# AI Bootcamp for Pharmacosmos ## Today's goal: one task, one clear prompt, one reusable GPT Pharmacosmos - Køge 30 April 2026 <div class="lab-grid"> <div class="lab-cell"><b>1 work task</b><span>Choose a real task from daily work where AI can help safely.</span></div> <div class="lab-cell"><b>1 clear prompt</b><span>Tell AI what the task is, what boundaries apply, and what the answer should look like.</span></div> <div class="lab-cell"><b>1 reusable GPT</b><span>Save the strong prompt as a GPT that can be used again.</span></div> </div> > Prompt = your instruction to AI. GPT = ChatGPT with fixed instructions for a specific task.

# Today's material is a website, so you can use it actively ### In the room - You can move forward and back at your own pace - We will use selected slides, exercises and live demos - The examples are templates; the real cases come from your own processes ### After today - You can return to prompts, sources and exercises - The material works as a reference site, not only as a presentation - The website format makes it easier for me and AI to improve the material together

# Today's work product Today each participant builds a small, safe AI work pattern: <div class="workflow-chain"> <span>Task<br>from daily work</span> <span>Subtask<br>AI may help with</span> <span>Prompt<br>first test</span> <span>Expert review<br>human owns it</span> <span>GPT instruction<br>can be reused</span> <span>Baseline<br>next test</span> </div> The goal is not to master all of AI. The goal is to use Enterprise ChatGPT for one concrete task in a better and more controlled way. > Baseline = how the task works today, before the AI test.

# Geoffrey Hinton: AI is powerful, but not predictable Geoffrey Hinton is one of the key figures behind the technology many modern AI tools build on. His point matters today: AI can be very useful, but it must be guided and controlled. <div class="lab-grid"> <div class="lab-cell"><b>Expectation of me</b><span>I manage the frame, pace and safety in the room.</span></div> <div class="lab-cell"><b>Expectation of the model</b><span>AI delivers suggestions, patterns and questions - not final truth.</span></div> <div class="lab-cell"><b>Expectation of you</b><span>Your expertise decides what is correct, relevant and safe.</span></div> </div> <a class="resource-link-card" href="https://www.youtube.com/watch?v=zKM-msksXq0" target="_blank" rel="noopener noreferrer"> <b>Open the Geoffrey Hinton clip</b> <span>Use it as perspective on why AI requires curiosity, rules and human control.</span> </a> [Nobel Prize facts](https://www.nobelprize.org/prizes/physics/2024/hinton/) · [University of Toronto](https://web.cs.toronto.edu/news-events/news/geoffrey-hinton-wins-nobel-prize-in-physics)

# AI does not replace your expertise - it moves where it is used > Less manual production. More clear task framing, expert review and decision-making. <div class="lab-grid"> <div class="lab-cell"><b>Before</b><span>I write, collect, sort and double-check most of it myself.</span></div> <div class="lab-cell"><b>Now</b><span>I give AI a bounded task and review the quality.</span></div> <div class="lab-cell"><b>Today</b><span>Each person chooses one process where AI can help without taking over expert responsibility.</span></div> </div>

# The work pattern changes: humans plan and evaluate <div class="workflow-chain"> <span>Human<br>defines task</span> <span>AI<br>drafts or structures</span> <span>Human<br>reviews</span> <span>AI<br>improves</span> <span>Human<br>decides</span> </div> <div class="lab-grid"> <div class="lab-cell"><b>Planning</b><span>The human defines purpose, boundaries, input and output.</span></div> <div class="lab-cell"><b>Production</b><span>AI can produce a draft, table, checklist or questions.</span></div> <div class="lab-cell"><b>Evaluation</b><span>The human checks evidence, quality, risk and fit for use.</span></div> </div> The key shift is not less expertise. It is moving expertise into task framing, review and decision-making.

# AI creates value only when expertise, tool and process fit together > Today we are not training clever one-off questions. We are training a clear way to instruct, review and improve AI work. <div class="lab-grid"> <div class="lab-cell"><b>Expertise</b><span>Your judgment decides what is correct, relevant and safe.</span></div> <div class="lab-cell"><b>Technology</b><span>AI suggests from patterns. It can help quickly, but it must be guided and checked.</span></div> <div class="lab-cell"><b>Process</b><span>We start with your current workflow, not with the AI tool.</span></div> </div>

# First: look at the work from 10,000 feet Do not ask AI to copy today's workflow. Use AI to inspect the work before doing the work. <div class="lab-grid"> <div class="lab-cell"><b>Purpose</b><span>What is this work meant to achieve, and who uses the output?</span></div> <div class="lab-cell"><b>Workflow</b><span>Which steps create value, risk, delay or repeated review work?</span></div> <div class="lab-cell"><b>AI support</b><span>Where could AI draft, structure, question or check without owning the conclusion?</span></div> </div> ```text Ask me questions about the purpose, users, steps, expert judgment and safe AI support. ```

