Five production AI systems, taken apart
Twenty-one interactive diagrams across five real enterprise deployments — every agent, API call, intent classifier, retrieval index, gateway, guardrail and human-in-the-loop gate that has actually been published. Diagrams are pannable, zoomable, and every box is clickable for the quote it rests on. Crucially, each component is marked confirmed by source or inferred — because most published “enterprise AI architectures” quietly mix the two.
Klarna AI Assistant
The most-cited deployment in the world and the thinnest public record. Reconstructed from vendor case studies, SEC filings, a CEO podcast and one hands-on teardown — with the intent-classification gap marked honestly rather than papered over.
Uber QueryGPT
The best-documented agent pipeline in public: Intent Agent → Search → Table Agent → human ACK → Column Prune Agent → SQL. Plus the GenAI Gateway and the 2026 agent-identity layer — the real answer to “how do agents reach external systems”.
Morgan Stanley
Enterprise RAG under supervision. The architecture matters less than the evaluation methodology — a 500-question regression suite and ~5,000 daily annotations that gate every change to retrieval.
LinkedIn Hiring Assistant
The deepest published multi-agent system anywhere: supervisor plus seven sub-agents, a skill registry built on “skill inversion”, agent mailboxes riding the production messaging stack, and billion-scale semantic retrieval trained purely on LLM teacher labels.
Salesforce Agentforce
The pattern most enterprises will actually deploy — and the clearest public example of intent classification: Topics are the classifier, and the topic also bounds the valid action set. ServiceNow included for contrast.
The shape they all converge on
Cross-cutting reference architecture distilled from all five. Each box names which company does what, so you can see where the industry agrees and where it genuinely does not. The two dashed feedback edges are the ones most reference architectures leave out — guardrails injected centrally into every prompt, and evaluation gating every change to retrieval.
Side by side
| Klarna | Uber QueryGPT | Morgan Stanley | Agentforce | ||
|---|---|---|---|---|---|
| Intent classification | not published LangGraph routes; mechanism unknown | Intent Agent — LLM call mapping to one or more of 12 workspaces (multi-label) | Topic router selecting a content-type sub-model | Implicit in the supervisor's plan; intake agent produces structured qualifications | Topics — prompt-matched against natural-language classification descriptions |
| Orchestration | LangGraph multi-agent | Fixed 4-agent chain, no re-planning | Single-turn RAG; multi-turn deliberately disallowed | Supervisor, plan-and-execute (ReAct explicitly rejected) | Atlas ReAct loop on an event-driven graph |
| Retrieval | RAG over human-agent documentation; store unnamed | Workspace-scoped kNN over SQL samples + schemas | Business-rules layer over search, co-tuned with OpenAI | MUSE dual-tower + IVFPQ ANN + DCNv2 rerank over 1.3B profiles | Data Cloud retrievers, hybrid keyword + vector |
| External systems | Klarna backend APIs; write depth undisclosed | Data access proxies → Presto. 2026: MCP Gateway | Salesforce, Zoom, Outlook. 2026: MCP for clients' own agents | Skill registry + unified integrations (BuildIn/BuildOut) → ATS | Six action types: Flow, Prompt, Apex, OpenAPI, MuleSoft, MCP |
| Model gateway | Klarna AI Gateway (FastAPI/K8s) | GenAI Gateway — Go, OpenAI-compatible, PII redact/un-redact | Direct OpenAI, ZDR. not Azure | GenAI proxy exposing OpenAI Chat Completions for all models | Einstein Trust Layer |
| Guardrails | inferred prompt injection has succeeded publicly | Gateway-level: PII redaction, safety, hallucination detection | No long multi-turn; topics restricted; citation reason codes | Injected universally at Prompt Source of Truth; moderation at the proxy | Structural — the topic bounds the valid action set |
| Human in the loop | Unconditional escalation, offered to all customers | Table ACK gate in the critical path | Advisor reviews everything; nothing direct-to-client | Two gates: qualification review, outreach review | Omni-Channel escalation with context |
| Observability | LangSmith traces + LLM-as-judge | Bespoke harness: 5 signals, Vanilla vs Decoupled flows | 500-q regression suite; ~5,000 annotations/day; daily trending | LangSmith pre-prod, OpenTelemetry in prod, LLM-judge cascade | Testing Center + OTel session tracing export |
| Source quality | vendor case studies + filings only | first-party engineering blog | exec interviews + press + patents | 9 eng posts + 2 full conference transcripts | product docs + one eng post |
Seven things that recur across all five
1. Intent classification narrows retrieval, it doesn't route to code
Uber's Intent Agent exists to “drastically narrow the search radius for RAG.” Agentforce's Topics bound the action set. Morgan Stanley's topic router picks a corpus-specific model. In none of these is the classifier a dispatch table — it is a scope reducer.
2. The corpus is the product
Morgan Stanley: “All the boring stuff that nobody cares about is what makes this successful.” Klarna: the manuals had to be clear enough. Uber: hand-curated workspaces. Every team that succeeded had already solved content governance before adding a model.
3. One chokepoint for model calls, always
Uber's GenAI Gateway, LinkedIn's GenAI proxy, Klarna's AI Gateway, Salesforce's Trust Layer. Every one is an OpenAI-shaped interface with PII handling, quota, audit and cost attribution behind it. Nobody lets applications call providers directly.
4. Human gates sit inside the critical path
Not as a fallback for failure — as a designed step. Uber's table ACK, LinkedIn's qualification review, Morgan Stanley's advisor sign-off. All three were added or kept because autonomy without them measurably underperformed.
5. Model tiering follows the scaling axis
LinkedIn states it cleanly: surfaces that scale with #users get the frontier hosted model; the path that scales with #items gets a fine-tuned small model. Their evaluation agent is 90% of all LLM calls — and it runs on an 8B model that is 75× cheaper than GPT-4.
6. Evaluation gates deployment, not the other way round
Morgan Stanley re-runs 500 questions against every change to search rules or prompts. A broken recency rule in production was caught by daily annotation trending, not by an alert. Their eval framework became the firm's reusable governance artefact.
7. Published architectures quietly mix fact and inference
The single most useful thing in this pack is the dashed border. Klarna has published no function names at all, yet detailed “Klarna architectures” circulate widely. Uber explicitly says its validation agent is not shipped, yet it appears in third-party diagrams as if it were. Morgan Stanley's “fine-tuning on 100k documents” was corrected on record to RAG, years ago.
How to read these diagrams
| Signal | Meaning |
|---|---|
| Solid border | Named or described in a first-party source (company engineering blog, docs, SEC filing, recorded talk, or a vendor case study carrying a company quote) |
| Dashed border + ? | Reference-architecture inference. The box makes the diagram legible, but the company has not published it. Use the hide inferred button to see only what is confirmed. |
| Grey box | Explicitly not shipped — the company described it as planned or needed, and third parties often draw it as real |
| Click any box | Opens an inspector with the exact quote, the confirmed/inferred badge, and a link to the source |
| Edge colours | amber = API / model call · teal = data · violet = human loop · red = escalation · blue = external |
On source quality
Uber and LinkedIn publish real engineering detail and are the most reliable pages here. Morgan Stanley's material is executive interviews and press — rich on method, thin on components. Salesforce and ServiceNow are product documentation, reliable on mechanics and uncontrolled on performance claims. Klarna is the outlier: no engineering source exists at all, which is itself the most important thing to know about it.