AI Kick-start — R&D Lederteam

# Copilot leadership kick-start for the R&D team
## Goal: calibration, skills, and next steps
Pharmacosmos — Holbæk
18 June 2026 · 09.00–13.00

<div class="lab-grid">
  <div class="lab-cell"><b>1 focus area</b><span>Where is Microsoft Copilot practically relevant for R&D leadership right now?</span></div>
  <div class="lab-cell"><b>1 work pattern</b><span>What does a safe skill + human review look like?</span></div>
  <div class="lab-cell"><b>1 decision</b><span>What does the leadership team prioritize after today?</span></div>
</div>

> Skill = reusable Copilot instruction for a recurring task. Agent = scheduled workflow that runs that skill on a cadence.

<div style="margin-top: .85rem; padding-top: .55rem; border-top: 1px solid rgba(25, 52, 73, .14); font-size: .48em; line-height: 1.32; color: #536473; font-style: italic;">
  Questions or follow-up after the session?<br>
  Torben Madsen · <a href="mailto:tm@kvalitor.dk">tm@kvalitor.dk</a> · Tel. 9390 8357
</div>

# What is an LLM?

<p style="margin:0 0 .35rem; font-size:22px; line-height:1.3; max-width:920px;"><b>LLM</b> = Large Language Model — the engine behind Copilot, ChatGPT and similar tools.</p>

<div class="six-terms-split">
  <button type="button" class="slide-image-expand" data-fullsrc="assets/horse-analogy.png" aria-label="Expand generative AI horse analogy diagram">
    <img src="assets/horse-analogy.png" alt="Horse analogy diagram explaining generative AI concepts for leaders">
    <span class="slide-image-expand-hint">Click diagram to enlarge</span>
  </button>

<div class="prompt-builder compact six-terms-vocab">
    <div class="prompt-step"><b>What it is</b><span>Predicts the next word from context — trained on books, articles, code and the web.</span></div>
    <div class="prompt-step"><b>What it is not</b><span>Not search, not a database, not a thinker — it generates plausible text.</span></div>
    <div class="prompt-step"><b>Why it goes wrong</b><span>Training cut-off date + hallucination — fluent answers that may be false.</span></div>
    <div class="prompt-step"><b>For R&D leaders</b><span>Like a well-read colleague who left two years ago — useful drafts, mandatory review.</span></div>
  </div>
</div>

<p class="deck-footnote">The model does not know it is wrong. Human review is the control mechanism.</p>

# Core vocabulary: AI terms

<p style="margin:0 0 .35rem; font-size:22px; line-height:1.3; max-width:920px;">Six terms you will hear today — and use after the workshop.</p>

<div class="prompt-builder compact six-terms-vocab">
  <div class="prompt-step"><b>Prompt</b><span>Task, role, context, boundaries and output format.</span></div>
  <div class="prompt-step"><b>Skill</b><span>Reusable Copilot instruction — saved, shared, versioned in OneDrive.</span></div>
  <div class="prompt-step"><b>Schedule</b><span>Skill on a recurring cadence — e.g. daily mail triage at 07.00.</span></div>
  <div class="prompt-step"><b>Co-work</b><span>Copilot Frontier — your files and folders inside the M365 boundary.</span></div>
  <div class="prompt-step"><b>Agent</b><span>Saved instruction + workflow steps; may include schedules and review gates.</span></div>
  <div class="prompt-step"><b>Hallucination</b><span>Fluent but wrong answers — human review is mandatory on regulated work.</span></div>
</div>

<p class="deck-footnote">Legacy note: ChatGPT GPTs map roughly to Copilot skills — same idea, different platform.</p>

# Generative AI helps — it does not know what is true

<p style="margin:0 0 .6rem; font-size:24px; line-height:1.35; max-width:920px;"><b>Main message:</b> AI is strong at drafting and structuring — weak at knowing whether the output is correct. Your review is the control mechanism.</p>

<div class="lab-grid">
  <div class="lab-cell"><b>What it can do</b><span>Suggest text, questions, summaries, checklists and draft workflows from the material you provide.</span></div>
  <div class="lab-cell"><b>What it cannot do</b><span>Verify facts, judge regulated content, or replace subject-matter ownership — especially on thin or missing input.</span></div>
</div>

**Work pattern:** task → material and rules → boundaries → desired answer → **human review**

> Models optimize for a useful answer — not necessarily the most critical one. Ground the task in your material. Ask for critique when you need it. Common sense and human review are not optional.