# New processes: from human alone to human plus AI <div style="display: grid; grid-template-columns: 135px 1fr; gap: 14px; align-items: center; max-width: 1120px; margin: 8px auto 18px;"> <div style="font-size: 25px; font-weight: 900; color: #64748b; text-transform: uppercase;">Before</div> <div class="workflow-chain" style="grid-template-columns: repeat(3, 1fr); margin: 0;"> <span>Human drafts</span> <span>Human checks</span> <span>Human improves</span> </div> <div style="font-size: 25px; font-weight: 900; color: #174ea6; text-transform: uppercase;">Now</div> <div class="workflow-chain" style="grid-template-columns: repeat(5, 1fr); margin: 0;"> <span>Human frames</span> <span>AI suggests</span> <span>Human validates</span> <span>AI iterates</span> <span>Human owns decision</span> </div> </div> You should not only find a good question for AI. You should find a process step where AI can help without removing responsibility, safety or expert review.

# Others in pharma already use AI - but we start with the work pattern <div class="lab-grid"> <div class="lab-cell"><b>AstraZeneca</b><span>AI and data science are used across discovery, clinical trials and development. <a href="https://www.astrazeneca.com/r-d/data-science-and-ai.html">Source</a></span></div> <div class="lab-cell"><b>Roche / Genentech</b><span>Roche and NVIDIA describe generative AI for drug discovery and development flows. <a href="https://www.roche.com/investors/updates/inv-update-2023-11-21">Source</a></span></div> <div class="lab-cell"><b>Pfizer / CytoReason</b><span>AI-driven disease models are used for development decisions. <a href="https://www.pfizer.com/news/press-release/press-release-detail/cytoreason-announces-expanded-collaboration-deal-pfizer">Source</a></span></div> </div> -> We are not doing the same thing today. We are learning the pattern: bound the task, set the data boundary, and review AI output.

# Novo Nordisk and OpenAI make generative AI part of pharma infrastructure On 14 April 2026, Novo Nordisk announced a strategic collaboration with OpenAI to integrate advanced AI across research, manufacturing, supply chain and commercial operations. <div class="lab-grid"> <div class="lab-cell"><b>Discovery and data</b><span>AI will help analyze complex data sets and identify promising drug candidates.</span></div> <div class="lab-cell"><b>Operations and scaling</b><span>Pilots cover R&D, manufacturing and commercial operations, with full integration as the goal by the end of 2026.</span></div> <div class="lab-cell"><b>Control</b><span>The collaboration is framed around data protection, clear rules and human oversight.</span></div> </div> The point for us: AI is not only a writing tool. It becomes work infrastructure that requires safe processes, strong instructions and expert review. [Berlingske: Novo Nordisk enters partnership with OpenAI](https://www.berlingske.dk/virksomheder/novo-nordisk-indgaar-samarbejde-med-openai--faar-aktien-til-at-stige?gaa_at=eafs&gaa_n=AWEtsqeR6JClWeEkSW_9FICLTdmbGS93sEJNoSOykyOcRGvCkIMRu5ZYsG-xcwevFY8%3D&gaa_ts=69f23c30&gaa_sig=WJiN0SMrfQKcxcXju76ArdgEHSAx42ftz1c09vOcX4CX4XpMtLjtJQqYCss8xFyiFo3TZU0ink0YsyHPW4JBZw%3D%3D) · [Press release via Nasdaq](https://www.nasdaq.com/press-release/novo-nordisk-and-openai-partner-transform-how-medicines-are-discovered-and-delivered)

# AI tools evolve faster than our work routines OpenAI's release notes are a useful reminder: ChatGPT is not a static tool. In a little over three years it has moved from simple chat to GPTs, images, browsing, files, apps, research and more automated workflows. <div class="lab-grid"> <div class="lab-cell"><b>Exercise now</b><span>Open the ChatGPT release notes and scroll down to the first ChatGPT notes from December 2022.</span></div> <div class="lab-cell"><b>Look for</b><span>How quickly the features move from chat to workspaces, tools and integrations.</span></div> <div class="lab-cell"><b>The point</b><span>We are not learning one button. We are learning a work pattern that can keep up.</span></div> </div> [Open ChatGPT release notes](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)

