Deployments — who runs what, and how
You asked where the infra lives and how agents are actually deployed, with real examples. This chapter maps each use-case to its deployment pattern (the infra shape) and lists 10+ real, publicly-known deployments for each. The mechanics behind these patterns are in Serving, Deploy & scale, Isolation, and the Landscape Reference architecture.
A dedicated fact-checking pass (108 agents, 24 claims confirmed 3‑0 against primary sources) verified the deep-dives in the next section — those carry Verified and a source link. The broad catalog tables further down list well-known public products to show market breadth; treat those as "these exist and do roughly this," not audited figures. Verified metrics are bolded throughout.
The universal deployment anatomy
Almost every production agent, regardless of domain, is assembled from the same tiers (see the reference architecture): a client → an API + AI gateway (routing, guardrails, cost/rate limits) → the agent runtime running the loop → tools / MCP servers → a sandbox for code/browser actions → a memory + vector tier → the model (hosted API or self-hosted vLLM on GPUs) → observability across all of it. What changes per use-case is which tiers dominate. Below, each section names that dominant pattern, then the examples.
Verified deep-dives — fact-checked, with architecture
Each of these was confirmed against primary sources in the follow-up research pass. This is the "how it's really built and how it went" layer.
Handled 2.3M conversations in month one (~⅔ of all chats), the work of ~700 agents; ~$40M projected 2024 profit lift; resolution 11 min → under 2 min, 25% fewer repeats, 23 markets, 35+ languages. Architecture: direct API access to accounts / transaction history / payment systems + a knowledge base for grounding + escalation logic routing complex cases to humans. klarna.com · openai.com/index/klarna
Klarna is not a pure success story. By May 2025 the CEO publicly walked back the AI-first approach, saying cost-cutting had lowered quality — the AI was strong on routine tier-1 volume but weak on the long tail (fraud, billing disputes, hardship cases), degrading CSAT. Klarna moved to a dual-track human+AI model (AI for tier-1, humans for escalations). It disputes "AI reversal" framing and says capacity reached ~853 roles' work with cost/transaction falling $0.32→$0.19 by late 2025. The honest lesson: average-case metrics looked great while long-tail failures hurt the brand — which is exactly why human-in-the-loop and escalation design matter.
$500M ARR, 330% YoY (Salesforce's fastest-growing product); 18,500 deals (9,500+ paid); handled 1.5M+ support requests and processed 3.2 trillion tokens. Architecture: Data Cloud (context unification) + MuleSoft (system access) + Agent Builder + Prompt Builder + Command Center (monitoring). Named outcomes: Williams-Sonoma's "Olive" handles ~60% of conversations; the IRS cut case-open time 10 days → 30 min; UK Police cut non-emergency calls 20%. Salesforce earnings + product docs
142,000+ lawyers, 1,500+ orgs, 60 countries, 92% monthly adoption; customers built 25,000+ custom Workflow Agents via an Agent Builder. Power users save 36.9 hrs/month (standard 15.7). Architecture: Assistant (drafting) + Vault (bulk analysis up to 100k docs/project) + Knowledge (research over 500+ sources) + Word/365/iManage integrations + LexisNexis for case law. harvey.ai
UPMC scaling Abridge to 12,000 clinicians by 2026 across 40+ hospitals / 800 sites / 44+ specialties; converts real-time conversation → billable documentation, with Linked Evidence mapping each summary line back to the source audio. Nuance DAX Copilot crossed 200,000 clinician users in 2026. Pattern: ambient speech → LLM summary → EHR, clinician-reviewed. abridge.com · Microsoft/Nuance
A multi-agent system on LangChain/LangGraph + RAG: intent classification → embedding retrieval → a knowledge graph → LLM re-rankers → a self-correction agent. ~95% of users rate accuracy "Passes" or above; embedded in LinkedIn's DARWIN platform (+ DataHub metadata), which lifted adoption 5–10×; serves hundreds of employees over millions of tables. LinkedIn Engineering
GPT-4 Turbo (128K) + vector DB + RAG + a multi-agent pipeline — Intent Agent (maps question → business workspace), Table Agent (picks tables), Column Prune Agent (drops columns to fit tokens). Query authoring ~10 min → ~3 min; ~78% of users saved time. Motivated by Uber's ~1.2M interactive queries/month. uber.com/blog/query-gpt
