TL;DR
We surveyed 87 Latin American enterprises (Mexico, Colombia, Argentina, Brazil, Chile) that have at least one LLM agent in production. Only 23% report clear ROI. The most-used framework is LangGraph (38%) followed by CrewAI (21%) and proprietary/custom (19%). For vector DB, Qdrant (34%) overtook pgvector (27%) and Weaviate (14%). For observability, Langfuse leads with 41% — tripling year-over-year. The three top barriers: integration with existing systems (46%), inference costs (38%), lack of systematic evaluation (34%). This report publishes the full data, sector conclusions, and the cross-tabulation between chosen stack and likelihood of reporting ROI.
Methodology
- Sample: 87 companies, 15 to 3,200 employees.
- Geography: Mexico (38), Colombia (19), Argentina (13), Brazil (11), Chile (6).
- Industries: fintech (24), healthcare (13), retail/e-commerce (17), legal/services (10), B2B SaaS (14), gov/edu (9).
- Inclusion: at least one LLM agent in production for 3+ months.
- Instrument: 40-question survey + 45-min interview with the technical lead.
- Field period: January-March 2026.
- Anonymous aggregated data; the curated dataset is published under CC-BY at
github.com/numoru-ia/latam-ai-agents-survey-2026.
Profile of deployed agents
Main use cases
| Use case | % companies |
|---|---|
| Customer support (text) | 58% |
| Internal content generation | 47% |
| Sales / SDR assistant | 31% |
| Internal-document search agent (RAG) | 44% |
| Backoffice operations automation | 28% |
| Voice agent (call center, reception) | 14% |
| Compliance / legal review | 9% |
| Code agent | 22% |
| Orchestrated multi-agent | 17% |
Companies average 2.3 agents each; that's why percentages exceed 100.
Deployment maturity
- 43% are prototypes exposed to <100 internal users.
- 38% are in production but in a single channel or team.
- 19% are multi-channel production with business metrics connected.
Only the last third reports measurable ROI.
Stack: what LATAM actually uses
Orchestration and frameworks
| Framework | % primary use |
|---|---|
| LangGraph | 38% |
| CrewAI | 21% |
| Custom (Go/TS) | 19% |
| AutoGen | 8% |
| Mastra (TS) | 7% |
| Pydantic AI | 5% |
| Other (Haystack, Semantic Kernel) | 2% |
LangGraph dominates regulated sectors (they need a checkpointer, replay and fine-grained controls). Custom shines in fintechs and senior engineering teams that already have robust Go infrastructure.
Vector databases
| Vector DB | % primary use |
|---|---|
| Qdrant | 34% |
| pgvector (Postgres) | 27% |
| Weaviate | 14% |
| Pinecone | 11% |
| Chroma | 6% |
| Milvus | 5% |
| Other | 3% |
Qdrant overtook pgvector for the first time in our panel. The cited reason: performance at >1M vector volumes and native hybrid-search support. pgvector remains strong on teams with a mature DBA function and volumes <500k.
Observability
| Tool | % primary use |
|---|---|
| Langfuse (self-hosted or OSS) | 41% |
| LangSmith | 18% |
| Phoenix (Arize) | 9% |
| OpenLLMetry + Grafana | 8% |
| Helicone | 5% |
| Nothing formal | 14% |
| Other | 5% |
Langfuse tripled share year over year. The reason: Apache 2.0 license, self-host option (critical for compliance), and an integration ecosystem that matured in 2025.
The 14% reporting "nothing formal" is almost identical to the group not reporting ROI. Interesting correlation.
Most used models
| Model (by inference volume) | % |
|---|---|
| Claude Sonnet | 34% |
| GPT-4o / GPT-4o-mini | 28% |
| Claude Haiku | 12% |
| Llama 3.3 (local) | 9% |
| Gemini | 8% |
| Qwen 2.5 (local) | 4% |
| Other | 5% |
Claude dominates regulated sectors due to retention policies and strong tool-use support. GPT leads voice agents and consumer chatbots. Local models (Llama, Qwen) grow fast in healthcare and legal.
Other stack pieces
- Redis as semantic cache or memory: 52% use it.