# This website was built with a small AI team <div style="display: grid; grid-template-columns: 230px minmax(0, 1fr); gap: 7px 12px; align-items: stretch; max-width: 1040px; margin: 4px auto 12px; font-size: 21px; line-height: 1.16;"> <div style="font-size: 14px; font-weight: 900; color: #64748b; text-transform: uppercase;">Person / agent</div> <div style="font-size: 14px; font-weight: 900; color: #64748b; text-transform: uppercase;">Task in the workflow</div> <div style="padding: 9px 12px; border-radius: 7px; background: #e8f1fb; font-size: 22px; font-weight: 900; color: #0f2f57;">Torben</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f8fafc; border: 1px solid #d8dee8; font-size: 21px;">Defines goal and task</div> <div style="padding: 9px 12px; border-radius: 7px; background: #eef4f1; font-size: 22px; font-weight: 900; color: #124437;">ProjectLead</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f8fafc; border: 1px solid #d8dee8; font-size: 21px;">Aligns goals, rules and direction</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f7f4ed; font-size: 22px; font-weight: 900; color: #5c3d00;">Thomas Director</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f8fafc; border: 1px solid #d8dee8; font-size: 21px;">Checks sponsor value and pharma fit</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f5eefb; font-size: 22px; font-weight: 900; color: #4b2375;">Medical Science Audience</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f8fafc; border: 1px solid #d8dee8; font-size: 21px;">Checks participant clarity</div> <div style="padding: 9px 12px; border-radius: 7px; background: #fff2e5; font-size: 22px; font-weight: 900; color: #6b3500;">Slide / Visual QA</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f8fafc; border: 1px solid #d8dee8; font-size: 21px;">Checks layout and readability</div> <div style="padding: 9px 12px; border-radius: 7px; background: #e8f1fb; font-size: 22px; font-weight: 900; color: #0f2f57;">Torben</div> <div style="padding: 9px 12px; border-radius: 7px; background: #f8fafc; border: 1px solid #d8dee8; font-size: 21px;">Guides and approves live version</div> </div> The point is not the technology. The point is the work pattern: clear roles, clear boundaries, multiple checks and human decision-making. > Today we are not building this kind of system for you. We start with the same pattern at a small scale: one task, one AI role and one review point.

# We work with process descriptions, not confidential documents <div class="safety-matrix"> <div class="safety-gate"> <h3>OK in the room</h3> <p>Role, process step, generic input type and desired output. No names, numbers, dates, products or documents.</p> </div> <div class="safety-gate caution"> <h3>Approved enterprise only</h3> <p>Confidential work content according to your internal rules. Use Enterprise ChatGPT, and keep human control.</p> </div> <div class="safety-gate stop"> <h3>Stop</h3> <p>Trade secrets, patient data, GDPR personal data, product formulas, contract terms, supplier names and non-public clinical details.</p> </div> </div> > When we practice today: describe the process without sharing the actual material.

# What does "sanitized" mean in practice? A sanitized process description may describe **what type of work it is**, but not reveal the actual content. <div class="prompt-builder"> <div class="prompt-step"><b>Yes</b><span>"We create an internal summary of a meeting."</span></div> <div class="prompt-step"><b>Yes</b><span>"Input is notes, output is decisions and open questions."</span></div> <div class="prompt-step"><b>Yes</b><span>"Expert review is done by the subject owner before sharing."</span></div> <div class="prompt-step"><b>No</b><span>Names, patients, products, numbers, dates and concrete document excerpts.</span></div> <div class="prompt-step"><b>No</b><span>Customer, supplier or deal details.</span></div> <div class="prompt-step"><b>No</b><span>Data or code that is itself IP or confidential.</span></div> </div>

# Mini exercise: where may the input be used? <div class="exercise-panel"> <div class="timer-display" data-seconds="360">6:00</div> <div> Place each example in one of three categories: - Public AI - Enterprise ChatGPT - Do not use in AI Examples: generic process description, protocol outline without product name, patient data, public research on a disease area, supplier name in contract context, product formula or recipe. </div> </div> > Today's exercises use only harmless process descriptions. When you work for real, follow your internal AI rules.

# Generative AI can help, but it does not automatically know what is true Generative AI can suggest: - text and structure - questions and review points - summaries and next steps - checklists and draft workflows That is why we always work with: **task -> material and rules -> boundaries -> desired answer -> human review**

# Basic language: six AI terms we use today <div class="prompt-builder"> <div class="prompt-step"><b>Model</b><span>The system that calculates a likely answer from patterns in training and context.</span></div> <div class="prompt-step"><b>Prompt</b><span>Your instruction: task, role, context, boundaries and desired output.</span></div> <div class="prompt-step"><b>Context</b><span>The information AI receives for the task: text, files, history, examples or rules.</span></div> <div class="prompt-step"><b>Context window</b><span>How much information the model can use at once. It is large, but not infinite.</span></div> <div class="prompt-step"><b>Hallucination</b><span>When AI answers fluently, but incorrectly or without a reliable source. That is why we review the answer.</span></div> <div class="prompt-step"><b>GPT</b><span>ChatGPT with its own instructions, knowledge and starter prompts for a recurring task.</span></div> </div>

# Three difficult AI terms that save review time <div class="lab-grid"> <div class="lab-cell"><b>Context window</b><span>AI only works with what is inside the window. Too much noise or old side tracks can make the answer worse.</span></div> <div class="lab-cell"><b>Conversation drift</b><span>When the chat slowly moves away from the original task because side tracks become new context. Technical term: prompt drifting.</span></div> <div class="lab-cell"><b>Hallucination</b><span>When the answer sounds confident, but is based on a wrong guess or missing source. Use it as a draft, not as truth.</span></div> </div> **Stop rule:** Start a new chat or restate the task, boundaries and source requirements when the work begins to drift.

