AI Across Functions: Legal, HR, Purchasing & More
Where value and risk show up in your day job
Listen to executive summary
Audio length ~2 min
Full lesson about ~40 min
Executive summary
~8 min to read the written sections · ~2 min to listen to this summary
This week maps AI to real functional work: Legal, HR, Purchasing, Marketing, Finance, Operations, and Sales — with different upsides, data sensitivity, and review bars in each.
You will build a short try / pilot / avoid portfolio with owners and metrics, and separate decision support from decision automation.
Outcome: a focused use-case list you can defend in budget and risk conversations, instead of tool sprawl.
Core concepts
The terms and comparisons you need for this week’s decisions — with examples by function.
Use-case portfolio (try / pilot / avoid)
A short managed list of AI uses with owners, metrics, and risk tiers — not a random zoo of tools. Try = low-risk personal productivity. Pilot = shared workflow with eval and kill criteria. Avoid = high stakes without controls.
Why this works: Without a portfolio, every demo becomes a pilot and every pilot becomes silent production. Portfolios restore executive choice.
Why it matters: Win by focusing on a few governed experiments, not every pitch.
Example: Legal: NDA first-pass (pilot). HR: JD drafts (try). Marketing: auto-publish claims (avoid).
One-page portfolio row
Weak / before: We’re exploring AI in HR.
Strong / after: Try: JD drafts in approved tool. Pilot: FAQ assist with citations. Avoid: automated reject without human path. Owner: CHRO ops. Metric: time-to-first-draft; quality sample weekly.
By function
Legal
Portfolio: try clause extract assist; avoid unsupervised external legal advice.
HR
Try JD drafts; pilot FAQ assist; avoid sole automated hire/reject.
Purchasing
Pilot RFP comparison assist with hour-saved metric — not seat count.
Marketing
Pilot variant generation with Legal gate; avoid auto-publish.
Finance
Try narrative drafts on approved numbers; avoid unreviewed external financial claims.
Operations
Pilot exception triage assist; avoid silent automation on safety-critical steps.
Sales
Try call summary copilots; avoid AI-invented commercial terms.
Risk
Force try/pilot/avoid with owner + metric + risk tier on every proposal.
Document AI
Using AI to extract, summarize, or compare text in contracts, invoices, policies, tickets, and RFPs — usually with a human still owning the final judgment.
Why this works: This is often the fastest value for Legal, HR, Purchasing, and Finance because the documents already exist and the pain is review time.
Why it matters: High fit with human review; low fit as unsupervised advice.
Example: Highlighting deviations from a playbook in a vendor contract before counsel reviews.
Purchasing
Weak / before: AI will review all supplier contracts.
Strong / after: AI flags deviations from our standard terms; category manager and Legal review before signature. Metric: review hours and miss rate on critical clauses.
Decision support vs decision automation
Support recommends options for a human. Automation acts or decides with little or no human step. Both can be valuable; they need different permissions and monitoring.
Why this works: Calling automation “just a recommendation engine” hides accountability. Naming the mode forces the right controls.
Why it matters: Automate low-risk steps first. Keep humans on rights, large money, and safety.
Example: Support: suggest a discount band. Automation: auto-send the discount to the customer.
vs Copilot vs agent: Same spectrum: help me decide vs act on my behalf across systems.
Deep dive lesson
~14 min readDeep dive: Where AI helps your function (and where it does not)
Learning objective. Build a three-item use-case portfolio with owners, metrics, and risk tiers — instead of collecting random tools.
Context. Literacy becomes useful when it maps to work. This lesson turns “interesting AI” into a try / pilot / avoid portfolio you can defend in a budget meeting.
1. From curiosity to a portfolio
Legal often gains from first-pass document review against a playbook. HR from job description drafts and FAQ assist — never sole hiring decisions. Purchasing from RFP comparison support. Marketing from variants and localization with brand/legal gates. Operations from SOP assistants grounded in approved content. Finance from analysis copilots with strict data classes.
The anti-pattern is tool sprawl: five overlapping copilots, no owner, no metric. Prefer a short portfolio: try / pilot / avoid — each with a named owner, a success metric, and a risk tier. Decision support recommends; automation acts. Choose deliberately which bucket you are in.