Microsoft rolled Claude Code + Copilot CLI to tens of thousands of engineers (early 2026, 4-month study): adopters merged ~24% more pull requests, sustained past the novelty phase, spreading via peer networks. A separate 430-engineer enterprise bakeoff (Faros AI): Copilot 78% adoption, 22% accept rate, ~$11.2M/yr value vs Amazon Q's 39% / 11% / $7.8M. arXiv 2607.01418 · Faros AI
Core loop = gather context → take action → verify work → repeat. Uses agentic filesystem search (grep/tail), automatic context compaction near the limit, isolated-context subagents for parallelism, MCP integrations, and verification via lint / visual feedback / LLM-as-judge. Anthropic uses it internally for almost all its major agent loops (deep research, video, notes). anthropic.com/engineering
Cleric: investigates incidents and returns root cause in Slack with evidence + confidence; integrates existing Datadog/Grafana/CI-CD (no re-instrumentation); adopters free 20–30% of engineering capacity; in production at BlaBlaCar. Resolve AI ($125M Series A, $1B valuation): an "agentic interface to production" — Coinbase cut critical-incident investigation time 72%; Zscaler cut engineers-per-incident 30%. businesswire · resolve.ai
Dropzone: investigates security alerts end-to-end across SIEM/EDR/cloud — −85% manual investigation, 5× faster MTTR; CBTS offloads 30–50% of alert volume. Skyvern: automates browser workflows with computer vision + code + LLM reasoning (no brittle CSS selectors), BYO-LLM, self-hostable via Docker on AWS/GCP/Azure — 10M+ workflows run, 99.9% uptime, 20k+ GitHub stars. dropzone.ai · skyvern.com
1 · Coding agents
Deploy pattern: IDE/CLI or cloud client → ReAct loop → repo indexing (embeddings/grep) + file/shell/git tools → sandboxed execution → model via gateway. Async variants run in a cloud sandbox and open a PR.
| Company / product | What it is · notable numbers |
|---|---|
| GitHub Copilot | Inline completions + agent mode (multi-file edits, terminal). The category's default; hundreds of millions in ARR. |
| Cursor | VS Code fork; builds a semantic index of the repo; agent mode does multi-file changes + terminal. |
| Claude Code | Terminal agent: grep/glob/read to navigate, runs tests & type-checkers to verify before committing. ReAct while-loop; file-based Markdown memory. |
| Devin (Cognition) | Autonomous async engineer in a sandbox: plans, codes, debugs, opens PRs. Example: 47 files migrated in ~20 min, tests passing. |
| Windsurf (Codeium) | Agentic IDE ('Cascade') with repo-wide context and multi-step edits. |
| Replit Agent | Builds & deploys full apps from a prompt inside Replit's cloud sandbox. |
| Amazon Q Developer | AWS coding/agent assistant; multi-file features, AWS-aware, autonomous upgrades. |
| Google Jules / Gemini Code Assist | Async cloud coding agent that works on a repo branch and proposes changes. |
| Sourcegraph Cody / Amp | Code search-grounded assistant + agent over large enterprise monorepos. |
| Aider | Open-source terminal pair-programmer with git-aware edits. |
| Tabnine | Privacy-focused completions/agent, self-hostable for enterprises. |
| JetBrains AI / Junie | In-IDE agent across the JetBrains suite. |
2 · Customer support & service
Deploy pattern: chat channel → RAG over help-center/KB + action tools (order/refund/account APIs) → guardrails + PII redaction → human handoff on low confidence. Hosted on a managed platform or K8s behind an AI gateway.
| Company / product | What it is · notable numbers |
|---|---|
| Klarna (OpenAI) | 2.3M conversations = ~700–853 agents' work; resolution 11 min → <2 min; 23 markets, 35+ languages; ~$40M profit impact. |
| Intercom Fin | Support agent grounded in help content; resolution-based pricing; widely deployed. |
| Sierra | Bret Taylor's conversational-AI-agent platform for branded customer experiences. |
| Decagon | Enterprise support agents (Duolingo, Notion, Rippling cited publicly). |
| Salesforce Agentforce | Service/SDR agents on the Salesforce platform. +40% lead follow-up efficiency reported. |
| Ada | No-code automated customer service across channels. |
| Zendesk AI Agents | Native resolution agents inside Zendesk's help desk. |
| Cresta | Real-time contact-center agent assist + automation. |
| Forethought | Support automation ('Solve') with intent routing. |
| Gorgias | E-commerce support automation (Shopify ecosystem). |
| Parloa / PolyAI | Voice-first contact-center agents. |
3 · Knowledge assistants & deep research
Deploy pattern: enterprise connectors → ingestion → vector index with ACL pre-filtering → hybrid retrieval + rerank → LLM. "Deep research" variants add an agentic multi-step retrieval loop.
