The challenge
A hospital pharmacist checking a drug monograph mid-shift. A resident working out which residency programs take her specialty and when applications close. A policy staffer who needs the society’s published position on a bill before a meeting starts.
These are not variations of one question. They come from different people, draw on different bodies of content, and belong to different departments inside the society. A single assistant trained on everything would answer all three badly.
The society’s problem was never a shortage of content. It was that each department needed its own narrow, accurate answer service, and none of them could justify a project of their own.
The solution
One corpus per assistant. Drug reference material and residency deadlines have nothing to do with each other, and an assistant holding both will blend them at exactly the wrong moment. Narrow scope is an accuracy decision before it is a design decision.
It reports what is published, and it does not practice. A society can tell you what its own monograph or position statement says. It cannot make a clinical judgment for a pharmacist at a bedside. Retrieval limited to published material keeps that boundary structural.
Standing up the tenth agent has to be trivial. If each new department needs its own contract, budget line and procurement cycle, the second one never happens.
Internal audiences count. Staff asking each other how the membership system works is the same problem as members asking how a program works, and internal agents need no brand review or legal sign-off.
What the nine actually are
Drug information came first, over the society’s formulary reference, and it remains the busiest by a wide margin. Residency programs followed, with a dedicated agent for the resources residents and preceptors ask about every cycle. Advocacy and practice advancement each got their own, one for policy positions and one for the society’s national survey work.
A general society assistant, a content search agent and a news agent cover knowledge discovery and member services. The last two sit over the association management system’s own documentation, so staff can answer their own questions about the platform they work in daily.
The capability that made it possible
What stops an organization putting an assistant in front of a second department is rarely the technology. The second one needs a business case, a budget line and somebody to own the procurement, and at that point the idea quietly dies.
Here, a department with a body of content and a recurring question can have an assistant over it, scoped to that content alone, without asking anyone for a new contract. No procurement between agents, no API build, no connector work, no engineering team standing behind it. The ninth agent cost the same as the second, which is why there are nine and not one.
The results
| MEASURE | TO DATE |
| Purpose-built assistants in production | 9 |
| Documents processed | 6,767 |
| Words processed | 35.2M |
| Questions in the busiest month | 1,006 |
| Support tickets raised | 0 |
The number that matters is nine. Most organizations that buy an assistant end up with one, because the second has to be argued for. This society has agents for drug information, residency programs, advocacy, practice advancement, news, member search and events, plus two pointed at internal staff, and none of them required a separate purchase.
The drug information assistant alone has carried the majority of every question asked across the estate, which is where the corpus weight sits.
Three years, nine agents, zero support tickets. All nine run on a single entry-level subscription, so the platform decision was made once and every department since has been a configuration rather than a purchase.
Why it worked
Scope is per agent, not per account. Each assistant is grounded in its own content and nothing else, so the residency agent cannot answer a drug question badly. For clinical content, that separation is an accuracy requirement.
Adding one is a configuration. No contract, no procurement, no engineering ticket. A department with content and a recurring question can be answered this week.
It keeps working unattended. Three years with no support tickets, across agents owned by departments rather than by IT.
The newest agent is the most capable. It uses retrieval, web search and code execution together, built years after the first one and on the same subscription.