# Mini exercise: where may this input be used?

Place each example in one of three categories — same gates as the previous slide:

- **OK in the room** — sanitized process description only; safe for today's exercises
- **Copilot / M365 only** — confidential R&D work content per Pharmacosmos rules; human review required
- **Do not use in AI** — stop; trade secrets, personal data, deals, formulas, non-public clinical details

**Examples to classify**

1. Generic process description ("we run a literature scan before internal review")
2. Protocol outline without product name
3. Patient or trial-subject data
4. Public research on a disease area
5. Supplier name in a contract context
6. Product formula, recipe or manufacturing know-how

Discuss in pairs, then compare with the room. Leadership question: *which gate would your team default to — and who owns the review point?*

</div>
</div>

> Today's exercises use only sanitized descriptions. In real work, follow Pharmacosmos AI rules and your M365 boundary.

# When frontier AI access can disappear overnight

In June 2025, Anthropic disabled **Fable 5** and **Mythos 5** for users outside the United States — following a US export-control directive. Organizations that relied on browser access to a frontier model lost capability **with no migration window**.

<div class="lab-grid">
  <div class="lab-cell"><b>What happened</b><span>A vendor suspended commercial frontier models for non-US users in hours — not after a planned transition.</span></div>
  <div class="lab-cell"><b>Why it matters here</b><span>R&D work must not depend on a single external model session for critical process steps.</span></div>
  <div class="lab-cell"><b>Leadership takeaway</b><span>Approved tools, clear data boundaries, and auditable artifacts — before you scale usage.</span></div>
</div>

<a class="resource-link-card" href="https://www.anthropic.com/news/fable-mythos-access" target="_blank" rel="noopener noreferrer">
  <b>When a frontier AI model was shut down for non-US users (new tab)</b>
  <span>Anthropic's statement on the Fable 5 and Mythos 5 access suspension — primary source for the room.</span>
</a>

# Keep your data where you control it

<p style="margin:0 0 .65rem; font-size:24px; line-height:1.35; max-width:920px;">Work files, outputs and institutional knowledge stay in <b>your</b> environment — not in a model provider's session or logs.</p>

<div class="prompt-compare">
  <div class="prompt-weak"><b>Browser-only / external model</b><span>Data leaves your boundary every session. The provider controls access — and can remove it overnight.</span></div>
  <div class="prompt-strong"><b>Co-work Frontier + OneDrive</b><span>Input and output stay local in OneDrive. Swap the model layer if you need to; your data does not move with it.</span></div>
</div>

<div class="lab-grid" style="margin-top: .6rem;">
  <div class="lab-cell"><b>R&D angle</b><span>Clinical data, process knowledge and institutional memory remain with Pharmacosmos — auditable and under your control.</span></div>
  <div class="lab-cell"><b>Architecture angle</b><span>OneDrive + SharePoint as the local data layer — retainable, reviewable, and replaceable without migrating work history.</span></div>
</div>

# Learn to recognize — inside vs. outside the frontier — Harvard · BCG study

<div class="prompt-compare">
  <div class="prompt-strong"><b>Inside frontier</b><span>~25% faster, >40% higher quality on suited consulting tasks (GPT-4 access vs. control).</span></div>
  <div class="prompt-weak"><b>Outside frontier</b><span>Performance <em>dropped</em> when consultants used AI on tasks beyond current capability — even strong performers.</span></div>
</div>

<div class="prompt-builder" style="margin-top:0.75rem;">
  <div class="prompt-step"><b>Reflect</b><span>Which two R&D tasks are likely inside today's frontier for your team?</span></div>
  <div class="prompt-step"><b>Reflect</b><span>Which task must stay human-led — and how do you spot misuse in a Monday meeting?</span></div>
  <div class="prompt-step"><b>Download</b><span>Full study PDF + reflection prompts — open in a new tab, then return to the deck.</span></div>
</div>

<a class="resource-link-card" href="assets/pdfs/jagged-technological-frontier.pdf" target="_blank" rel="noopener noreferrer">
  <b>Download study PDF (new tab)</b>
  <span>Full HBS / BCG working paper — analyze with your own notes.</span>
</a>

<a class="resource-link-card youtube-link" href="https://www.youtube.com/watch?v=ZhWlGqkLC_M&t=5s" target="_blank" rel="noopener noreferrer">
  <b>How to Use Generative AI in Consulting</b>
  <span>Opens in YouTube — new tab (starts at 0:05).</span>
</a>