| Company / product | What it is · notable numbers |
|---|---|
| Morgan Stanley (GPT-4) | RAG over ~100,000 research docs; 98% advisor adoption; access 20% → 80%; synthesis 30+ min → seconds. |
| Glean | Enterprise search + assistant over all company apps with permission-aware retrieval. |
| Perplexity | Consumer answer engine: live retrieval + citation-grounded generation. |
| OpenAI Deep Research | Agentic multi-step web research producing long cited reports. |
| Google Gemini Deep Research | Plans, browses, and synthesizes multi-source reports. |
| Hebbia | Document-heavy research for finance/law ('Matrix' over filings). |
| Elicit | Research assistant over academic papers with structured extraction. |
| Notion AI / Dropbox Dash / Guru | Workspace assistants with permission-aware retrieval over org content. |
| Amazon Q Business | Enterprise RAG assistant over connected data sources. |
| Microsoft 365 Copilot | Assistant grounded in the Microsoft Graph (mail, docs, chat). |
4 · Legal
Deploy pattern: RAG over case law/contracts with citation grounding + strict eval/guardrails + mandatory human review; often single-tenant/VPC for confidentiality.
| Company / product | What it is · notable numbers |
|---|---|
| Harvey | Gen-AI for elite law firms & PwC; research, drafting, review workflows. |
| CoCounsel (Thomson Reuters / Casetext) | Legal assistant grounded in authoritative legal databases. |
| Robin AI | Contract review & drafting copilot. |
| Legora (ex-Leya) | Collaborative legal AI for research and review. |
| Spellbook | Contract drafting/review inside Microsoft Word. |
| Luminance | Contract analysis ('Autopilot') for legal teams. |
| Ironclad AI | Contract lifecycle management with AI redlining. |
| DraftWise | Drafting from a firm's own precedent. |
| Legalfly / Definely | EU-focused legal drafting & review assistants. |
| Paxton AI | Legal research & drafting assistant with citation grounding. |
| vLex Vincent AI | Agentic legal research across a global case-law corpus. |
| Clio Duo | AI assistant embedded in Clio's practice-management suite. |
5 · Finance & banking
Deploy pattern: private VPC / on-prem or zero-retention APIs, RAG over filings/research, heavy governance, HITL. Highest security bar.
| Company / product | What it is · notable numbers |
|---|---|
| JPMorgan Chase | 450+ AI use cases in production; LLM Suite ~200k daily users; M&A-memo agent: hours → <30 sec. |
| Morgan Stanley | GPT-4 RAG assistant + AI @ MS Debrief meeting summaries; zero-data-retention governance. |
| BloombergGPT | Domain LLM trained on financial data for finance NLP tasks. |
| Hebbia | Agentic analysis over financial documents for asset managers. |
| Ramp / Brex | AI agents for spend, expense categorization, and finance ops. |
| Rogo | AI analyst for investment banking research/modeling. |
| AlphaSense | Market-intelligence search + generative summaries over filings/transcripts. |
| Bridgewater | Reported internal 'AIA' analyst assistant for investment research. |
| Klarna (internal ops) | Beyond support — internal knowledge and marketing agents. |
| Greenlite AI | Autonomous agents for AML / KYC compliance review at banks & fintechs. |
| Sardine | AI agents for fraud, AML, and compliance operations. |
| Casca | AI loan-origination agent automating lending workflows. |
6 · Healthcare
Deploy pattern: ambient capture (speech → LLM summary) or evidence RAG, HIPAA-grade infra, clinician review before anything is finalized.
| Company / product | What it is · notable numbers |
|---|---|
| Abridge | Ambient clinical documentation from patient conversations; major US health systems. |
| Nuance DAX Copilot (Microsoft) | Ambient scribe integrated with Epic/EHRs. |
| Hippocratic AI | Safety-focused patient-facing voice agents for non-diagnostic tasks. |
| OpenEvidence | Evidence-grounded medical Q&A for clinicians. |
| Suki | Voice AI assistant for clinical notes. |
| Ambience Healthcare | Ambient documentation + coding across specialties. |
| Nabla | Ambient AR (assistant) for clinicians. |
| Corti | Real-time clinical decision/documentation support. |
| Microsoft Dragon Copilot | Unified ambient clinical documentation (DAX + Dragon lineage). |
| Regard | Automated chart review that surfaces diagnoses from the EHR. |
| Commure / Athelas | Ambient scribe + revenue-cycle and ops automation. |
7 · Sales, marketing & GTM
Deploy pattern: CRM/data tools + enrichment + RAG, often scheduled/async background runs (prospecting, outreach) with HITL approval before sends.
