one an agent will never be able to hold. What changes is where their time goes: out of the systems they used to navigate and into directing the digital staff that navigate for them.
3. The Control Plane is what is new. If digital staff are the workers, the control plane is the management layer that directs them. It is where humans set what each digital staff member is allowed to do, assign their work, and oversee the results, sitting between the people giving direction and the systems where the work lands. At the outset it is a distributed set of control planes, each scoped to a function or cluster of digital staff: individual supply chain, commercial operations, and MSAT control plane(s). These will not be fully integrated at first, but over time, as governance matures, those distributed planes converge toward a unified enterprise-wide control plane.
4. The Integrated Data Fabric binds all three together. Without it, manufacturers end up with capable digital staff operating in silos: each control plane integrated differently, governance applied inconsistently, audit trails that stop at functional boundaries. The collective intelligence layer is the underlying mesh of data flows, integration standards, identity and trust protocols, guardrails, and governance infrastructure that allows humans, the control plane, digital staff, and systems of record to interoperate as a coherent system. In AT it also must hold chain of identity and a continuous audit trail across every system, because a break is a compliance and patient-safety event, not just a data-quality issue. Someone needs to own it, and the decisions made about it now are among the hardest to reverse later.

Calling Them Agents Understates What They Are and What They Require

Classical enterprise software is deterministic. The same input produces the same output, every time, because its behavior is fully specified at build time. Digital staff are not built this way. They reason over context, weigh options, and decide when a situation calls for escalation rather than action, probabilistic in the same sense a human judgment call is probabilistic: not unpredictable, but exercising judgment within guardrails rather than fixed logic.

A system which reasons cannot be managed the way a system that executes rules is managed. It must be directed, not configured. Its outputs must be interrogated, not checked. Its failure mode is not a bad output under known conditions, it is a misjudgment under a novel one, and in AT that misjudgment is not a support ticket, it is a batch that cannot be remade and a patient who cannot wait for you to notice.

That is what makes where digital staff sit on a given task the highest-stakes design decision in the operating model, not just an implementation detail. Every role has an automation line, a boundary between work that requires irreducible human judgment and work that does not, and where that line sits determines how digital staff and humans actually interact. That interaction is not a single mode. It plays out across a workflow much like the working relationship between a manager and a capable junior team member.

When a digital staff member identifies a problem, it does not just flag it. It brings a proposed solution, already assembled with context and a recommended path, because a purely alerting capability adds little value over deterministic code. The human reviews the recommendation and, if they agree, approves it. If they disagree, they interrogate the digital staff member directly: challenge the assumptions, redirect the approach, provide context the agent did not have. The digital staff member incorporates that direction and revises. Once the human is satisfied, it executes and logs the outcome. The human can also initiate the interaction themselves, posing a question or directing an action without waiting for the agent to surface something first.

In practice, a given workflow involves a mix of these patterns depending on the situation. A routine exception might move quickly from identification to approval with minimal interrogation. A complex supply disruption, a cryo-shipment excursion or a manufacturing delay that threatens a patient's treatment window, might involve multiple rounds of direction before the human is ready to confirm. What matters is that the interaction pattern is designed deliberately for each workflow rather than be left to emerge on its own.
 

In Practice: A Slot Recovery That Actually Works

An AT manufacturer with a single therapy program running across two geographies generates a volume of monitoring, synthesis, and exception management work across apheresis scheduling, slot management, logistics management, manufacturing scheduling, and release coordination that already requires a significant number of dedicated, practitioners to execute manually and provide an expected “white glove” service. Scale that to five therapies across three regions each, and the staffing requirement compounds faster than any hiring plan can absorb it. Most manufacturers do not have the time or budget to hire and maintain practitioners at that scale, and in many cases those practitioners do not exist in the numbers the work would require even if the budget did. Digital staff close that gap not by working faster than humans, but by enabling humans to operate at a scale no human team alone could staff.