# Temperature: how predictable should the answer be? Temperature is a model setting that changes how much variation the AI may use when it chooses the next words. <div class="lab-grid"> <div class="lab-cell"><b>Low</b><span>More predictable and consistent. Use for tables, checklists, rewriting, extraction and controlled summaries.</span></div> <div class="lab-cell"><b>Medium</b><span>Balanced. Use for sparring, process improvement ideas and comparing options.</span></div> <div class="lab-cell"><b>High</b><span>More varied and creative. Use for brainstorming, alternative framings and early idea generation.</span></div> <div class="lab-cell"><b>Important</b><span>Temperature is not a truth setting. Sources, boundaries and human review still matter.</span></div> </div> If you do not see a temperature knob, prompt for the behavior: "Be conservative and consistent" or "Give me varied alternatives."

# The 3 major language models <img src="assets/large-language-models.png" alt="Logos for ChatGPT, Gemini and Claude as examples of large-language-model workspaces" style="display: block; width: min(1080px, 94%); max-width: 1080px; max-height: 520px; height: auto; object-fit: contain; margin: 8px auto 14px;">

# From idea to improved solution - together with AI <div class="workflow-chain"> <span>Idea or problem</span> <span>Discuss with AI</span> <span>AI drafts</span> <span>Review and adjust</span> <span>Test together</span> <span>Improve and repeat</span> </div> <div class="lab-grid"> <div class="lab-cell"><b>Start early</b><span>Use AI before the draft is finished: ideas, alternatives and review questions.</span></div> <div class="lab-cell"><b>Stay in control</b><span>Define the task, inspect output and keep the expert decision with the human.</span></div> <div class="lab-cell"><b>Improve the pattern</b><span>When the answer is wrong, improve the instruction instead of only fixing the answer.</span></div> </div>

# ChatGPT: first chat, then GPTs We start with ordinary chat because it is the fastest way to learn the AI work pattern. But if a task repeats, it should not be reinvented every time. <div class="lab-grid"> <div class="lab-cell"><b>1. Chat</b><span>Test the task quickly in ordinary ChatGPT.</span></div> <div class="lab-cell"><b>2. Prompt</b><span>Make the task, boundaries, format and review point clear.</span></div> <div class="lab-cell"><b>3. GPT</b><span>Save the strong prompt as a GPT for recurring work.</span></div> </div> Today's flow: **write a clear prompt -> test in chat -> save the pattern as a GPT -> choose one task after the workshop**.

# A GPT is ChatGPT with fixed instructions A GPT is a version of ChatGPT where role, rules and output format are saved in advance. It is not a full agent that can act on its own. But it can give you: - the same role every time - the same rules every time - the same output format every time - the same safety boundary every time - the same review questions every time -> Today we build a simple Process Card GPT together.

# How to create a GPT <div class="prompt-builder"> <div class="prompt-step"><b>1</b><span>Open GPTs in ChatGPT and choose Create.</span></div> <div class="prompt-step"><b>2</b><span>Write what the GPT should be used for.</span></div> <div class="prompt-step"><b>3</b><span>Insert role, rules, stop rules and desired output.</span></div> <div class="prompt-step"><b>4</b><span>Add 2-3 starter questions that fit the task.</span></div> <div class="prompt-step"><b>5</b><span>Test on a harmless process description.</span></div> <div class="prompt-step"><b>6</b><span>Adjust the instruction and save only when the review point is clear.</span></div> </div> [OpenAI Help: GPTs](https://help.openai.com/en/articles/8554397-creating-a-gpt) · [OpenAI Help: Data Controls](https://help.openai.com/en/articles/7730893-how-to-use-custom-instructions)

# Template: from prompt to GPT Copy the parts that fit your process. ```text This GPT helps with: [one recurring work task] Role: You are [critical sparring partner / review assistant / structure assistant]. Rules: - Ask first if the task is unclear. - Never ask for trade secrets, patient data or non-approved confidential information. - Do not give final expert conclusions. - Always mark what the human must review. Output format: [table / checklist / outline / questions] ```