2. Red lines worth writing down
People decisions (hire, fire, promote, discipline) need heightened care and often legal involvement. External claims (pricing, product, regulated wording) need verification. Material non-public information and sensitive personal data do not belong in unapproved tools. Write the red lines once for your team; update when tools change. A one-page red line list beats a 40-page policy nobody opens under deadline pressure.
3. Metrics that mean something
“Seats used” is not value. Prefer cycle time for a defined task, rework rate, error rate on high-stakes fields, customer or employee satisfaction, or hours returned to judgment work. Define the metric before the pilot starts so nobody rewrites success after the fact.
Method: Try / Pilot / Avoid canvas
Try
Low risk, reversible, personal productivity — approved tools only.
Pilot
Shared workflow with eval, owner, kill criteria, and end date.
Avoid
High stakes without controls (people decisions, unrestricted external claims, Red data).
Metric
One number you will look at on the go/no-go date.
Worked example: Marketing’s “AI campaign engine”
Situation
Agency proposes generative ads plus automated audience selection. CMO wants speed for a product launch.
How an executive thinks it through
- Split analytics (who to target) from generative (what to say).
- Claims and regulated wording need Legal gate.
- Brand voice needs human creative lead on final.
- Audience models can encode bias — check sensitive attributes.
Decision / what to say
Approve pilot for draft variants only; targeting stays on existing platform with current controls; Legal review for claims; kill criteria: brand incidents or claim errors above threshold.
Apply in your function
Legal
Pilot: clause extraction vs playbook. Avoid: unsupervised legal advice to clients.
HR
Try: JD drafts. Avoid: sole automated reject without human review path.
Purchasing
Pilot: RFP response comparison assist. Metric: review hours, not “AI used.”
Marketing
Pilot: variant generation. Avoid: auto-publish claims without Legal.
Common mistakes
- Starting with the tool instead of the workflow metric.
- No avoid list — everything becomes a pilot.
- Owners who are “the AI team” rather than the function that lives with outcomes.
Practice (15 minutes)
- List three AI uses: one to try, one to pilot, one to avoid in your function.
- Add owner, metric, and risk tier (low/medium/high) for try and pilot.
- Share with one peer for challenge: what did you miss?
Check your understanding
Key takeaways
- Same technology, different stakes by function and decision type.
- Document AI and scoring AI dominate early enterprise value.
- Cross-functional coordination prevents duplicate tools and conflicting policies.
- Pick a small portfolio: try, measure, govern.
Practical applications
Prioritize clause extraction and research assist over fully automated advice.
Separate internal productivity tools from hiring decisions that affect candidates’ rights.
Use AI for market scans and contract metadata; keep award decisions accountable.
Pilot predictive maintenance or quality alerts with clear false-alarm handling.
Portfolio: try brief variants; pilot localization with brand gate; avoid auto-publish of product claims.
Library materials for this week
Extend this week’s decisions with briefs, cases, and primary sources that matter for your next meeting.
AI for Legal Leaders
Where legal teams gain speed with AI (research assist, clause extraction, first-pass review) and where professional responsibility draws hard lines.
Klarna customer service AI — scale, then rebalance (2024–2025)
Published 18 November 2025
Klarna’s 2024 AI assistant was widely reported as handling work equivalent to hundreds of full-time agents. In 2025 the company publicly rebalanced toward more human support after quality concerns — while still claiming large AI productivity. The live lesson: cost-first automation can overshoot customer trust.
Optional focus hour
After the core lesson (~40 min), spend about 45 minutes on one long-form source matched to this week’s level.
PwC 29th Global CEO Survey 2026 — full PDF
PwC
Published 19 January 2026
Why this week
Use-case week: CEO survey language helps you pitch or challenge AI work in frames ELT already uses.
What to take away
How CEOs frame AI vs growth and trust; one statistic for your leadership update.
Check your understanding
Week 4 · 3 short questions · no grades shared outside this device
Select an answer for each question.
Hands-on
~15 minStart / Scale / Stop card
Shared vocabulary prevents shadow IT and blind spots.
- Draw three columns: Start, Scale, Stop/Restrict.
- Place two real or proposed AI uses from your function into the grid.
- For each, write the decision owner and the worst realistic failure.
- Share with a peer in another function — note overlaps or conflicts.
Reflection
- Which other function’s AI tool already affects your team’s work?
- Where might two departments automate the same process with conflicting rules?