Slot Recovery is where that gap shows up most starkly. A botched manufacturing slot recovery in AT has two consequences: a patient who needed that slot did not get it, and the manufacturer absorbs the full cost of goods for a batch that was manufactured and never infused. It is one of the highest-stakes, most time-sensitive workflows in commercial operations, and in most AT organizations today it is overwhelmingly manual.

Today. A cancellation or reschedule notification lands in the order orchestration system. Someone has to read through it manually to figure out whether it falls within a window that demands action: an imminent cancellation needs an immediate response, a distant one might not. That triage alone eats time and judgment.

If action is warranted, the team member has to figure out which slot actually opened up, cross-referencing the orchestration system against a separate slot management system to see what's really available. To work out how to fill it, and which sites and patient profiles even qualify, they dig through approval records and reference data scattered across that second system.

Then comes the manual pass through the patient queue: matching profiles against eligibility criteria using a mix of structured data and knowledge that lives mostly in people's heads. Even once candidates are identified, the harder question remains: who's not only technically eligible but actually schedulable, given clinical status, site readiness, and program-specific constraints.

In practice, most of that analysis never happens. There isn't time. So, the default becomes blanket outreach to treatment sites about patients who may or may not be approved or viable for the slot. The result: wasted site time, outreach that goes nowhere, relationship capital spent on calls that shouldn't have been made while the patients who were genuinely strong candidates may never surface at all.

Tomorrow. A digital staff member, watching slot status and the patient queue continuously, catches the cancellation the moment it's logged. It determines the window and flags whether action is required automatically.

If recovery is warranted, it pulls the released slot, cross-references site approvals and reference data, and works the patient queue against eligibility criteria on its own. It surfaces a ranked shortlist to the coordinator: patients who are approved for the slot, clinically appropriate, and realistically schedulable, each with the supporting rationale already assembled.

The coordinator reviews the shortlist, applies judgment on anything the digital staff member couldn't assess, chooses the recovery path, and confirms. The outreach still happens (and is still performed by the coordinator), but now it's precise and pre-qualified instead of a blanket call. The digital staff member documents the outcome and updates the schedule everywhere it needs to be updated, leaving an auditable trail.
 

Bolting Digital Staff Onto an Existing Operating Model Will Fail

The manufacturers who struggle with this transition will fail because they bolted digital staff onto an operating model designed for direct human-to-system interaction. Three dimensions need to be defined explicitly.

1. The organizational dimension. Digital staff operate inside the same functional structures as their human counterparts. They have owners, governance structures, and escalation paths. The distinction from conventional software is meaningful: software tools have administrators, digital staff have managers. Someone is accountable for how a digital staff member performs, what guardrails it operates within, and what happens when it fails. Where digital staff operate across functional boundaries, governance cannot sit within a single function. No one team can own them unilaterally.
2. The process dimension. Not every workflow is a good candidate for digital staff, and deploying them in the wrong ones means choosing an expensive tool when a simpler one would do the job better. Routine processes, ones that are deterministic and rules-governed, are better served by traditional automation. It is cheaper, simpler, and fully auditable. Digital staff are best deployed on complex workflows: situations where context and judgment matter such as chain-of-identity reconciliation across systems, supply risk management, apheresis-to-infusion scheduling, contract development and manufacturing organization (CDMO) partner alignment, and slot recovery. These are workflows where the right answer depends on factors that cannot be fully anticipated at build time. This distinction is not new. It is the same judgment a manager applies when deciding which tasks to delegate to a junior team member versus which to handle personally. The difference is that the junior team member is now a digital staff member.
3. The role dimension. Digital staff eliminate tasks, not roles. The cognitive load of functional roles redistributes toward work that requires irreducible human judgment: interpreting ambiguous signals in context, navigating the relationships that underpin AT operations, directing digital staff and escalating when they fail, and making calls under uncertainty. For the analysts and coordinators whose expertise has been defined by knowing how to navigate these systems, the transition is real and will take time. The skills gap to close is not technical literacy but the ability to direct, interrogate, and maintain accountability alongside digital staff. Closing the gap requires deliberate investment in role redesign, training, and recruitment.