# Working with AI requires a choice of control style S1-S4 are four collaboration styles: from precise instruction to more delegation. <div class="lab-grid" style="grid-template-columns: repeat(2, 1fr); max-width: 1080px;"> <div class="lab-cell" style="border-top: 7px solid #174ea6;"> <b>S1: Directing</b> <span>You decide steps, boundaries and format. Use when the answer must be precise and easy to review.</span> <span style="margin-top: 10px; color: #111827;"><strong>Example:</strong> Create a table with 3 columns: process step, AI contribution and review point.</span> </div> <div class="lab-cell" style="border-top: 7px solid #1b7f3a;"> <b>S2: Coaching</b> <span>You give direction and ask AI to think with you, explain choices and suggest improvements.</span> <span style="margin-top: 10px; color: #111827;"><strong>Example:</strong> Suggest 3 ways to improve this process, and explain benefits and risks.</span> </div> <div class="lab-cell" style="border-top: 7px solid #f2a900;"> <b>S3: Supporting</b> <span>AI asks questions and helps you clarify options, risks and next steps.</span> <span style="margin-top: 10px; color: #111827;"><strong>Example:</strong> Ask me 5 questions before you suggest how AI can help safely.</span> </div> <div class="lab-cell" style="border-top: 7px solid #6f3cc3;"> <b>S4: Delegating</b> <span>You set the goal and let AI suggest method and draft. Use only when risk is low and the review point is clear.</span> <span style="margin-top: 10px; color: #111827;"><strong>Example:</strong> Create a first draft checklist for this sanitized task.</span> </div> </div> Often start in S2 or S3. Use S1 for precision. Use S4 only when the task is familiar and human review is clear.

# A weak prompt creates more review work than it saves ### Weak "Help me with my process." ### Strong ```text You are my critical sparring partner. I describe a workflow without confidential details. Find 3 safe AI contributions. Return a table with: process step, AI contribution, data risk and human review. ``` -> The difference is not the model. The difference is the work instruction.

# Different ways to work with generative AI <div class="lab-grid" style="grid-template-columns: repeat(2, 1fr); max-width: 1080px;"> <div class="lab-cell"><b>Ask</b><span>Use AI for questions, explanations and alternatives.</span></div> <div class="lab-cell"><b>Draft</b><span>Use AI to create an outline, table, checklist or first version.</span></div> <div class="lab-cell"><b>Review</b><span>Use AI to find gaps, assumptions, unclear claims and review needs.</span></div> <div class="lab-cell"><b>Orchestrate</b><span>Use fixed steps, roles and stop rules when the task repeats.</span></div> </div> Today's work: concrete workflows -> safe AI work patterns.

# A good AI task starts with process, not technique <div class="prompt-builder"> <div class="prompt-step"><b>Process step</b><span>What happens before, during and after the task?</span></div> <div class="prompt-step"><b>Harmless description</b><span>What can be described without confidential details?</span></div> <div class="prompt-step"><b>AI contribution</b><span>Structure, draft, review questions or checklist?</span></div> <div class="prompt-step"><b>Stop rule</b><span>What must AI not see or decide?</span></div> <div class="prompt-step"><b>Output format</b><span>Table, list, questions or next steps?</span></div> <div class="prompt-step"><b>Review point</b><span>Where does AI stop, and where does the human take over?</span></div> </div>

# Give AI the right workspace AI performs better when it receives the right information - and not too much noise. This is also called context engineering. <div class="lab-grid"> <div class="lab-cell"><b>The workspace is limited</b><span>More input is not always better. Too much noise makes it harder to hit the target.</span></div> <div class="lab-cell"><b>The prompt is only one part</b><span>Instruction, data, examples, tools and history together shape the AI answer.</span></div> <div class="lab-cell"><b>High-signal and tight</b><span>Give the smallest set of information that makes the task, boundary and output clear.</span></div> </div> [Anthropic: Effective context engineering for AI agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)

# Question technique: from short answers to deeper reflection <div class="prompt-builder"> <div class="prompt-step"><b>1. Simple answer</b><span>Ask for a short answer or list.</span></div> <div class="prompt-step"><b>2. Clarification</b><span>Ask AI to explain assumptions and missing information.</span></div> <div class="prompt-step"><b>3. Analysis</b><span>Ask for options, trade-offs, risks and evidence needs.</span></div> <div class="prompt-step"><b>4. Deep reflection</b><span>Ask AI to challenge the process and identify blind spots.</span></div> <div class="prompt-step"><b>5. Transformation</b><span>Ask AI to help redesign the workflow or turn it into a reusable pattern.</span></div> </div> The deeper the level, the more you must give AI relevant material, role, boundary and review point.

# Choose your work track Choose the track that looks most like your daily work. Use the same track through the rest of the exercises. <div class="lab-grid"> <div class="lab-cell"><b>Medical Writing</b><span>Protocol outline, document structure, review questions and checklists.</span></div> <div class="lab-cell"><b>Science / Medical</b><span>Study idea, evidence, assumptions, uncertainties and decision questions.</span></div> <div class="lab-cell"><b>Business Development</b><span>Public research on a disease area, treatment landscape and information needs.</span></div> <div class="lab-cell"><b>Statistics / programming</b><span>Analysis process, validation, error sources and explanation of code flow.</span></div> <div class="lab-cell"><b>Leadership / adoption</b><span>Baseline, micro-habit, next test and follow-up.</span></div> <div class="lab-cell"><b>Other</b><span>Choose a recurring task with fixed inputs, fixed steps and a clear review point.</span></div> </div>

# Exercise 1: Make a process AI-ready without sharing confidential data <div class="exercise-panel"> <div class="timer-display" data-seconds="600">10:00</div> <div> Open Enterprise ChatGPT. Choose a current workflow and describe it in 5 lines without names, numbers, patients, products or confidential documents. Ask ChatGPT for sparring on where AI could help safely. Output: - 3 process steps - 1 safe AI question - 1 thing AI must not see - 1 human review point </div> </div> > Process description only: no names, numbers, products or confidential documents.

