What AI Really Is (and Isn’t)
A working mental model for boardrooms, contracts, and team decisions
Listen to executive summary
Audio length ~2 min
Full lesson about ~35 min
Executive summary
~8 min to read the written sections · ~2 min to listen to this summary
This week builds the operating definition of AI you will use for the rest of the course: software that finds patterns in data to predict, classify, recommend, or generate outputs — not a colleague and not magic.
You will sort workplace tools into analytics vs generative (different failure modes, different controls), and apply four leadership questions to any proposal: What decision changes? What data is required? What happens when it is wrong? Who is accountable?
Outcome: board- and vendor-ready language so demos no longer set the agenda for Legal, HR, Purchasing, Marketing, Finance, or Operations.
Core concepts
The terms and comparisons you need for this week’s decisions — with examples by function.
Artificial intelligence (AI)
At work, AI is software that finds patterns in data and uses those patterns to predict, classify, recommend, or generate outputs. It is not a person, not conscious, and not automatically trustworthy. Think of it as pattern machinery that can be very fast and very wrong.
Why this works: If you treat AI as magic, you approve demos without controls. If you dismiss it entirely, staff adopt consumer tools in the shadows. A clear definition lets you govern it like any other operational capability.
Why it matters: You already own budgets, risk, and people outcomes. Extend that ownership: decide where pattern machinery helps, where it must stay supervised, and where it should not be used.
Example: Spam filters, fraud alerts, demand forecasts, ChatGPT drafting a memo, and Claude summarizing a contract are all AI — different jobs, same idea: patterns → outputs.
What to say in a meeting
Weak / before: We need AI to transform the department.
Strong / after: We will pilot AI to draft first-pass contract extracts against our playbook, with counsel review before any advice goes out. Success = hours saved without quality drop.
Replace vague transformation language with a decision, a control, and a metric.
vs Human expertise: AI scales pattern matching; humans still own values, accountability, and unusual judgment calls.
By function
Legal
Clause extraction and playbook matching on NDAs — analytics + generative hybrid, not “the AI understood the law.”
HR
Job description drafts (generative) vs résumé ranking (analytics) — different risk and review bars.
Purchasing
Vendor “AI suite” demos: force the sort analytics vs generative before scoring the RFP.
Marketing
Audience scoring (analytics) vs ad copy variants (generative) — claims review only applies to the second.
Finance
Forecast models score demand; generative tools draft commentary — never invent numbers from chat.
Operations
Exception queues ranked by risk score (analytics); SOP copilots draft steps (generative).
Sales
Next-best-action scores vs AI-drafted proposals — CRM truth for pricing, not the model.
Risk
Inventory every AI touchpoint as analytics, generative, or both before one blanket control set.
Analytics AI vs generative AI
Analytics AI scores, ranks, forecasts, or flags (credit risk, churn, fraud). Generative AI drafts new content (emails, summaries, images, code). Many products mix both, but the failure modes differ — and so should your controls.
Why this works: One vague “AI policy” under-controls high-stakes scoring and over-controls harmless drafting. Sorting the bucket first tells you what to test and who must review.
Why it matters: In every vendor pitch, force the sort out loud: analytics, generative, or both? Then ask what happens when it is wrong.
Example: Analytics: a model scores supplier risk. Generative: Claude or Grok drafts a customer email. Hybrid: extract clauses (classify) then draft a redline summary (generate).
Same word “AI,” different risk
Weak / before: Approve the AI tool — Marketing and HR both want it.
Strong / after: Marketing’s generative copy tool needs claim review. HR’s analytics ranking tool needs fairness documentation and human override. Two packages, two gates.
vs Each other: Wrong score vs invented paragraph — different detection and accountability.
Models from AI labs
Major labs train the engines behind familiar apps: OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), SpaceXAI (Grok), and others. Enterprise products often wrap those engines with admin controls and contracts.
Why this works: Staff say “we already use AI at home.” You need to translate: which lab, which data path, which review step is missing at work?
Why it matters: Ask which lab and model tier you are buying — vendors rebrand wrappers constantly.
Example: ChatGPT vs Claude vs Gemini vs Grok look similar on screen; enterprise agreements and data handling often do not.
vs Your internal tools: Consumer chat is not the same as an enterprise deployment with logging and retention controls.
Copilot (not autopilot)
A copilot assists a human who stays accountable for the decision or outgoing message. An agent or autopilot acts across systems with less step-by-step human control.