# AI can intensify work — not only save time

<div class="lab-grid">
  <div class="lab-cell"><b>Task expansion</b><span>People take on wider scope because AI makes it feel accessible.</span></div>
  <div class="lab-cell"><b>Review load</b><span>Experts spend more time correcting AI-assisted output from others.</span></div>
  <div class="lab-cell"><b>Workload creep</b><span>More speed and multitasking — not necessarily less busyness.</span></div>
  <div class="lab-cell"><b>Leadership</b><span>Plan for an intentional AI practice — pauses, sequencing, human grounding.</span></div>
</div>

<a class="resource-link-card" href="articles/ai-intensifies-work.html">
  <b>Read: HBR article + manager reflection questions</b>
  <span>Download PDF, then answer: what should a manager of a team using AI be aware of?</span>
</a>

<a class="resource-link-card" href="assets/pdfs/ai-intensifies-work.pdf">
  <b>Download article PDF</b>
  <span>Ranganathan & Ye — <em>AI Doesn't Reduce Work—It Intensifies It</em> (2026).</span>
</a>

# Knowledge work industrializes — from hand loom to automatic loom

<button type="button" class="slide-image-expand" data-fullsrc="assets/toyota-loom.jpg" aria-label="Expand Toyota automatic loom photo">
  <img src="assets/toyota-loom.jpg" alt="Sakichi Toyoda automatic loom on display at a Toyota museum in Japan">
  <span class="slide-image-expand-hint">Click photo to enlarge</span>
</button>

<div class="lab-grid">
  <div class="lab-cell"><b>1890s hand loom</b><span>One worker, every motion manual — expertise in the fingers.</span></div>
  <div class="lab-cell"><b>1924 Type G</b><span>Automatic shuttle change, auto-stop on broken thread — one operator oversees the system.</span></div>
  <div class="lab-cell"><b>R&D today</b><span>Skills and scheduled agents in the Copilot ecosystem handle bounded steps; leaders design the workflow and monitor output.</span></div>
</div>

<a class="resource-link-card" href="articles/knowledge-work-industrialization.html">
  <b>Read: industrialization of knowledge work (Japan museum photo)</b>
  <span>Sakichi Toyoda's loom lesson — supervise the machine, don't re-weave every thread.</span>
</a>

# Pharmacosmos Copilot stack — from chat to scheduled agents

<div class="workflow-chain">
  <span>Copilot<br>daily assistant</span>
  <span>Co-work<br>Frontier</span>
  <span>Skills<br>saved instructions</span>
  <span>Schedules<br>recurring runs</span>
  <span>Agents<br>workflow + gates</span>
</div>

<div class="lab-grid" style="margin-top: .6rem;">
  <div class="lab-cell"><b>Copilot</b><span>Chat in M365 — test a task, refine the prompt, stay inside approved boundaries.</span></div>
  <div class="lab-cell"><b>Co-work Frontier</b><span>File-aware work in OneDrive — skills read and write traceable artifacts.</span></div>
  <div class="lab-cell"><b>Skill</b><span>Reusable instruction saved in a OneDrive folder — invoke with a slash command.</span></div>
  <div class="lab-cell"><b>Schedule</b><span>Run a skill on a cadence — morning mail triage, weekly prep, and similar.</span></div>
  <div class="lab-cell"><b>Agent</b><span>Multi-step workflow with stop points — still with a human owner, not autonomous scale.</span></div>
</div>

Start with chat. Graduate to a **skill** when the prompt is stable. Add a **schedule** when the win is daily. An **agent** is the next step — not a bot without leadership.

# Inside a Copilot skill — what Thomas built yesterday

<p style="margin:0 0 .55rem; font-size:22px; line-height:1.35; max-width:920px;"><b>Core vocabulary</b> — plain words before the live demo:</p>

<div class="prompt-builder">
  <div class="prompt-step"><b>Skill folder</b><span>One named folder in OneDrive — e.g. <code>daily-mail-triage/</code>. The folder <em>is</em> the skill.</span></div>
  <div class="prompt-step"><b>Markdown (.md)</b><span>Simple text with headings and bullets — like a Word outline or SOP. <em>Not</em> code; just readable instructions for the AI.</span></div>
  <div class="prompt-step"><b>Instruction file</b><span>The markdown file inside the folder — the recipe: what to read, what to do, what to produce.</span></div>
  <div class="prompt-step"><b>Slash command</b><span>Type <code>/daily-mail-triage</code> in Copilot to run that skill.</span></div>
  <div class="prompt-step"><b>Input / output</b><span><code>input/</code> = files it may read · <code>output/</code> = where results are saved.</span></div>
  <div class="prompt-step"><b>Artifact</b><span>The saved result in <code>output/</code> — timestamped, in OneDrive, ready for your review.</span></div>
  <div class="prompt-step"><b>Governance</b><span>Who owns the skill, who may edit it, and when it is retired — especially on shared folders.</span></div>
</div>

