AI Product Taxonomy
Executive summary
The enterprise AI market describes a wide set of solution classes — from corporate LLM access to optimization and digital twins. This page captures 18 classes as a reference / research taxonomy: it helps classify business requests and decide when a class deserves its own AI product.
The AI operating model keeps the real catalog at no more than 10 delivery tracks — the routes through which AI initiatives move, get owners, and pass stage gates. Other classes serve as reference material, industry slots, or cross-cutting layers.
For the operational catalog, see AI products.
How to read the taxonomy
This page uses different levels of abstraction:
| Level | What it is |
|---|---|
| Model / platform | Reusable technical capability (LLM gateway, RAG, ML platform) |
| AI product | Managed solution with an owner, SLA, and usage rules |
| Capability / layer | Function inside a product (guardrails, eval, data readiness) |
| AI initiative | A specific use case in the business funnel |
| Delivery track | Implementation route in the delivery funnel |
The key question: does this need a separate AI product and delivery track, or is a capability inside an existing track enough?
Full taxonomy: 18 classes
Compact reference table. Tier reflects typical enterprise demand maturity — not mandatory standalone tracks.
| # | Class | Tier | What it is | Verbs | Include when | Do not create separately when |
|---|---|---|---|---|---|---|
| 1 | LLM Self-Service & Model Gateway | Core | Governed access to approved foundation models via UI/API: chat, playground, routing, templates, cost controls, governance | generate, summarize, classify, extract, rewrite, translate, reason, draft | 3+ functions need GenAI or multiple initiatives need reusable model access | M365/Google Copilot covers the need with adequate controls, or it is only a feature inside one app |
| 2 | Enterprise Knowledge Search / RAG / Q&A | Core | Grounded search and Q&A over enterprise data with citations, permission-aware retrieval, and traceability | find, search, answer, summarize, compare, retrieve | Same Q&A/search pattern appears in 3+ functions or 5+ knowledge domains | One-document chatbot, data is not governed, or enterprise search already solves it |
| 3 | Agentic Workflow / Routine Automation | Core | Orchestrates LLM agents, deterministic workflows, tools, APIs, humans, and approvals for multi-step tasks | automate, orchestrate, collect, decide, act, monitor, escalate | High-volume routine work repeats across functions; APIs or clear approval gates exist | One-off expert work, process is undefined, no system access, or RPA/iPaaS already solves it |
| 4 | Document Intelligence / IDP | Core | Document classification, splitting, OCR, extraction, validation, and routing into business systems | extract, classify, process, split, compare, validate, route | 3+ processes or document types need repeatable extraction and validation | Low volume, data available via API, or it is a one-off form |
| 5 | Meeting & Conversation Intelligence | Advanced | Captures, transcribes, and summarizes meetings/calls; extracts decisions and action items | transcribe, summarize, extract actions, coach | Meetings/calls are high-volume work artifacts; follow-up is a recurring pain | Legal/culture prevents recording, or native M365/Zoom capability is sufficient |
| 6 | Coding Agent / Software Engineering Copilot | Core | IDE/CLI/cloud agents for repo understanding, code generation, refactoring, testing, debugging, PR assistance | write code, explain, refactor, test, debug, document, migrate | Software delivery is material; enough developers and repos to justify governance | Few developers, or a centrally procured IDE add-on with no portfolio management need |
| 7 | ML Platform / Model Factory / Decision AI | Core | Platform for classical ML, scoring, forecasting, anomaly detection, recommendations, and model lifecycle | predict, score, classify, forecast, recommend, detect anomaly | Multiple predictive/decision use cases recur; historical data exists | Problem is search/generation; no labels, business owner, or action path |
| 8 | BI / Analytics Copilot / Decision Intelligence | Advanced | Natural-language and agentic analytics over governed semantic data, metrics, and dashboards | ask data, analyze, compare, explain variance, monitor KPIs | KPI/analytics questions repeat; governed semantic data exists | Data is not trusted, or each analysis is bespoke consulting |
| 9 | Customer & Employee Service Agent | Advanced | AI agents for support/service desks: answer, triage, resolve, tickets, escalation | answer, classify, route, resolve, automate, escalate | High support volume; KB and actions are reusable | Low-volume expert advice or no KB/action integration |
| 10 | Voice / Speech AI & Contact Center | Optional | Real-time speech, voice agents, agent assist, call summarization, QA, analytics | transcribe, summarize, answer by voice, route, coach | Voice is a core channel or contact-center volume is high | Meeting transcription is enough, or legal consent is infeasible |
| 11 | AI Governance / Portfolio Management / Control Tower | Core | System of record for AI initiatives, models, agents, risks, approvals, ownership, value, compliance | govern, inventory, approve, prioritize, monitor, report | Large or regulated company with multiple AI products/agents | Fewer than ~10 AI uses; start with lightweight GRC/PMO workflow |
| 12 | AI Security / Guardrails / Red Teaming | Core | Security controls and testing for AI apps, LLMs, agents, prompts, data, model supply chain | protect, detect, block, redact, monitor, test, respond | AI touches sensitive data, external users, or tool/actions | Low-risk prototype; embed minimum controls inside the LLM platform |
