Where AI shines — and which AI

A simple map for AI use cases

Two questions place any task on the map. How repetitive is it? (does the same shape of work happen many times?) How reasoning-heavy is it? (does each instance require judgement, synthesis, multi-step thought?) The top-right quadrant — where work is both repetitive and reasoning-dense — is where AI changes the economics of the work. But knowing where AI fits is only half the answer; you also need to know which AI to use.

Part 1 · Where

The 2 × 2 below plots every use case in the gallery. Hover for the title. Blue dots augment a human; orange dots automate the step.

Reasoning →
Repetition →
Augment — AI assists, human decides Automate — AI runs the step end-to-end

Part 2 · Which AI

Modern AI workflows are not "just an LLM." They layer three families — generative at the unstructured boundary, predictive at the classification step, symbolic at the guardrails. Each plays a different role; together they do work that none can do alone.

Symbolic AI

Rules & logic

Explicit rules, lookup tables, knowledge graphs, constraint solvers. Deterministic. What you write is what runs.

Examples: validation (totals match), routing trees, policy gates, schema constraints, business-rule engines, master-data joins.

Why it stays: when wrong is expensive — compliance, posting to a system of record, blocking unsafe outputs — you want a layer the auditor can read.

Predictive AI

Pattern learners

Classifiers, regressors, forecasters, anomaly detectors trained on labelled data. Statistical. Narrow, accurate, replayable.

Examples: intent classification, OCR text detection, time-series forecast, churn score, isolation-forest anomaly, sentiment, fraud scoring.

Why it stays: for high-volume classification or scoring it's cheaper, faster, more reliable than generative — and easier to evaluate against a labelled hold-out.

Generative AI

Artifact makers

LLMs, image & voice models that produce new content: text, code, structured records, images, speech.

Examples: field extraction from messy docs, drafting SOPs, headline writing, image generation, SQL-from-question, narration of insights.

Why it's needed: the unstructured boundary — when input is a free-form PDF, voice memo, or vague brief — only a generative model can understand or produce the unconstrained shape.

Part 3 · How the three combine in each use case

Almost every interesting AI workflow uses all three. The bold tag (Heart) marks the AI type that does the load-bearing work for that use case.