# AI is strongest as a critical assistant before it is used as a writer Ask before you request text: - "Find gaps in this process description." - "Which clarifying questions should an expert ask?" - "Which parts require sources, data or expert approval?" - "What can AI help with without seeing confidential material?" > The best first output is often a review checklist, not a polished draft.

# Prompt exercise: get AI to ask the right questions Copy the prompt and insert your sanitized process description at the bottom. ```text You are my critical sparring partner. I describe a work task without confidential details. Before you suggest a solution, ask 5 clarifying questions. Ask only about process, purpose, desired answer and expert review. Do not ask for names, patients, products, numbers, document excerpts or trade secrets. My process description: [insert 5 lines here] ``` > A good prompt often starts with better questions, not with a polished answer.

# Process lab: document work and Medical Writing If your process is about documents, AI can help without seeing the document itself. <div class="lab-grid"> <div class="lab-cell"><b>Structure</b><span>"Create an outline from this sanitized task description."</span></div> <div class="lab-cell"><b>Expert review</b><span>"Find unclear assumptions and questions for expert review."</span></div> <div class="lab-cell"><b>Checklist</b><span>"Create a checklist before the document moves forward."</span></div> </div> > Describe the document type and task, not the document.

# Process lab: science, BD, data or tables If your process is about analysis, research or tables, describe the purpose before the content. <div class="lab-grid"> <div class="lab-cell"><b>Science</b><span>"Find claims, evidence needs, assumptions and questions before the next decision."</span></div> <div class="lab-cell"><b>BD research</b><span>"Create key questions for public research on a disease area."</span></div> <div class="lab-cell"><b>Validation</b><span>"Suggest which checks a human should make before output is used."</span></div> </div> > No raw data, code IP, deal details or non-public clinical details in the exercise.

# From process description to reusable GPT instruction <div class="workflow-chain"> <span>Role<br>critical sparring partner</span> <span>Input<br>sanitized process</span> <span>Rules<br>no confidential data</span> <span>Output<br>process card table</span> <span>Review<br>human decides</span> <span>Test<br>harmless case</span> </div> A process card is a simple table: process step, AI contribution, data risk and human review point. Later, the same instruction can become a GPT for recurring work.

# A GPT improves when errors are written back into the instruction We first test with a harmless process description. If the GPT answers too broadly, we improve it: ```text Always answer as a table with four columns: Process step | AI contribution | Data risk | Review point ``` Then we test again. > This is the meta-improvement: errors become better instructions.

# Agent thinking starts as a fixed workflow An agent is not magic. Think of it as AI in a fixed workflow with clear steps and stop rules. It is a work process with: - fixed steps - clear inputs - stop rules - review point - documentable output Today we work with the part you can use now: better GPTs and better work instructions.

# AI in the browser advises - agent tools can also act <div class="lab-grid" style="grid-template-columns: repeat(2, 1fr); max-width: 1080px;"> <div class="lab-cell"><b>AI in the browser</b><span>Advises, drafts and discusses inside the chat. It does not access your local files or run commands.</span></div> <div class="lab-cell"><b>Agent tools</b><span>Can read, write and run actions in a controlled environment if they are given access.</span></div> <div class="lab-cell"><b>Today</b><span>We use ChatGPT and GPTs as the practical starting point.</span></div> <div class="lab-cell"><b>Future direction</b><span>Agent tools require stronger governance, permissions and review routines.</span></div> </div>

# For coding: use a harness, not a loose prompt For resourceful coding users, the main lesson is technical: do not ask AI to "just build it". Give it a controlled engineering harness. <div class="lab-grid" style="grid-template-columns: repeat(2, 1fr);"> <div class="lab-cell"><b>State</b><span>Use git history, progress notes and a feature list so the next AI session knows what happened.</span></div> <div class="lab-cell"><b>Scope</b><span>Ask for one feature or fix at a time. Avoid one-shotting a whole application.</span></div> <div class="lab-cell"><b>Tests</b><span>Require unit checks, end-to-end checks or browser automation before marking work as done.</span></div> <div class="lab-cell"><b>Clean handoff</b><span>End with committed code, a short progress update and clear next steps.</span></div> </div> Recent engineering guidance: long-running agents need a setup that preserves state across context windows and prevents premature "done". [Anthropic Engineering: Effective harnesses for long-running agents](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents)

# Today we do not install agents - we learn the pattern What you should recognize: <div class="workflow-chain"> <span>Plan<br>what should happen?</span> <span>Input<br>what may be used?</span> <span>Execute<br>what can AI help with?</span> <span>Review<br>what must the human check?</span> <span>Learn<br>what should improve?</span> <span>Next test<br>what do we try?</span> </div> This is not an installation requirement. It is a way to think about work.