| 13 | AI Evaluation, Observability & LLMOps | Core | Testing, tracing, monitoring, and improving LLM apps, RAG, agents, prompts, costs, quality | evaluate, monitor, debug, trace, compare, optimize | 3+ production LLM/RAG/agent apps | Vendor black-box without instrumentation; use vendor monitoring instead |
| 14 | Data & Knowledge Readiness / AI Data Product | Core | Reusable layer: catalog, access, lineage, quality, classification, semantic definitions, knowledge readiness | connect, govern, classify, enrich, curate, retrieve, reuse | Data readiness is a recurring bottleneck across AI initiatives | Mature enterprise data platform already owns it; treat as capability, not duplicate |
| 15 | Multimodal Content Generation / Brand Creative | Optional | Controlled generation/editing of text, image, video, audio, design with brand/legal controls | generate, edit, localize, personalize, version, review | High creative/localization volume across business units | Occasional design work, or agency/DAM workflow already covers it |
| 16 | Computer Vision / Visual AI / Edge AI | Optional | Analyzes images/video/physical environments: detection, inspection, safety, process visibility | detect, classify, count, inspect, monitor, alert | Physical operations have repeatable visual problems | Office knowledge work, or sensor/API data is enough |
| 17 | Research / Expert Workbench / High-Stakes Analysis | Advanced | Expert workbench for corpus analysis, source comparison, evidence-backed memos, high-stakes decisions | research, compare, diligence, synthesize, cite | High-value expert analysis repeats across functions | General RAG or BI copilot already covers the task |
| 18 | Optimization / Simulation / Digital Twin / Prescriptive AI | Optional | AI/OR/simulation for optimal plans under constraints and scenario simulation | optimize, simulate, recommend, allocate, schedule, prescribe | Decisions are high-value, repeated, codifiable with constraints/data | Ad hoc decisions, unreliable data, or planning system already optimizes well |
Mapping to the operational catalog
The operational catalog stays at no more than 10 delivery tracks. Industry slots A and B are reserved for recurring company-specific demand (analytics, voice, creative, CV, research, optimization — depending on context).
The 10 operational tracks:
| # | Delivery track |
|---|---|
| 1 | Corporate LLM |
| 2 | RAG / Knowledge Assistant |
| 3 | ML platform |
| 4 | Code agent |
| 5 | Automation AI (agents + orchestration) |
| 6 | Document Intelligence |
| 7 | Meeting intelligence |
| 8 | Service Agent |
| 9 | Industry slot A |
| 10 | Industry slot B |
18 classes → default operational handling:
| Research class | Default operational handling |
|---|---|
| LLM Self-Service & Model Gateway | Track 1: Corporate LLM |
| Enterprise Knowledge Search / RAG / Q&A | Track 2: RAG / Knowledge Assistant |
| ML Platform / Model Factory / Decision AI | Track 3: ML platform |
| Coding Agent / Software Engineering Copilot | Track 4: Code agent |
| Agentic Workflow / Routine Automation | Track 5: Automation AI |
| Document Intelligence / IDP | Track 6: Document Intelligence |
| Meeting & Conversation Intelligence | Track 7: Meeting intelligence |
| Customer & Employee Service Agent | Track 8: Service Agent |
| BI / Analytics Copilot | Track 9/10 if analytics is a repeated strategic demand; otherwise applied capability over the data/BI stack |
| Voice / Speech AI | Track 9/10 for contact-center-heavy companies; otherwise a channel inside Service Agent |
| Multimodal Content Factory | Track 9/10 for marketing/content-heavy companies |
| Computer Vision / Edge AI | Track 9/10 for manufacturing, retail, logistics, security, field operations |
| Research / Expert Workbench | Track 9/10 for legal, risk, investment, regulatory contexts; otherwise advanced RAG pattern |
| Optimization / Digital Twin | Track 9/10 for planning-heavy industries; often an ML/OR extension |
| AI Governance / Portfolio Management | Cross-cutting governance layer — not a separate track by default |
| AI Security / Guardrails / Red Teaming | Cross-cutting security layer — not a separate track by default |
| AI Evaluation, Observability & LLMOps | Cross-cutting quality/release layer — not a separate track by default |
| Data & Knowledge Readiness | Cross-cutting data/knowledge layer — not a separate track by default |
Governance, AI security, eval/observability, and data readiness are research classes and cross-cutting layers by default. A separate delivery track is justified only with a clear owner and repeated company demand.
Name synonyms
The market uses different names for the same classes. Synonyms help map market language to operational tracks without catalog sprawl:
| Operational track | Market synonyms |
|---|---|
| RAG / Knowledge Assistant | Enterprise Knowledge Search, grounded Q&A, knowledge assistant |
| Automation AI | Agentic Workflow, agentic automation, routine automation platform |
| Code agent | Coding Agent, Software Engineering Copilot, dev copilot |
Sources and market anchors
The links below are market anchors: they help orient in the current landscape but are not timeless methodology facts. Forecasts and vendor positioning change.
- Stanford AI Index 2025
- Gartner: over 40% of agentic AI projects will be canceled by end of 2027
- OWASP Top 10 for LLM Applications
Product examples (to clarify classes only — not recommendations):
- Azure AI Document Intelligence — Document Intelligence
- GitHub Copilot coding agent — Code agent
- Microsoft Fabric Data Agent — BI / Analytics Copilot
- ServiceNow AI Control Tower — AI Governance
- LangSmith — AI Evaluation & LLMOps
- Microsoft Purview data governance — Data & Knowledge Readiness
Related sections
- AI products — operational catalog and two product levels
- Selecting an AI product — routing initiatives playbook
- Delivery track — delivery funnel
- Glossary — terminology canon
- Industry insights overview