<div class="workflow-chain" style="margin-top: .55rem;">
  <span>Folder</span>
  <span>Instruction<br>(markdown)</span>
  <span>/slash</span>
  <span>Read input</span>
  <span>Save output</span>
  <span>You review</span>
</div>

# Skills governance — who owns what in the library?

<div class="lab-grid">
  <div class="lab-cell"><b>Default skills</b><span>Approved, named, and maintained — available to everyone. Requires a named owner and a review cycle.</span></div>
  <div class="lab-cell"><b>Naming convention</b><span>Verb-noun format: <code>daily-mail-triage</code>, <code>meeting-summary</code>, <code>doc-review</code>. Consistent naming = findable and shareable.</span></div>
  <div class="lab-cell"><b>Owner</b><span>Every skill must have a named owner responsible for correctness, updates and security boundary. No owner = not default.</span></div>
  <div class="lab-cell"><b>Update cycle</b><span>Skills decay as processes change. Review on a cadence — quarterly is a starting point.</span></div>
  <div class="lab-cell"><b>Retirement</b><span>A skill with no update or no active owner after 90 days is archived or deleted. Stale skills create risk.</span></div>
</div>

Leadership question: **Who decides what becomes a default? Who owns the library?** Architecture guidance covers sharing, security boundary and naming at scale.

# Scout team — serial sprints, same team, compounding learning

<div class="workflow-chain">
  <span>Process<br>hypothesis</span>
  <span>Baseline<br>measure</span>
  <span>Skill<br>pilot</span>
  <span>Gain<br>verify</span>
  <span>KPI<br>gate</span>
  <span>Rollout<br>decision</span>
</div>

<div class="lab-grid" style="margin-top: .6rem;">
  <div class="lab-cell"><b>Same team, serial</b><span>Run one sprint at a time with the same small team. Parallel sprints split the learning — serial sprints compound it.</span></div>
  <div class="lab-cell"><b>Compounding</b><span>Sprint 3 is faster than sprint 1 because the team has the pattern. Each failure gets written back into the skill.</span></div>
  <div class="lab-cell"><b>KPI gate</b><span>Quality, time or cost must improve ≥10% before the process change is approved for rollout. Enthusiasm is not a gate.</span></div>
  <div class="lab-cell"><b>Leadership role</b><span>Cascade decisions from scout team findings to the full R&D team. Scout team proves the gain — leadership approves the change.</span></div>
</div>

# GenAI implementation journey — four phases to sustainable value

<button type="button" class="slide-image-expand" data-fullsrc="assets/genai-implementation-phases.png" aria-label="Expand GenAI implementation phases diagram">
  <img src="assets/genai-implementation-phases.png" alt="Four phases from understand and prioritize to standardize and embed">
  <span class="slide-image-expand-hint">Click diagram to enlarge</span>
</button>

<p style="margin:0.25rem 0 0.5rem; font-size:20px; line-height:1.35; max-width:920px;">Today's kick-start is <b>Phase 1</b> — with Copilot skills and KPI thresholds for Phase 2: understand, explore, prioritize. The leadership output is one safe pilot that moves you into <b>Phase 2</b> — not a jump to scale.</p>

<div class="workflow-chain" style="grid-template-columns: repeat(4, 1fr);">
  <span>1<br>Prioritize</span>
  <span>2<br>Pilot</span>
  <span>3<br>Scale</span>
  <span>4<br>Standardize</span>
</div>

# KPI threshold card

<p style="margin:0 0 .4rem; font-size:22px; line-height:1.3; max-width:920px;">Four lenses before you change a process: <b>cost</b>, <b>time</b>, <b>quality</b>, <b>job satisfaction</b>.</p>

<button type="button" class="slide-image-expand" data-fullsrc="assets/ai-metrics-four.png" aria-label="Expand four AI metrics diagram">
  <img src="assets/ai-metrics-four.png" alt="Four AI metrics for leaders: cost, time, quality and job satisfaction">
  <span class="slide-image-expand-hint">Click diagram to enlarge</span>
</button>

<div class="prompt-compare" style="margin-top:.5rem;">
  <div class="prompt-strong"><b>Threshold</b><span>≥10% on quality, time <em>or</em> cost after a bounded pilot — review gates intact.</span></div>
  <div class="prompt-weak"><b>Gates</b><span>Pilot → baseline data → leadership decision.</span></div>
</div>

> Exploratory wins need no threshold. <b>Process change</b> needs a number leadership can follow up on.