- LiteLLM as gateway: 29% (grew fast in 2025).
- MCP in production: 18% (vs 2% a year ago).
- Mem0 / Zep for semantic memory: 14%.
- Temporal / Inngest for durable flows: 11%.
- Guardrails (NeMo, Guardrails AI): 23%.
ROI: what separates the 23% from the rest
Comparing the "clear ROI" group against the rest, five factors appear with statistical significance:
1. They have systematic evaluation (p < 0.01)
86% of the ROI group has automated evals in CI/CD. Of the rest, only 19%. Moral: without measuring, you don't improve.
2. Production observability (p < 0.01)
95% of the ROI group uses Langfuse, LangSmith or Phoenix. Of the rest, 52%. Visibility in production is the difference between fixing in time and accumulating debt.
3. Tiered memory stack (p < 0.05)
58% of the ROI group combines working memory + sessions + semantic memory. Of the rest, 21%. Agents without structured memory lose relevance from turn 10-15 on.
4. Production guardrails (p < 0.05)
48% of the ROI group runs active guardrails (content, topic, PII). Of the rest, 17%. Fewer user-visible errors = more operational trust.
5. Single agent owner (p < 0.05)
74% of the ROI group has a named technical-functional owner (an AI PM, AI Ops lead or similar). Of the rest, 36%. Without an owner, the agent degrades and nobody notices.
Direct conclusion: it's not the framework. It's operational discipline. A team with Langfuse + evals + tiered memory + guardrails + an owner can use any framework and do well.
Main barriers (self-reported)
| Barrier | % companies mentioning |
|---|---|
| Integration with existing systems | 46% |
| Inference costs | 38% |
| Lack of systematic evaluation | 34% |
| Data quality / preparation | 32% |
| Compliance / regulation | 27% |
| Scarce talent | 24% |
| Uncontrolled hallucinations | 22% |
| Latency | 19% |
| Internal adoption | 18% |
| Infrastructure cost | 12% |
Integration with existing systems is #1 by a clear margin. That's why our recent work points to standardized MCP templates: standardizing that layer offers the highest ecosystem-level leverage.
By sector
Fintech (n=24)
- Modal stack: LangGraph + Qdrant + Langfuse + Claude Sonnet.
- Main use: assisted scoring, customer support, KYC.
- Reported ROI: 38%. Leads the panel.
- Differentiator: they already had MLOps; adding LLMOps was incremental.
Healthcare (n=13)
- Modal stack: LangGraph + Qdrant + Ollama (Llama 3.3 local) + Langfuse.
- Main use: clinical summaries, triage, documentation.
- Reported ROI: 15%. Slower due to compliance.
- Pain: HIPAA + LFPDPPP + (if exporting) AI Act. 100% on-prem stack for sensitive cases.
Retail / e-commerce (n=17)
- Modal stack: CrewAI + pgvector + LangSmith + GPT-4o-mini.
- Main use: customer support, recommendations, product descriptions.
- Reported ROI: 24%.
- Pain: high inference volume = worrying costs. Redis semantic cache is now a strong trend.
Legal / services (n=10)
- Modal stack: Custom Go/TS + Qdrant + Langfuse + Claude Opus.
- Main use: drafting, review, doc search.
- Reported ROI: 30%.
- Pain: trust. Heavy investment in evals and verifiable citations.
B2B SaaS (n=14)
- Modal stack: LangGraph + Qdrant + Langfuse + Claude Sonnet.
- Main use: feature embedded in their own product (in-app assistant).
- Reported ROI: 21%.
- Pain: differentiating their assistant when Cursor/Claude are now commodity.
Gov / education (n=9)
- Modal stack: mixed; many on GPT-4o with little custom stack.
- Main use: educational assistants, citizen services.
- Reported ROI: 11%. The lowest.
- Pain: long procurement cycles, lagging infrastructure.
Reported primary orchestration framework per company. Multiple responses allowed in sub-projects, primary reported here.
Numoru LATAM AI Agents Survey 2026. Dataset CC-BY.
Primary vector DB per company. Claude / pgvector lead differently depending on data volume.
Numoru LATAM AI Agents Survey 2026.