| Company / product | What it is · notable numbers |
|---|---|
| Salesforce Agentforce | Autonomous SDR & marketing agents on the CRM. +40% lead follow-up. |
| 11x (Alice / Julian) | Autonomous 'digital workers' for sales development and voice. |
| Clay | Data-enrichment + AI outbound orchestration. |
| Jasper | Enterprise marketing content generation at scale. |
| Regie.ai | Autonomous prospecting/sequencing agents. |
| Qualified (Piper) | AI SDR for inbound website conversion (Salesforce ecosystem). |
| Gong | Revenue-intelligence + AI call analysis and follow-ups. |
| Cognism / Apollo AI | Prospecting data + AI outreach. |
| Copy.ai | GTM workflow agents for marketing/sales ops. |
| Artisan (Ava) | Autonomous AI SDR ('digital worker') for outbound prospecting. |
| Writer | Enterprise generative-AI platform with marketing/GTM agents. |
8 · DevOps, SRE & security
Deploy pattern: read-only access to observability/telemetry + runbook RAG; HITL required before any remediation (least-privilege, see security).
| Company / product | What it is · notable numbers |
|---|---|
| Microsoft Security Copilot | SOC assistant over Microsoft's security signal. |
| Datadog Bits AI | Incident investigation + query assistant across Datadog telemetry. |
| PagerDuty AI | Incident triage, summarization, and response assistance. |
| Cleric | Autonomous SRE agent that investigates alerts. |
| Resolve.ai | Production engineering / on-call investigation agent. |
| Dropzone AI | Autonomous SOC analyst for alert triage. |
| Torq HyperSOC | Agentic security automation. |
| Honeycomb (Query Assistant) | Natural-language observability querying. |
| Cursor/Claude for infra | IaC (Terraform/K8s) generation & review via coding agents. |
| CrowdStrike Charlotte AI | Agentic SOC triage across the CrowdStrike security platform. |
| Traversal | Autonomous AI SRE for root-cause investigation. |
9 · Computer use, browser agents & RPA
Deploy pattern: sandboxed VM/browser + screenshot-vision loop (or DOM), strict per-session isolation (microVM) and HITL for high-risk actions (see Isolation).
| Company / product | What it is · notable numbers |
|---|---|
| OpenAI Operator / Agent | Cloud browser agent that navigates and completes web tasks. |
| Anthropic Computer Use | Claude controls a desktop via screenshots + mouse/keyboard coordinates. |
| Manus | Autonomous general agent running in per-session cloud sandboxes (Firecracker). |
| Google Project Mariner | Gemini browser agent for multi-step web tasks. |
| MultiOn | API-driven autonomous web actions. |
| UiPath Agentic Automation | Agents layered on enterprise RPA workflows. |
| Automation Anywhere | Agentic process automation for back-office ops. |
| Skyvern | Open-source browser automation via vision + LLM. |
| Adept (ACT-1 lineage) | Action models for software UIs (now part of Amazon). |
| Browser Use | Open-source library that lets LLMs drive a real browser (huge GitHub traction). |
| Genspark | Autonomous general agent for multi-step web tasks. |
10 · Data & analytics (text-to-SQL / BI)
Deploy pattern: schema RAG → text-to-SQL → validate & execute against the warehouse → HITL on results; tight guardrails to prevent bad/expensive queries.
| Company / product | What it is · notable numbers |
|---|---|
| LinkedIn SQL Bot | Natural-language → SQL over LinkedIn's data lake, in production. |
| Uber QueryGPT | Internal text-to-SQL cutting query authoring time for thousands of employees. |
| Snowflake Cortex Analyst | Managed text-to-SQL over Snowflake with semantic models. |
| Databricks Genie / AI-BI | Conversational analytics over the lakehouse. |
| Google Gemini in BigQuery | NL data analysis and SQL generation in BigQuery. |
| Microsoft Fabric / Power BI Copilot | NL → DAX/SQL and report generation. |
| Julius AI | Consumer/analyst data-analysis agent over uploaded data. |
| Hex Magic | Notebook-native AI for analysts. |
| Vanna AI | Open-source RAG-based text-to-SQL framework. |
| ThoughtSpot Spotter | Agentic analytics search for business users. |
Notice the same infra recurring: RAG + tools + guardrails + human handoff for knowledge/support/legal/finance/health; sandbox + ReAct loop + self-verification for coding and computer-use; schema-grounded generation + validation for data. The domain changes the tools, the data, and the risk controls — the deployment skeleton barely changes. Master the skeleton (the reference architecture) and you can deploy in any of these verticals.