Use case Symbolic — rules & logic Predictive — pattern learners Generative — artifact makers
OCR: PO → SOworkflow automation · automate Catalog lookup, totals-match, customer-known checks; SAP posting via deterministic mapping. OCR text detection & layout classification (which template). HeartVision-LLM extracts free-form fields (vendor name variants, multi-line addresses) into structured JSON.
IVR → STT routingworkflow automation · automate Confidence threshold, routing tree (intent → queue), instant-bail keywords ("manager"), PII redaction rules. HeartSTT acoustic model + intent classifier (top-12 intents on labelled history). LLM disambiguates ambiguous utterances; TTS synthesises self-serve replies.
Anomaly detect & notifydata analytics · automate Alert rules (z-score > 3), suppression windows, escalation paths (who is paged when). HeartTime-series forecast (Prophet / DeepAR) + isolation forest / autoencoder reconstruction error. LLM writes the natural-language alert summary ("sales dipped 22% on Friday because…").
Auto-generate dashboarddata analytics · augment Chart-selection rules (categorical + temporal → bar by month), KPI threshold colouring, layout grid. Column-type inference, distribution profiling, trend / seasonality detection. HeartLLM reads schema + question, writes SQL, picks layout, and narrates each chart's headline.
Generate SOPcontent generation · augment SOP template structure (purpose / scope / steps / approvals); policy-language guardrails. STT transcription + speaker diarisation on the source recording. HeartLLM rewrites the rough transcript into formal numbered procedure with role assignments.
Generate posterscontent generation · augment Brand grid: type scale, palette, safe zones, logo placement — enforced after gen. Aesthetic / brand-fit scoring of candidate compositions; image quality classifier. HeartLLM writes headline + body copy; image model generates background visuals from the brief.
OCR documents → Excelworkflow automation · automate Column-alias dictionary; schema validator (type checks, ranges); .xlsx writer. Table detection on each page; layout classifier picks the extractor strategy. HeartVision-LLM extracts rows from heterogeneous tables and reconciles column names across pages.
Resume → Excel + JD-fitworkflow automation · automate Eligibility floor rules (years, location, lang); name redaction guard during scoring; output schema. NER for name/dates/skills/education; resume layout parser; bias-audit checks. HeartLLM scores fit vs JD using the rubric prompt and writes the "Why" evidence line for each candidate.
Data → action plans & ticketsdata analytics · augment RACI map (metric → owner); ticket schema (Jira API); de-duplication windows; SLA gates. Cluster related findings; impact ranker (magnitude × baseline × decay); priority scoring. HeartLLM drafts hypothesis + step-by-step plan + acceptance criteria + due date for each ticket.
Pricing from internal + externaldata analytics · augment Floor / ceiling enforcement; promo-overlap rules; constraint solver respecting business limits. HeartLog-linear elasticity per cluster from 12+ weeks of history; demand forecast at the new price. LLM narrates the rationale ("sit just under cluster avg") so the manager can review.
Virtual management meetingworkflow · augment Role personas + meeting protocol (round-robin, time-box, topic gates); RACI rules per agent. Sentiment / consensus scoring across rounds; flags agents who shifted position. HeartMulti-agent LLM debate — each agent stays in role and applies its mandate; final synthesis writes the meeting brief.
Talk to databasedata analytics · augment Schema graph (tables, joins, FKs); SQL executor; row-count + result-type checks; PII column blocklist. Disambiguation of ambiguous entities; auto-correct misspelled column names. HeartLLM reads schema + question, picks tables, writes SQL, then narrates the result in business language.
Meeting → minutes & actionsworkflow · augment Meeting template (decisions / actions / parking-lot); attendee-list → owner mapper; SLA gates. STT with speaker diarisation + speaker identification (matched to attendee list). HeartLLM extracts decisions vs discussions vs parking-lot; drafts minutes in house style; writes each task.
Email triage + draft replyworkflow · augment Routing rules: priority sender list, escalation paths, urgent-keyword overrides, auto-archive list. Intent classifier (FYI / question / scheduling / action) + priority predictor. HeartLLM drafts reply in your voice using past-thread context; flags the ones that need your judgement.
Invoice 3-way matchworkflow · automate HeartThe matching rules: qty(GR) ≤ qty(PO); price(invoice) = price(PO); total within ±2%; vendor approved. OCR text detection on each doc; document-type classifier. Vision-LLM extracts free-form fields (vendor name variations, multi-line addresses) into the schema.
KYC document verificationworkflow · automate HeartKYC tier rules; sanctions / PEP / watchlist matching; ID expiry + name-on-address-proof check. OCR + ID layout classifier; face-match similarity score; document-tampering detector. LLM normalises name variants and address formatting across docs.
VOC clusteringdata analytics · augment Topic dedup rules (merge near-duplicate clusters); business-domain glossary; PII redaction. HeartEmbedding model (sentence-transformers) + HDBSCAN clustering + per-cluster sentiment scoring. LLM names each cluster ("Slow checkout — promo codes") and writes a one-line summary.
Churn prediction + outreachdata analytics · automate Signal aggregation pipeline; intervention playbook (which action for which segment + risk-level). HeartGradient-boosted classifier (XGBoost) on tabular signals; uplift model picks the right save offer. LLM personalises outreach with the customer's actual usage drop and recent ticket history.
Code review on PRscoding · augment Linter / formatter / type-checker; test-coverage threshold gates; CI must-pass rules. Pattern matchers for common bug shapes (off-by-one, null deref, race, SQL injection). HeartLLM reads diff in context, writes per-line comments, suggests fixes as code blocks; explains why.
Test case generationcoding · augment Test runner (pytest / jest); coverage tool; CI gate that requires generated tests to pass. Edge-case predictor (boundaries, type-mismatch, null inputs) trained on bug history. HeartLLM reads signature + docstring, generates pytest cases including property-based via Hypothesis.
Product description from speccontent · augment Brand voice rules (banned words, tone, length); regulatory claim list; SEO meta length cap. Spec-sheet table extractor; image classifier confirms photo matches description. HeartLLM writes the description, 3 bullets, and SEO meta from the spec + brand voice prompt.
Forecast → replenishmentdata analytics · automate Reorder formula (ROP = avg demand × lead time + safety stock); approval thresholds; supplier MOQ. HeartHierarchical time-series forecast (Prophet / DeepAR / Croston for intermittent); seasonality + promo lift. LLM writes the rationale on each suggested PO ("demand up 14% MoM, stock < 7-day cover").
Outbound voice agentworkflow · automate Conversation tree (confirm / reschedule / opt-out / handoff); compliance disclaimers; calling-window rules per timezone. STT for utterances; intent classifier; sentiment scoring for handoff signal. HeartTTS in friendly natural voice; LLM handles unexpected utterances within policy scope.
Bank statement reconciliationworkflow · automate HeartMatching rules (ref equality; amount within ±1฿; date ±3d); GL coding; auto-clear thresholds; FX policies. String similarity / fuzzy match scoring on memos; memo categoriser. LLM proposes probable match for unstructured memos by date proximity + invoice match.
Lead scoring + routingdata analytics · automate Routing playbook (territories, segments, queue caps); SLA timers; CRM record schema. HeartLead-scoring model (XGBoost on conversion history) + intent score from web/email behaviour. LLM drafts the personalised first-touch email when lead lands in AE queue.
Sales tacit → customer knowledgeworkflow · augment CRM knowledge schema (decision-makers, pain, competitive landscape); interview question library; PII redaction. STT with sales jargon vocab; named-entity recognition for people / company / product / currency. HeartLLM generates adaptive questions; synthesises tacit observations into structured CRM facts + narrative.
Production planning optimizationdata analytics · augment HeartMILP solver: minimise cost subject to capacity, labour, material, demand, changeover penalties, inventory caps. Demand forecast per SKU × week; machine downtime predictor; quality yield model. LLM writes the plan rationale and explains tradeoffs to the planner.
Customer support reply draftworkflow · augment Brand voice rules; escalation rules (refunds > ฿X to senior); required disclaimers; PII redaction in retrieved context. Vector search over policy KB (semantic retrieval) + ticket-similarity match for past resolutions. HeartLLM drafts the reply with the right blend of empathy, answer, next-step, and brand voice.
Sales call coachingworkflow · augment Must-cover topics (price, timeline, decision-process, next-step); banned phrases; rep tier rules. HeartSTT diarisation; talk-ratio computation; question density; sentiment trajectory; filler-word counter. LLM writes 3 specific coaching observations + a positive lead — never punitive.
Translation Thai ↔ ENcontent · augment Brand glossary (locked translations); tone rules; banned-phrase list per language; integrity check. Glossary term detection (NER-style); language detection; sentence segmentation. HeartLLM translates Thai ↔ EN preserving tone, idiom, and brand voice.
KB article from ticket clustercontent · augment KB style template (Problem / Cause / Fix / FAQ); link-validity checker; obsolete-content flag. Embedding + clustering (HDBSCAN) on solved tickets; cluster size + recency for prioritisation. HeartLLM reads ticket cluster + resolution notes; writes the article in standard KB voice.