Why this works: Most early enterprise failures come from skipping the human ownership step while still calling the work “assisted.”
Why it matters: Default functional work to copilot until permissions, logs, and kill switches exist for automation.
Example: Copilot drafts a board email; you edit and send. An agent might also update CRM and schedule meetings without you.
Permission language
Weak / before: Let the AI handle refunds end-to-end.
Strong / after: Phase 1: AI recommends refund decisions for human approval. Phase 2: auto-execute only under hard limits, with logs and a kill switch.
vs Agent: Copilot helps you act; agent acts across systems. Agents need automation-grade controls.
Deep dive lesson
~16 min readDeep dive: Build a usable mental model of AI
Learning objective. Explain what AI is (and is not) in one minute, sort any work tool into analytics vs generative, and run four leadership questions before you approve a pilot.
Context. Your advantage is judgment and ownership of outcomes — not model training. This lesson gives you a mental model that stops demos from running your calendar, with examples from Legal, HR, Purchasing, Marketing, and Ops.
1. What AI actually is at work
Marketing stories often promise machines that think, or wholesale department replacement next quarter. A sharper operating definition is narrower and more useful. Artificial intelligence in business is software that finds patterns in data and uses those patterns to predict, classify, recommend, or generate outputs.
It does not understand your company the way a seasoned colleague does. It does not care about your values, your brand promise, or your legal duty of care. It scales pattern-matching — sometimes brilliantly, sometimes confidently wrong. That single sentence should change how you listen in meetings. When someone says “the AI decided,” translate: “software produced a score or a draft based on patterns; a human still owns the outcome.”
Two traps open immediately. Trap one: treat AI as magic that will “figure it out.” Trap two: dismiss everything as hype so completely that staff quietly adopt ChatGPT, Claude, Gemini, or Grok on personal accounts with no guardrails. You already own decisions, risk, and people. Extend that ownership to where pattern machinery helps, where it must stay supervised, and where it should not be used at all.
2. Analytics vs generative — why one policy is not enough
Most workplace tools fall into two buckets. Analytics AI scores, forecasts, ranks, or flags — credit risk, demand, fraud, churn, résumé ranking. Generative AI drafts new content — text, images, code, meeting notes, marketing variants. The failure modes differ. Analytics fails as a wrong score or a biased ranking. Generative fails as a fluent but invented paragraph, citation, price, or case name.
If Legal, HR, Purchasing, and Marketing all write one vague “AI policy,” you will under-control one bucket and over-control the other. Example: a spellcheck-style copilot for internal drafts needs light rules. A hiring ranker or a claims-generating marketing tool needs heavier review. Same word “AI,” different stakes.
In every pitch meeting, force a sort out loud: “Is this primarily analytics, generative, or both?” Then ask four questions that matter more than the algorithm name: What decision changes if we use this? What data does it need? What happens when it is wrong? Who is accountable in our org chart? If those answers are fuzzy, the project is not ready — regardless of how polished the demo is.
3. Labs, products, and “we already use AI at home”
Staff already use consumer products: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Grok (SpaceXAI). Those are product experiences built on models from different labs. Enterprise tools often wrap the same underlying engines with admin controls, data agreements, and retention settings.
When someone says “we already use AI at home,” translate carefully: which lab, which data path, which review step is missing at work? Consumer tools may train on inputs, store chats, or lack enterprise DLP. Your legal and privacy teams care about those differences even when the screen looks similar.
Copilot means assist-and-stay-accountable: the human edits and owns the send. Agent or autopilot means the system acts across tools — update CRM, issue refund, schedule vendor outreach. Default most functional work to copilot until permissions, logs, and kill switches exist. That single distinction prevents many early disasters.
In the meeting
A short exchange you can reuse when the conversation gets vague.
Vendor
Our AI reviews every contract in seconds so Legal doesn’t have to.
You
Is that analytics scoring, generative drafting, or both? And who still owns signature risk?
Vendor
It flags risky clauses automatically — 98% accuracy.
You
Accuracy of what, on whose NDAs, and what happens on the 2% when a liability clause is missed?
Method: The four questions (use in every AI meeting)
Decision
Name the decision or output that changes — not “efficiency” in the abstract.
Data
List what goes in (contracts, HR files, customer lists) and where it is stored or trained.
Failure
Write the worst plausible wrong answer for your function in one sentence.
Owner
Name the human who is accountable if that wrong answer ships — by role, not “the AI.”
