Last week we drew the line between generative AI and agentic AI. Generative writes. Agentic acts. That is the category.
This week is the next problem. How do you tell what you are actually buying when every vendor in your inbox calls their product agentic?
Print these five capabilities. Bring them to the next vendor demo. Mark which ones the vendor can demonstrate, not just describe. That is your filter.
Capability 1: Goal-setting
A real agentic system accepts an outcome, not just a prompt. You can tell it what you want lifted, by how much, and inside what guardrails. It plans the steps to get there.
In pharma marketing this means a brand manager does not paste "write me a subject line" into a tool. They set a goal of "lift unbranded HCP open rate on the diabetes content set by 15 percent in Q3, within the approved channel mix." The system breaks that into work.
The vendor-demo test. Ask the vendor to show the same product running against two different goals, with the system actually deciding on different paths. If both demos look identical regardless of goal, the system is responding to prompts. It is not pursuing outcomes.
Goal-setting is what separates a tool you have to drive from a system that drives toward something.
Capability 2: Tool use
A real agentic system can call other systems. CRM, marketing automation, the approved content library, analytics. It acts on what it finds. It does not generate output for a human to copy and paste.
In pharma marketing this means the system can pull a list from Veeva, segment it inside the rules of your approved audience definitions, pull approved content from your DAM, schedule a multi-channel campaign in your MA tool, and report back. Each of those is a real tool call. The brand manager touches the dashboard. The system touches the in-between steps.
The vendor-demo test. Ask the vendor to show one workflow that touches three of your real systems live in the demo. Generative tools can describe integrations. Agentic systems prove them in one screen-share.
If the demo lives entirely inside the vendor's own product, the system is generative. If the demo crosses systems your team already runs, the system is closer to agentic.
Capability 3: Memory and state
A real agentic system remembers what it tried, what worked, what failed, and what the human reviewed. It builds a state over time. Without memory, the system makes the same mistakes weekly.
In pharma marketing this means the system knows that two months ago it tried a variant that did not pass MLR. It does not try it again. It also knows which content variants you have already approved this quarter, and what the open-rate baseline was before it started measuring.
The vendor-demo test. Ask what the system remembers from one workflow run to the next, and where that memory is stored. If the answer is "the prompt history," it is generative with persistence. If the answer involves a state store and explicit recall, you are closer.
Memory is what makes the difference between a system that improves and a system that resets. Without it, your team becomes the memory. That is back to where you started.
Capability 4: Evaluation
A real agentic system grades its own output against a metric you defined, and adjusts. It can tell when it is below the bar and when it has cleared it. It does not need a human to inspect every artifact to know what to do next.
In pharma marketing this means the system sends a content variant, watches open rate at the four-hour mark, compares to the baseline you set, and either lets it run or swaps it for the next best variant from the approved library. The human reviews the system's decisions weekly. The system handles the daily.
The vendor-demo test. Ask the vendor to define, inside their product, what "good" looks like for a specific commercial goal. Then ask them to show the system actually deciding based on that definition. If "good" is defined by the human in every run, the system is a tool. It is not an agent.
Evaluation is the capability that closes the loop. Generative AI produces. Agentic AI produces, checks, and adjusts.
Capability 5: Escalation
A real agentic system knows when to stop and route to a human. It identifies edge cases, ambiguity, regulatory boundaries, and unexpected results. It pauses rather than guessing.
In pharma marketing this means the system tries to send a piece of content to an audience segment that falls outside the approved definitions for that brand. It does not silently expand the segment. It stops, flags the conflict, and routes to the brand manager with the specific reason and the relevant policy reference.
The vendor-demo test. Ask the vendor to break their own product on purpose. Set a goal that triggers an edge case or a guardrail violation. Watch what happens. If the system pushes through, it is not safe to run in a regulated environment. If the system stops cleanly with a human-readable explanation, it is.
Escalation is the capability that makes agentic safe to deploy in pharma. The other four make the system fast. This one makes it accountable.
The takeaway
Five capabilities. Goal-setting, tool use, memory and state, evaluation, escalation. Each one has a question you can ask in a demo that the vendor cannot fake on screen.
These are the questions we are pressure-testing at the Agentic Pharma Summit in November, with the commercial teams actually deploying agentic systems today. The vendors will keep stretching the label. Your job is to know what good looks like before you sign.
Print the five. Bring them to the next demo. See what you actually have on the table.
This is the second post in Agentic Pharma Insights, a weekly briefing for pharma commercial leaders deploying agentic AI across marketing, sales, customer engagement, patient experience, and market access. Published by the Agentic Pharma Summit, November 9-10, 2026, Philadelphia.