Operations & Supply Chain

purchase request → RFQ/quote comparison → contract & PO → receipt/invoice match → supplier scorecard

• Automate
  Intake & policy check — AI validates purchase requests against budget, approved-vendor, and policy rules, and routes for approval · rules-based, reversible before money moves, high repetition
• Draft
  RFQ drafting & quote comparison — AI drafts the RFQ from the requirement and normalizes returned quotes into a comparison matrix · unit conversions and hidden-fee extraction are verifiable, but apples-to-apples judgment calls need the buyer
• Assist
  Negotiation & award — the buyer negotiates terms and picks the supplier, with AI summarizing leverage points and market benchmarks · relationship- and leverage-driven, high stakes, and the counterparty is also human
• Draft
  Contract review — AI redlines against the clause playbook and flags nonstandard terms for legal · deviations are verifiable against the playbook, but signature risk demands human ownership
• Automate
  Three-way match — AI reconciles PO, receipt, and invoice, auto-clearing matches and queuing exceptions · the canonical rules-based, high-volume, fully verifiable back-office task
• Automate
  Supplier scorecarding — AI compiles delivery, quality, and pricing performance into quarterly scorecards · aggregation of recorded data; the review conversation stays human

process walkthrough (recording/screens) → drafted SOP → validated & published runbook → drift audit report

• Automate
  Process capture — AI converts a screen recording or narrated walkthrough of the process into a stepwise draft with screenshots · transcription-shaped work; the output is a draft by definition, so errors are harmless until validation
• Draft
  SOP drafting & formatting — AI structures the capture into the house template: prerequisites, steps, exceptions, escalation contacts · templated and checkable, but the process owner must confirm the exceptions, which the recording never shows
• Avoid
  Validation run — a person who has never done the task executes the SOP as written · this test only works because a human hits the gaps AI smoothed over; automating it deletes the point
• Automate
  Drift detection — AI compares SOPs against system UI changes, recent tickets, and actual logged behavior, flagging stale steps · mechanical diffing, high repetition, flags are cheap
• Draft
  Translation & localization — AI produces site-specific and language-specific variants of validated SOPs · derivative of verified content, but safety-relevant wording needs local review

historical + signal data → statistical forecast → consensus forecast → replenishment orders → forecast-accuracy postmortem

• Automate
  Data assembly & cleansing — AI reconciles sales history, promotions calendar, and channel inventory feeds, flagging anomalies before they poison the forecast · high-volume hygiene work, verifiable, errors caught in review
• Automate
  Baseline forecasting — ML models produce SKU-level statistical forecasts with confidence intervals · the model beats human guessing on volume and is graded weekly against actuals
• Assist
  Consensus adjustment — planners overlay market intelligence the model can't see: a competitor exit, a big deal in the pipeline, a port strike · the whole value of this step is context that lives outside the data
• Automate
  Replenishment execution — AI converts the consensus forecast into POs within preset min/max and budget guardrails · rules-bounded and reversible for fast-moving SKUs — but keep long-lead-time and seasonal buys on Draft because those orders can't be unwound
• Draft
  Exception review — AI surfaces the twenty SKUs where forecast and reality are diverging fastest, with a hypothesis for each · the ranking is mechanical; the hypotheses are guesses the planner must confirm
• Draft
  S&OP narrative — AI drafts the monthly pack: forecast changes, inventory position, risks, decisions needed · numbers are verifiable, but the pack drives executive commitments so the planning lead owns it

booking request → carrier tender → tracking timeline → exception alert & customer notice → freight invoice audit

• Automate
  Booking & tendering — AI selects carriers per the routing guide and tenders loads, escalating when primary carriers reject · rules-based against the routing guide, high volume, rejections are self-announcing
• Automate
  Document handling — AI extracts and files BOLs, customs docs, and PODs from emails and portals · document extraction at volume, verifiable against the source, the classic ops back-office win
• Automate
  Track & trace — AI monitors telematics and carrier feeds, maintaining a live ETA per shipment · pure aggregation; humans were never good at watching 500 dots on a map
• Draft
  Exception triage & customer notification — AI detects late/at-risk shipments, drafts customer notices, and proposes recovery options · the detection is automatic, but a wrong promise to a customer about recovery is costly, so a human approves the send
• Assist
  Expedite & recovery decisions — a human decides whether to eat a $4,000 air-freight bill to save a customer commitment · high-stakes tradeoff requiring relationship and margin context AI lacks
• Automate
  Freight invoice audit — AI checks carrier invoices against contracted rates and accessorial rules, auto-disputing variances · rules-vs-contract verification, high volume, disputes are reversible

deviation/incident report → triage classification → root-cause analysis (RCA) → CAPA plan → effectiveness check

• Automate
  Detection & intake — AI monitors inspection data, sensor streams, and complaint feeds, auto-filing structured deviation reports · pattern detection at volume; a false alarm costs minutes, a missed drift costs recalls, so tune for sensitivity
• Draft
  Triage & severity classification — AI classifies against the severity matrix and routes; anything touching safety or compliance is forced up to a human · rules-based, but a misclassified safety issue is close to irreversible, so a human confirms the tier
• Draft
  Investigation & RCA drafting — AI assembles the timeline from logs, batch records, and interviews, and drafts the 5-why/fishbone analysis · the timeline is verifiable, but AI reliably produces a plausible cause rather than the true one; the quality engineer must own the conclusion
• Assist
  CAPA design — humans design the corrective action; AI checks it against past CAPAs for repeat-failure patterns · the fix changes physical process and often carries regulatory exposure
• Draft
  Regulatory documentation — AI drafts the compliance-formatted incident record for the QMS · templated against known formats, but a regulator reads it, so sign-off is human and named
• Automate
  Effectiveness check — AI monitors post-CAPA metrics and flags recurrence automatically · scheduled statistical monitoring, exactly what gets forgotten when humans do it

supplier/geo risk register → monitoring alerts → scenario impact brief → continuity playbook → post-event review

• Automate
  Risk monitoring — AI scans news, weather, port congestion, financial filings, and sanctions lists for events touching mapped suppliers and lanes · watching the whole world is a volume problem; alerts are cheap to dismiss
• Automate
  Exposure mapping — AI joins the alert to the bill of materials and order book: which SKUs, which customers, how much revenue, how many weeks of buffer · deterministic joins over internal data, fully verifiable
• Draft
  Scenario impact brief — AI drafts the "if this lasts 2/6/12 weeks" analysis with mitigation options and costs · arithmetic is checkable, but the option set and probabilities need someone who knows the suppliers
• Draft
  Playbook maintenance — AI drafts and refreshes continuity playbooks per site/supplier, flagging ones untested in 12 months · templated documents against a checklist, human-validated like any SOP
• Avoid
  Activation decisions — leadership decides to trigger alternate sourcing, allocate constrained supply, or notify customers · rare, high-stakes, largely irreversible calls made under uncertainty AI can't price — it briefs, it does not decide
• Draft
  Post-event review — AI reconstructs the response timeline against the playbook and drafts the lessons-learned report · the reconstruction is verifiable from logs; the lessons need the people who lived it