Agentic AI in Commercial Pharma: Fewer Tools, More Follow-Through

A patient start rarely depends on a single team. Field engagement, payer requirements, case work, and pharmacy fulfillment all shape whether someone begins therapy, so improving only one step in the process can often have little-to-no impact on the overall goal. That problem shapes how I think about AI agents in commercial pharma.

Design around decisions, not departments

Since starting at ConcertAI, I have become a big believer that AI should be designed around decisions that cross functions. The goal is to carry the evidence, the proposed action, and the responsible owner through the entire decision.

A real-world example I have seen several times in my 25+ years working in pharma: imagine leadership sees paid starts falling behind plan in a particular segment. The first job is to work out whether the constraint is demand, access, fulfillment, or incomplete data. Connecting hub, pharmacy, and payer evidence might reveal a new documentation requirement. Market access confirms the policy and the affected population. Patient access prioritizes cases and helps specialists complete the required work. The brand team sees which provider and payer segments need support and adjusts resources.

Each team needs a view suited to its role and permissions, but they should share one understanding of the situation, who owns the next step, and which dependencies remain open. That prevents duplicate or premature outreach. It also helps improve efficiency and coordination: often a problem detected by one function needs action from another function.

Make “agent” earn its name

I’d encourage every commercial leader to be demanding about the term “agent”. A language model can interpret a problem and propose steps. An agent also needs approved tools, a record of what has happened, and a way to check the result before choosing its next move.

Consider preparing a reimbursement request. The system needs to establish whether it was approved, whether the receiving system accepted it, and what remains unresolved. A generated statement that the work is complete is not independent evidence of completion. We evaluate the whole sequence across repeated trials, including conflicting information, failed tool calls, and cases requiring human approval.

Prove the value, and be honest about what you’ve proven

We separate model quality, execution quality, and business impact. An accurate summary or a successful tool call doesn’t establish that access improved. The record must connect the original barrier, the intervention, and an independently observed outcome.

Our own reported patient access results show why this matters. We did a small study with a 15-person team with one of our customers who used ConcertAI’s agentic commercial platform Cadence Suite from May through August 2026, but had been using a non-agentic approach from May through August 2025. We found that resolution time fell from 22.2 to 14 days by using the agentic approach, and 156 more patients were on therapy. I find that encouraging operating evidence, though to be fair, the comparison does not completely isolate AI’s contribution.

My advice is to agree on the baseline and comparison before deployment, accounting for case mix, payer policy, and staffing. Then include integration, review, and ongoing operating costs, and count each outcome once. That gives leaders a defensible basis for deciding what to expand.

Context comes before cleverness

An agent needs the right context for a decision, and that context has to stay current as the work progresses. Retrieving a payer policy paragraph is useful, but applying it requires knowing the plan, effective date, affected population, and whether the requirement has already been satisfied.

Data semantics matter just as much. A paid claim, a shipment, and treatment initiation are different events, and a delayed feed can’t automatically become evidence that a patient stopped therapy. Systems must preserve those distinctions and recognize uncertainty. We favor a reusable foundation that gives each model or agent the relevant, permitted evidence for its task, so teams can interpret the same situation consistently.

Accountability travels with the decision

Model confidence is neither authorization nor proof of correctness. Preparing an internal summary carries different consequences from sending a patient communication or making a contractual commitment. The software executing the action should enforce permissions and approvals outside the language model. Clinical decisions remain with clinicians.

Our approach is expanding delegation as a workflow demonstrates reliable performance, including on difficult exceptions. Reviewers need enough context and time to challenge a recommendation, and the organization needs a way to stop or recover the work. Shared controls make responsible scaling more practical, because each new workflow can reuse the same mechanisms for evidence, authorization, and review.

Start with one workflow

Looking ahead, I expect lasting advantage to come from an organization’s ability to retain and improve its decision knowledge: which intervention was chosen, under what conditions, and what happened afterward. Accumulating interactions alone doesn’t establish improvement. The learning mechanism needs to be explicit, with changes tested against representative cases and assessed through a controlled rollout.

For a leader starting now, here is what I’d do:

1. Choose one consequential workflow.
2. Establish its baseline, owner, and decision rights.
3. After 90 days, review whether the evidence supports expansion or whether more observation is needed.
4. Build it so another role or brand can reuse the trusted context and controls.

Over time, the organization can develop a commercial operating model in which local expertise and enterprise learning reinforce each other.

Connecting the dots

This is the thinking behind how ConcertAI approaches commercial intelligence: shared context, specialized tools for each role, and workflows that close the loop from insight to verified outcome.