Worked example: The “AI for contracts” pitch
Situation
A vendor demos a tool that “reviews all NDAs in seconds.” Your GC and a Purchasing lead are impressed. They ask you — as a function head — for a reaction in the room.
How an executive thinks it through
- Bucket: mostly generative + classification (extract clauses, flag deviations) — not pure magic understanding of law.
- Decision changed: first-pass review speed, not final legal judgment or signature authority.
- Data: contracts may include confidential commercial terms — training rights and storage matter.
- When wrong: missed liability clause or false “clean” flag is worse than a slow human pass.
- Accountable: still counsel for advice; business owner for accepting residual risk.
Decision / what to say
Say this: “Useful as a copilot for first-pass against our playbook with human review. Before pilot: data rights, sample eval on our NDAs, clear fail criteria. Not autopilot for signature.”
Apply in your function
Legal
Inventory AI that touches contracts or advice. Label each analytics/generative. Require human review for advice.
HR
List screening, survey, and writing tools. Mark any that influence hire/promote/fire as high stakes — not “productivity.”
Purchasing
Add the four questions to your RFP template for AI vendors. Reject “98% accuracy” without definition.
Marketing
Separate audience analytics tools from generative content tools — different claim-review bars.
Operations / Finance
Map forecasting and exception tools: who overrides a score when reality diverges?
Common mistakes
- Equating a polished demo with readiness for production.
- Assuming consumer ChatGPT/Claude rules apply to enterprise tools (they often do not).
- Writing one AI policy that never distinguishes analytics vs generative risk.
- Letting “the tech team owns AI” remove functional accountability for outcomes.
Practice (10 minutes) — do this before next week
- List three tools your team already touches that might use AI (email suggest, screening, forecasting, chat).
- Label each analytics, generative, or both.
- For each, write the worst plausible error for your function in one sentence.
- Circle any tool where nobody owns verification today — that is your first governance gap.
- Bring that list to your next staff meeting and assign owners for the circled items.
Check your understanding
Key takeaways
- AI finds patterns and predicts or generates — it does not think or feel like a human.
- Separate hype from two practical buckets: analytics AI and generative AI.
- Your role is judgment, governance, and prioritization — not coding.
- Every AI system has failure modes; leadership means planning for them.
Practical applications
When counsel says 'AI contract review,' ask what patterns it was trained on and what still needs human review.
Treat résumé-screening tools as assistants with bias risk, not as neutral judges of talent.
Demand plain answers on data use, accuracy claims, and exit terms from AI vendors.
Map which decisions are pattern-based (good AI fit) vs. one-off judgment calls (human-led).
Separate campaign tools into analytics (who to target) vs generative (what to say) — different review bars for each.
Library materials for this week
Extend this week’s decisions with briefs, cases, and primary sources that matter for your next meeting.
AI 101 for Business Leaders
A clear map of artificial intelligence: what it is, the two buckets most businesses use (analytics vs generative), and the leadership questions that matter more than algorithms. Built for experienced executives who need board-ready fluency, not a computer science degree.
DBS Bank — AI value at enterprise scale
Published 2025 results coverage (Business Times)
Singapore’s DBS Bank has publicly reported large economic value from AI programs — including Business Times coverage that the bank unlocked about S$1 billion in AI-related economic value in 2025 — after years of digital transformation and governed AI use cases across the franchise.
Optional focus hour
After the core lesson (~35 min), spend about 55 minutes on one long-form source matched to this week’s level.
Stanford HAI — AI Index Report 2026 (full PDF)
Stanford Institute for Human-Centered AI
Published 2026 edition
Why this week
Week 1 builds a plain mental model of AI. The Index PDF gives non-vendor numbers so hype does not define your baseline.
What to take away
Three facts for a board pack; one chart on adoption or capability; one methodology caveat.
Check your understanding
Week 1 · 3 short questions · no grades shared outside this device
Select an answer for each question.
Hands-on
~10 minThe 10-minute AI inventory
Most organizations already use AI without a shared map. Awareness is the first control.
- List every tool your team uses that claims 'AI,' 'smart,' or 'automated recommendations.'
- For each, write one sentence: what decision does it influence?
- Mark each as High / Medium / Low impact on people, money, or compliance.
- Circle the one you understand least — that becomes your first deep-dive next week.
Reflection
- Where do people in your team already trust a system’s recommendation without checking it?
- What decision in your world must never be fully automated — and why?