Part 4 · Vocabulary on the cards

Every card carries the same metadata so use cases are comparable, not just memorable.

The four modalities

Every AI workflow swaps in for one of four human senses. Naming the modality first sharpens the design.

  • Read — extract structure from text, images, scans, screens.
  • Write — generate text, code, structured records, images.
  • Listen — turn speech or audio into text or signal.
  • Speak — turn text into speech, voice agents, narration.

The four domains

Where the work lives in a business. Each domain has a different bar for accuracy vs creativity.

  • Workflow Automation — moving data between systems. Bar: accuracy.
  • Content Generation — making artifacts (docs, posters, posts). Bar: creativity within brand.
  • Data Analytics — sensing, summarizing, alerting on data. Bar: accuracy + explainability.
  • Coding — writing, reviewing, testing software. Bar: correctness.

Augment vs automate

The same task can be either. The choice is about cost of being wrong and volume.

  • Augment when each instance carries reputational or legal risk, or when judgement is needed at the boundary. Human reviews, AI drafts.
  • Automate when volume is high, error cost is low or recoverable, and the success rule is clear. AI runs end-to-end with sampling QA.
  • Most use cases ride the seam — start as augment, graduate to automate as confidence builds.

How to read each card

  • Modality + Domain tags — which sense, which department.
  • Mode badge — Augment (blue) or Automate (orange).
  • Repetition × Reasoning dots — three filled dots = high.
  • Standalone page — interactive demo (Input → Process → Output) plus problem, approach, when it works, when it doesn't.
  • AI-type pills inside the demo — slate=symbolic, teal=predictive, violet=generative.