# Exercise 2: Build a fixed instruction for one recurring task <div class="exercise-panel"> <div class="timer-display" data-seconds="960">16:00</div> <div> Choose one process: - expert review of an argument - summary of meeting decisions - structuring non-confidential notes - preparing questions for expert review - translating a process into a checklist Define role, input, stop rules, output format, review point and first test. </div> </div> > The instruction must be shareable without confidential content. Confidential work belongs only in an approved environment and with human review.

# Adoption happens through micro-habits, not big transformations Choose small fixed practices: - "AI-assisted?" as a point in the meeting template - 10-minute case demo in a team meeting - shared GPT or prompt library - baseline and after-measurement on one task - review rule: AI output must never move forward without expert approval The leader's question: **How did you use AI, and where was the review point?**

# Build-Measure-Learn: how AI work improves AI improvement is a small learning loop, not a large project. <div class="lab-grid"> <div class="lab-cell"><b>Build</b><span>Try one prompt, GPT instruction or workflow on a harmless case.</span></div> <div class="lab-cell"><b>Measure</b><span>Measure whether output became shorter, clearer, faster or required less review.</span></div> <div class="lab-cell"><b>Learn</b><span>Write the error back into the prompt, GPT instruction or process card.</span></div> </div> Example: If AI answers too long, add a length requirement. If AI guesses, add source requirements and a stop rule.

# We end with three process cards from the room For each process card: | Field | Decision | |------|------------| | Process | Which workflow will we test? | | AI contribution | Where can AI help? | | Stop rule | What must AI not see? Confidential material is out. | | Review | What must the human validate? | | Baseline | How long, how many iterations or which error types today? | | Next test | What do we try first? |

# Output from today <div class="takeaway-dashboard"> <div class="takeaway-item"><b>Process card</b><span>One workflow described safely.</span></div> <div class="takeaway-item"><b>Safe prompt/GPT</b><span>One prompt or GPT instruction without confidential content.</span></div> <div class="takeaway-item"><b>Review point</b><span>A clear stop where expertise takes over.</span></div> <div class="takeaway-item"><b>Baseline</b><span>Time, iterations or error type before the AI test.</span></div> <div class="takeaway-item"><b>Next test</b><span>One task to try after the workshop.</span></div> </div>

# What do I test after the workshop? Write one sentence: > "I test AI on [process] by using [GPT/prompt] for [bounded contribution]. My baseline is [time/iterations/error type]. I evaluate after [first follow-up]." This is not a major transformation. It is one safe test that the Pharmacosmos team can follow up on.

# The next module can build on today's process cards Possible continuations: - mature one process into a shared GPT - create a team checklist for safe AI use - test process cards against actual workflows in an approved environment - build a shared prompt library - choose a supervised pilot with clear gates

# Extra tool: Perplexity for source-aware web research Perplexity is useful when the task starts with a question and needs current public sources, not only a fluent answer. <div class="lab-grid"> <div class="lab-cell"><b>What it does</b><span>Searches the web, synthesizes answers and shows source links so you can inspect the evidence.</span></div> <div class="lab-cell"><b>When to use it</b><span>Good for public landscape scans, competitor context, terminology, recent news and finding starting sources.</span></div> <div class="lab-cell"><b>How to prompt it</b><span>Ask for source types, date limits, uncertainty and a short list of claims that require human validation.</span></div> <div class="lab-cell"><b>Stop rule</b><span>Do not paste confidential material unless your organization has explicitly approved that environment and use case.</span></div> </div> Use the citations as a starting map. The expert still decides whether the source is relevant, current and suitable for the work. [Open Perplexity](https://www.perplexity.ai/) · [How Perplexity works](https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work) · [Pro Search](https://www.perplexity.ai/help-center/en/articles/10352903-what-is-pro-search)

# Extra tool: NotebookLM for sources NotebookLM is useful when the task starts with material that must be understood, connected and made usable. <div class="lab-grid"> <div class="lab-cell"><b>Sources first</b><span>Upload or add materials such as PDFs, web pages, YouTube, audio, Google Docs or Slides.</span></div> <div class="lab-cell"><b>Answers with citations</b><span>Chat with the notebook and get answers grounded in the sources with citations.</span></div> <div class="lab-cell"><b>Multiple formats</b><span>Create a briefing, study guide, mind map or audio overview from the same source set.</span></div> </div> Safe use: add only material that your internal rules allow in the specific Google/Workspace environment. [Open NotebookLM](https://notebooklm.google.com/)