The Manufacturers Who Start Now Will Scale a Working Model. Those Who Wait Will Retrofit a Broken One.

As more advanced therapies reach market, operational execution becomes the competitive differentiator. In AT that execution is unforgiving: a living product tied to a single patient, batch sizes of one, scarce manufacturing slots, and a vein-to-vein clock that turns every exception into a patient-level risk rather than a delayed order. The manufacturers who can orchestrate that reliably, catch exceptions before they reach the patient, and hold chain of identity across a growing patient population will build a structural advantage. 

Building the capability to do that takes longer than most organizations expect. Designing workflows, establishing governance, calibrating digital staff, and training practitioners to direct rather than navigate is organizational work, not technology work. In a GxP environment where every action is audited and chain of identity is non-negotiable that calibration cannot be rushed and does not come with a platform purchase. Manufacturers who start while their pipelines are still manageable will have a working model when commercial scale pressure arrives. Those who wait will be retrofitting under that pressure, and the retrofit always costs more and takes longer than building it right the first time.

As governance matures across the industry, the distributed control planes manufacturers are building today will converge into a unified enterprise-wide layer. Manufacturers who build their integrated digital fabric deliberately from the start will be better positioned to lead that consolidation. The place to start is not a platform selection or a technology roadmap. It is picking one high-stakes workflow, slot scheduling, order orchestration, or deviation handling, designing the human-digital staff interaction deliberately, and building the governance muscle from there. 
 

About This Perspective

This point of view reflects the accumulated perspective of Deloitte's NextGen Therapies practice, developed through more than a decade of commercial, supply chain, and digital infrastructure engagements for AT manufacturers across therapy types and geographies.

The operating model described here is grounded in the practice's proprietary integrated digital capability framework for value chain management, and in the digital staff architecture the practice has developed through active client deployments across commercial and operations functions.

This publication contains general information only and Deloitte is not, by means of this publication, rendering accounting, business, financial, investment, legal, tax, or other professional advice or services. This publication is not a substitute for such professional advice or services, nor should it be used as a basis for any decision or action that may affect your business. Before making any decision or taking any action that may affect your business, you should consult a qualified professional advisor.

Deloitte shall not be responsible for any loss sustained by any person who relies on this publication.
 

References
 

1. Young CM, Quinn C, Trusheim MR. Durable cell and gene therapy potential patient and financial impact: US projections of product approvals, patients treated, and product revenues. Drug Discov Today. 2022;27(1):17-30. doi:10.1016/j.drudis.2021.09.001
2. Quinn C, Young C, Thomas J, Trusheim M; MIT NEWDIGS FoCUS Writing Group. Estimating the clinical pipeline of Advanced Therapies and their potential economic impact on the US healthcare system. Value Health. 2019;22(6):621-626. doi:10.1016/j.jval.2019.03.014
3. Salesforce. Introducing Salesforce Headless 360. No browser required. Salesforce News. Published April 15, 2026. Accessed June 29, 2026. https://www.salesforce.com/news/stories/salesforce-headless-360-announcement/
4. ServiceNow. ServiceNow opens its full system of action to every AI agent in the enterprise. ServiceNow Newsroom. Published May 5, 2026. Accessed June 29, 2026. https://newsroom.servicenow.com/press-releases/details/2026/ServiceNow-opens-its-full-system-of-action-to-every-AI-Agent-in-the-enterprise/default.aspx

About Deloitte

Deloitte refers to one or more of Deloitte Touche Tohmatsu Limited, a UK private company limited by guarantee (“DTTL”), its network of member firms, and their related entities. DTTL and each of its member firms are legally separate and independent entities. DTTL (also referred to as “Deloitte Global”) does not provide services to clients. In the United States, Deloitte refers to one or more of the US member firms of DTTL, their related entities that operate using the “Deloitte” name in the United States and their respective affiliates. Certain services may not be available to attest clients under the rules and regulations of public accounting. Please see www.deloitte.com/about to learn more about our global network of member firms.