Independent · free · vetted · no vendors

info@pharmaaicollective.org

Pharma AI Collective
Menu
Week 1 · unit 2 of 7

  1. Brief

    1. The confident intern
    2. From answering to acting
    3. Where it lands in pharma
    4. Follow the money

  2. Try

    1. Same question, three ways

  3. Ask

    1. Questions for Monday

  4. Build

    1. Page 1: the opportunity map

From answering to acting

Introduction

Chat, assistant, workflow, agent: what changes is who chooses the next step and what the system may touch.

Exercise

The demo is tomorrow

A vendor wants to demo an "autonomous deviation agent" to Meridian's Head of Quality. It reads the electronic quality system (eQMS), queries the manufacturing execution system (MES) and the lab system (LIMS), drafts the investigation and routes it for approval. You get one question before the demo. Which do you ask?

The demo is tomorrow

Show answers
Worth knowing for cost and contract terms, but the model is the most replaceable part of the system. It tells you nothing about what could go wrong in Meridian's eQMS.
This sets Meridian's exposure. Reading the MES is one risk; routing an investigation or changing a record is another. The answer tells QA what it would have to govern.
A fair question, but an accuracy figure from other companies' deviations says little about Kingsford's, and accuracy matters less than what the system does when it is wrong.
Validated for what intended use? Until you know what it may do on its own, nobody can define the use to validate. Ask the scope question first, then this one.

What the system may do without a person is the decision that sets your risk, your validation effort and your accountability. Everything else follows from it.

Generative AI describes what the model does: it produces content. That can happen inside a chat, a fixed workflow or an agent, so classify each deployment, not the technology. In a chat or assistant, a person asks and decides what to do with the answer. In a workflow (AI), people designed the path, possibly with fixed branches, and a model does one or more steps. An agent gets a goal and tools, such as searching a system, running a query or sending a message, chooses its next step, looks at the result and repeats.

Autonomy is a design choice, not a product feature. The same model can sit behind a chat box, a fixed workflow or an agent. Widely cited guidance for builders advises starting with the simplest option that works, which can mean no agent at all (Anthropic, 2024 (opens in a new tab)).

The agent loop, and where you decideCycle diagram: a person sets a goal; the agent plans, acts through tools such as the MES and eQMS, observes the results, which may contain wrong or planted content, and repeats until the goal is met or a limit is reached. It then hands over to a person. A marker shows the review point can be placed at the plan, before any write action, or at the end.GoalSet by a person:Draft theinvestigationfor thisdeviationPlan1Act2MES:batch alarmseQMS: similardeviationsObserve!May include wrongor planted contentGoal met,or limitreached?No, repeat3YesHand overA person reviews and signs, or doesn’t.Your decision: where the person sits1Approve the plan2Approve before any write action3Review at the end The agent loop, and where you decideCycle diagram: a person sets a goal; the agent plans, acts through tools such as the MES and eQMS, observes the results, which may contain wrong or planted content, and repeats until the goal is met or a limit is reached. It then hands over to a person. A marker shows the review point can be placed at the plan, before any write action, or at the end.GoalSet by a person: “Draft theinvestigation for this deviation”Plan1Act2MES:batch alarmseQMS: similardeviationsObserve!May include wrongor planted contentGoal met,or limitreached?No, repeat3YesHand overA person reviews and signs, or doesn’t.Your decision: where the person sits1Approve the plan2Approve before any write action3Review at the end
Where a person reviews is a design decision. The numbered marks show three places the review can sit.
Text description

Goal. A person sets it: "Draft the investigation for this deviation."

Plan. The model decides what it needs next.

Act. It calls a tool: pull the batch's alarms from the MES, search the eQMS for similar deviations.

Observe. It reads what came back, including anything wrong or planted in that data.

Repeat or stop. It loops until it judges the goal met, or hits a limit someone set.

Hand over. A person reviews and signs, or doesn't. Where this step sits is your decision.

Exercise

Sort Meridian's AI uses

Classify each deployment as described: who chooses the steps, and what can it touch? (Many chat products can also search or use tools; here, assume only what the description says.)

A QA director asks the enterprise assistant to explain, in general terms, how a deviation differs from a nonconformance.
A technical writer pastes a draft SOP and asks for a plain-language summary for operators.
The digital lead uploads two versions of a batch-record template and asks what changed.
Every new product complaint is classified by a model into one of eight categories and routed to the matching queue.
Each month a scheduled job pulls open corrective and preventive actions (CAPAs), a model writes a one-paragraph status for each, and the pack goes to the quality review board.
Given a deviation number, the system decides which batch records, alarms and past CAPAs to pull, queries the systems and drafts root-cause hypotheses.
A supply planner asks for alternative resin suppliers; the system searches the web, compares options and emails three suppliers for quotes.

Show answers
General knowledge from training. Fine for orientation; Meridian's own definitions live in its SOPs, which the model hasn't seen.
It works on what it was given, and a person decides whether to use the result. Low autonomy; the writer stays accountable.
Same pattern: supplied material, a person checks the answer. Useful and cheap to govern.
The path is predefined; the model does one step. A fixed path is usually easier to test for repeatable work.
Tempting to call it an agent because it runs unattended. Running on its own isn't the same as choosing its own steps; the path is predefined.
It chooses what to look at and calls the systems itself. That flexibility is the value, and it is why QA must know exactly what it can reach.
It acted outside the conversation, in Meridian's name. Sending things is where governance changes, whatever the use case.

Governance effort rises from chat to agent, whatever the demo looks like. Two questions sort any proposal: who chooses the next step, and can the system change anything outside the conversation?

Four deployments on two axesPlot of four AI deployments by who chooses the next step and what the system can touch. Chat and assistant sit where a person chooses and the system touches only the conversation or material it is given. A workflow follows a designed path. An agent chooses its own steps and can act in other systems. Governance effort rises toward the agent.What it can touchSystems it writes toSystems it readsMaterial you give itThe conversation onlyChatTraining onlyAssistantPerson acts on itWorkflowFixed path; mayread and writewithin its stepsAgentChooses its ownsteps and actsin other systemsThe same modelcan sit behindall four.Governance effortrises toward the agentPersonPath fixedby designersModel at run timeWho chooses the next step Four deployments on two axesPlot of four AI deployments by who chooses the next step and what the system can touch. Chat and assistant sit where a person chooses and the system touches only the conversation or material it is given. A workflow follows a designed path. An agent chooses its own steps and can act in other systems. Governance effort rises toward the agent.Governance effort risesChatChooses: a personTouches: the conversation onlyAssistantChooses: a personTouches: material you give itWorkflowChooses: path fixed by designersTouches: systems it readsand writes, within fixed stepsAgentChooses: the model, at run timeTouches: systems it reads andwrites to, through toolsThe same model can sitbehind all four.
Classify the deployment, not the technology.

Note

Sources for this unit

  • Building effective agents — Anthropic, 2024-12-19. Read the source (opens in a new tab) The most cited practical distinction between workflows and agents, and the advice to start simple. Written for builders.
  • Equipping agents for the real world with Agent Skills — Anthropic, 2025-10 (updated 2025-12-18). Read the source (opens in a new tab) How a skill is packaged and loaded only when needed, and why some steps belong in ordinary code.
  • Agent Skills: an open format — agentskills.io, accessed 2026-11-02. Read the source (opens in a new tab) The vendor-neutral home of the skills format.