# Extra demo: AI and demand management in pharma We use a concrete notebook as an example of how AI can make a source set more usable. <div class="prompt-builder"> <div class="prompt-step"><b>1</b><span>Open the notebook and see which sources it uses.</span></div> <div class="prompt-step"><b>2</b><span>Ask for key concepts, assumptions and possible decision points.</span></div> <div class="prompt-step"><b>3</b><span>Use the answer as overview - not as truth without expert review.</span></div> </div> The purpose is not to automate demand management. The purpose is to show how sources can become questions, overview and next steps. [Open notebook: AI and demand management in the pharmaceutical industry](https://notebooklm.google.com/notebook/2208448e-3476-4a50-a89a-2703b5cd5267)

# Source: Co-Intelligence is about working with AI Ethan Mollick's book *Co-Intelligence: Living and Working with AI* describes generative AI as something you learn by collaborating with it in practice - not only by reading about it. <div class="lab-grid"> <div class="lab-cell"><b>Invite AI early</b><span>Use AI for ideas, alternatives and review questions before the draft is finished.</span></div> <div class="lab-cell"><b>Be the human in the loop</b><span>Bound the task, review output and keep expert decisions with humans.</span></div> <div class="lab-cell"><b>Give a clear role</b><span>Tell AI what kind of assistant it should be, and which boundaries it works within.</span></div> </div> [Co-Intelligence on Amazon](https://www.amazon.com/Audible-Co-Intelligence-Living-Working-AI/dp/B0CNFCRJ77) · [Wharton interview about the book](https://knowledge.wharton.upenn.edu/article/co-intelligence-how-to-live-and-work-with-ai/)

# Resource: AI access has become a leadership resource Jensen Huang from NVIDIA points out that access to compute and AI consumption becomes a real constraint for knowledge work. <a class="resource-link-card" href="https://www.youtube.com/watch?v=Hg0Vus39I60" target="_blank" rel="noopener noreferrer"> <b>Open the video in a new tab</b> <span>Jensen Huang on compute, AI consumption and AI access as a leadership resource.</span> </a> Use it as perspective, not as a technical requirement for today's exercises.

<!-- section: EXTRA --> # Extra tracks if time allows Use these slides only if the topic comes up in the room. <div class="lab-grid" style="grid-template-columns: repeat(2, 1fr);"> <div class="lab-cell"><b>AI's expert boundary</b><span>Why review points and human expertise are necessary.</span></div> <div class="lab-cell"><b>Pharma case</b><span>Bayer as an example of AI for concrete document processes.</span></div> <div class="lab-cell"><b>Adoption</b><span>Why skills, governance and micro-habits matter more than inspiration.</span></div> <div class="lab-cell"><b>Agent/GPT measurement</b><span>Why the success criterion is value, quality and review load - not the number of agents.</span></div> </div>

<!-- section: EXTRA --> # Extra: AI has an expert boundary Jagged frontier means: AI can be very strong inside some tasks and significantly weaker just outside them. <div class="lab-grid"> <div class="lab-cell"><b>Inside the frontier</b><span>Studies show faster solutions and higher quality when the task fits the model.</span></div> <div class="lab-cell"><b>Outside the frontier</b><span>The result can become worse, even when the answer sounds convincing.</span></div> <div class="lab-cell"><b>Consequence</b><span>Use AI as suggestion and sparring. Keep expert review as a fixed stop.</span></div> </div> Connection to today: a good prompt also tells AI where to stop.

<!-- section: EXTRA --> # Extra: Bayer shows the point about concrete process Bayer has described using generative AI for drafts in clinical study reports. <div class="lab-grid"> <div class="lab-cell"><b>Not "AI for everything"</b><span>Choose a specific document process with known format and known review requirements.</span></div> <div class="lab-cell"><b>Medical Writing</b><span>AI can help with structure, first draft, consistency and review questions.</span></div> <div class="lab-cell"><b>Human validation</b><span>The subject owner checks claims, data, sources and regulatory suitability.</span></div> </div> Question for the room: which document type in your work has the most repetition and the clearest review point?

<!-- section: EXTRA --> # Extra: adoption requires skills and governance Deloitte's AI market signal points in the same direction: the large barrier is not only tools, but skills and governance. <div class="lab-grid"> <div class="lab-cell"><b>Skills gap</b><span>Teams must learn to frame tasks, assess output and know data boundaries.</span></div> <div class="lab-cell"><b>Training</b><span>Training is often the primary strategy for bringing AI safely into work.</span></div> <div class="lab-cell"><b>Governance</b><span>Autonomous agents require clearer rules than ordinary chat and GPTs.</span></div> </div> Connection to today: start with micro-habits, process cards and one safe test.

<!-- section: EXTRA --> # Extra: measure the value of GPTs and agents The McKinsey/EY/PwC point: the number of agents is not the success criterion. <div class="lab-grid"> <div class="lab-cell"><b>Value</b><span>Which task becomes faster, better or easier to follow up on?</span></div> <div class="lab-cell"><b>Quality</b><span>Does output become more consistent, complete or easier to review?</span></div> <div class="lab-cell"><b>Review load</b><span>Does AI save time, or does it simply move the work into checking and corrections?</span></div> </div> Good test: one GPT, one process, one baseline and one follow-up.