Every company answering an AI survey says it's "investing in AI". Job postings can't bluff: a req for an engineer to "migrate our RAG pipeline off LangChain" is budget, a headcount, and a dated commitment — published in public. So instead of asking anyone, we read the hiring record.
We analyzed 4.7 million currently-active job postings across 761,828 companies, extracted the technologies each one mentions, and — the part that matters — classified the direction of change: is the company using the technology, adopting it, evaluating it, or replacing it? Filtered to the AI build stack — retrieval, agent frameworks, model platforms, AI dev tools — that gives a live census of who is actually building, not who is talking about it. Three findings stood out.
The build-out doubled in six months — and it's still accelerating
In September 2025, 0.32% of all new job postings mentioned hands-on work with the AI build stack. By June 2026 it was 0.71% — 2.2× in nine months, with the slope steepening in Q2. New-adoption signals (postings that describe introducing a technology, not maintaining it) grew even faster: from 0.030% to 0.074% of all postings.
Two things make this a hard signal. First, it's paid for: every posting in the series is an open req with budget behind it. Second, it's current — 87% of the active adoption signals in this report come from postings published within the last 90 days. When the migration ends, the postings close and the signal decays by construction. This is not a survey of intentions from last year.
AI building escaped the tech sector — half the RAG wave is banks, pharma, and factories
The lazy assumption is that "companies building with AI" means startups in San Francisco. The hiring record disagrees: 51% of companies currently adopting RAG are outside the software and internet industries. The single largest non-tech vertical is finance — 190 banks, insurers, and financial-services firms have an active AI build signal — followed by healthcare, telecom, automotive, and defense.
And these aren't pilot-project press releases. Morgan Stanley is hiring for LangChain, LangGraph, Semantic Kernel, CrewAI and RAG work simultaneously. Ford and AB InBev are staffing LangGraph agent teams. GM is hiring for vLLM — self-hosted inference, several layers deeper than "we bought Copilot licenses". Nike, Capital One, Cigna and J.Crew are rolling out Cursor to their engineering orgs. John Deere and 3M are building RAG systems.
| Adopting | Companies |
|---|---|
| RAG | Microsoft · Uber · SAP · Huawei · Santander · Itaú Unibanco · Morgan Stanley · GSK · Airbnb · 3M · John Deere |
| MCP | DHL · Uber · Orange · Crédit Agricole · Woolworths · AMD · Bloomberg · Synopsys · Fortinet · FactSet · Just Eat Takeaway |
| LangGraph | Ford · AB InBev · Itaú Unibanco · Morgan Stanley · Broadcom · Mapfre · Santander Brasil · Nielsen · Indra Group · URBN |
| Cursor | Nike · Capital One · DoorDash · Grab · Cigna · Urban Outfitters (URBN) · Wolters Kluwer · Palo Alto Networks · HelloFresh · J.Crew |
The stack nobody predicted: no framework, boring retrieval, and agents growing up
Ask social media what a RAG stack looks like and you'll get LangChain plus a vector database. The hiring record describes a different building: 56% of companies adopting RAG show no LangChain, LlamaIndex, or LangGraph anywhere in their stack — they're wiring retrieval directly against model APIs. And the retrieval layer they reach for first is Elasticsearch, not a vector-native database.
The dedicated vector databases tell the second half of that story. Their usage is broad — Pinecone shows up in 1,285 companies' stacks — but new adoption has nearly stopped: 14 companies adopting Pinecone, 9 Weaviate, 3 Qdrant, 2 Milvus. The vector-DB land grab of 2023–24 is over; retrieval became a feature of infrastructure companies already own.
Where adoption is very much alive is the agent layer — and it's growing up. LangGraph is the most enterprise-weighted technology in the entire AI stack: 46% of its adopters have over 1,000 employees (Ford, Morgan Stanley, Itaú Unibanco, Broadcom). It's also cannibalizing its own ecosystem — 52 of its 131 adopters already run LangChain and are moving up the stack. CrewAI and n8n live in smaller companies; Cursor spreads bottom-up through the mid-market before the enterprise signs off.
| Technology | Adopting | Evaluating | Using | Non-tech |
|---|---|---|---|---|
| RAG | 386 | 157 | 4,396 | 51% |
| Cursor | 320 | 98 | 4,665 | 41% |
| LangChain | 211 | 156 | 4,272 | 46% |
| LangGraph | 131 | 61 | 2,204 | 53% |
| n8n | 104 | 68 | 3,242 | 60% |
| MCP | 99 | 37 | 1,118 | 33% |
| Copilot Studio | 69 | 24 | 1,099 | 75% |
| Codex | 66 | 26 | 1,310 | 50% |
| CrewAI | 64 | 34 | 932 | 50% |
| Azure OpenAI | 57 | 30 | 1,014 | 56% |
| LlamaIndex | 51 | 56 | 1,447 | 45% |
| AutoGen | 49 | 34 | 800 | 49% |
| Vertex AI | 48 | 51 | 1,588 | 69% |
| AWS Bedrock | 39 | 25 | 742 | 67% |
| Agentforce | 34 | 13 | 332 | 56% |
Take the dataset with you
The free CSV has the 50 largest companies with an active AI build signal — name, domain, industry, size, and which technologies they're adopting. The full slice (1,273 companies × full stack context) is available on request, and we'll run a custom pull against your own account list at no cost.
How this was measured
Echoloc continuously indexes public job postings and company pages; at analysis time the index held 15,207,852 postings, of which 4,747,583 were active, across 761,828 companies. An LLM extraction pipeline reads each posting and records every technology mentioned (15,193 distinct) together with its context: using, adopting, evaluating, or replacing. Signals aggregate to company level from active postings only — when the hiring ends, the signal expires with it.
- AI build stack here means 20 technologies across retrieval (RAG), agent frameworks (LangChain, LangGraph, CrewAI, AutoGen, LlamaIndex, Semantic Kernel), protocols (MCP), model platforms (Azure OpenAI, Vertex AI, AWS Bedrock), AI dev tools (Cursor, Codex, Windsurf, Devin), and agent/automation builders (n8n, Copilot Studio, Agentforce, vLLM, LangSmith).
- Excluded from company counts: consulting, IT-services, staffing, and outsourcing industries (10 LinkedIn categories) — their postings describe client projects, not their own stack. Named professional-services firms are additionally filtered from company tables. Residual misclassification is possible where a consultancy self-reports as another industry.
- Out of scope: consumer assistants (ChatGPT, GitHub Copilot, Claude as products) — postings mention them as tools of daily work, not systems being built; our extractor doesn't track them as build-stack technologies. Model-provider names (OpenAI, Gemini, Anthropic) appear in the data but are excluded from rankings due to marketing-use contamination.
- Momentum series uses shares of postings published per month, not raw counts — index volume grew unevenly over the period. Months before Sep 2025 are omitted (low volume).
- Verification: we spot-check extraction against raw posting text; in this report's core signals, 97–99.7% of "adopting/replacing" labels co-occur with explicit corresponding language in the posting.
Questions about the method, or a number you'd like re-cut? [email protected] — happy to share query logic.