2026-2027 predictions (based on the panel)
- Langfuse exceeds 60% share in the panel before Q4 2026 — self-host becomes default for compliance.
- MCP moves from 18% to >40% in the panel — standardization accelerates adoption.
- Local models (Llama, Qwen) cross 25% combined, driven by costs, privacy and tool maturation (Ollama + vLLM).
- CI/CD eval pipelines become standard practice: they correlate so strongly with ROI that serious teams stand them up day 1.
- Two dominant patterns: a) single agent with MCP + tiered memory (SMBs), b) multi-agent orchestrated by LangGraph + Temporal (mid-to-large companies).
Actionable recommendations
If you're starting (0 agents in production)
- Pick a use case with a clear metric (e.g. response time, CSAT).
- Initial stack: LangGraph + Qdrant + Langfuse + Claude Sonnet + Redis.
- Stand up evals and observability from day 1.
- Single named owner.
If you have 1-2 agents in prototype
- Resolve integration with the system of record (CRM, ERP) via MCP.
- Introduce basic guardrails.
- Migrate from notebook/script to a formal orchestrator.
If you have production but uncertain ROI
- Observability audit — are there full traces?
- Review prompt and memory — tiered memory or retire.
- Implement regression evals in CI/CD.
Business & commercial impact
What the report sells
This data is a product on its own. There are four ways Numoru commercialises it: the premium 80-page PDF for investors and ops teams, a benchmarking module where a customer compares itself against the panel, a licensed research subscription for consultancies that want the raw CSV, and the lead generation that feeds every other Numoru service (RAG rescue, GEO, AI Act, voice). Enterprise analysts pay multiples of the ticket for McKinsey / Gartner reports that are less specific to LATAM.
Who buys research like this
Research pricing by buyer (Numoru, 2026)
Comparable public benchmarks
McKinsey — State of AI (annual)
LangChain — State of AI Agents
Illustrative case — research subscription reseller
Boutique consultancy (Bogotá) licensing the research to layer into their own strategy practice
ROI calculator — enterprise licensing the panel
CIO office benchmarking against the 87-company panel (12 months)
| Report + engagement fee | −$9,500 |
| Internal time | −$4,400 |
| Reallocated over-spend captured | +$57,600 |
| Better stack choices (12% × $320k) | +$38,400 |
| Vendor-negotiation leverage | +$24,000 |
| Net year-1 contribution | +$106,100 |
Pricing tiers
- TL;DR + methodology
- Core stat highlights
- Charts embeddable with citation
- Link to CC-BY dataset
- Cross-tab: stack × ROI
- Per-sector breakdowns
- 12 appendix charts
- Raw CSV under CC-BY-NC
- Citation license for internal use
- Stack & cost comparison report
- 60-min exec workshop
- Custom 15-page Miro board
- Vendor-negotiation playbook
- Follow-up call at 90 days
- 1-year re-benchmarking rights
Reseller and vendor-partnership licenses quoted separately. Contact sales.
FAQ
Is the sample representative?It's the largest available LatAm sample on this topic. It's not probabilistic (LinkedIn recruitment + client referrals), so conclusions should be read as directional, not population statistics.
How comparable are the ROIs across companies?We asked for "clear ROI as defined by the company" without imposing a formula. Some measure tickets reduced, others conversion, others NPS. The 23% is the group that could defend numbers in front of their committee.
Why does Qdrant beat pgvector now?Data volume: when collections cross 1M vectors or hybrid search is required, Qdrant's Rust architecture scales better. pgvector remains ideal under 500k and on teams already operating Postgres.
Claude vs GPT-4?Each dominates niches. Claude in sectors with strong tool-use needs and strict retention policies. GPT for voice and consumer chat. Many teams use both via LiteLLM with fallback.
Does the OSS stack really compete with SaaS?Yes when the team has 1 person with DevOps experience and volumes justify the fixed cost. Below 50k inferences/month, SaaS usually wins on simplicity.
Next steps
The full dataset and reproducible script are at github.com/numoru-ia/latam-ai-agents-survey-2026. If your company wants to participate in the 2026-H2 edition (field September-October), the form is open. We'll publish additional segments if sample size allows: foodtech, edtech, proptech and logistics.