AI tools and frameworks,
picked per workflow, not by listicle.

The honest answer: it depends on the workflow shape, not the framework's marketing page. Three rules of thumb. (1) Under ~6 steps, no branching, narrow tool use → plain Python with the vendor SDK directly. A framework adds weight you don't need. (2) Branching graph, durable replay, multi-vendor routing → LangGraph. It's the most production-tested orchestration layer in the 2026 stack. (3) Multi-agent role-based scenes → CrewAI for the prototype, then re-platform to LangGraph before production. The $3K audit picks the framework per workflow with your eval data — we'd rather you not build on a framework that's wrong for the shape.

LangChain is the older general-purpose toolkit — useful pieces (loaders, retrievers), but production code on top of it tends to fight the abstractions. LangGraph is the same team's stateful graph framework — what we reach for when the orchestration is non-trivial; vendor-agnostic. CrewAI is role-based multi-agent — great for prototypes of "researcher + writer + reviewer" scenes, brittle as the cast grows. OpenAI Agents SDK is GPT-native — clean ergonomics if your stack is GPT-only and you don't need multi-vendor failover. Most langchain alternatives questions on this SERP collapse to: "use LangGraph for graphs, OpenAI Agents SDK for GPT-only, plain Python for everything under 6 steps." Read the full <a href="/services/ai-agent-development/">AI agent development pillar</a> for our agent engagement model.

We ship all three. Claude vs GPT vs Gemini, briefly: Claude Sonnet 4.6 wins most of our internal coding-agent and long-context tool-use evals — it's the default for production agent work and the model we'd reach for on best llm for coding. GPT-5.4 wins when the workflow needs the Realtime API for sub-second voice, vision pipelines, image gen, or the Codex engineering workflow. Gemini 2.0 wins on massive-context document QA (1M+ tokens beats chunking on some workloads) and tight Google Workspace integration. Open-weights win on sovereign data and cost-floor at scale. The right answer to anthropic vs openai is almost always "both" — pick per workflow on the eval, not per partner badge. See the dedicated pillars: <a href="/services/claude-development/">Claude development</a> and <a href="/services/openai-development/">OpenAI development</a>.

Both, often. The rule we ship: if your retrievable corpus is under ~200K tokens and changes rarely, paste it into the system prompt and use prompt caching — no vector database needed. If it's larger, changes, or needs precise retrieval, you need vector search. The best vector database question is less interesting than the chunking and reranking question — we default to pgvector on existing Postgres for ≤2M chunks (operationally simpler than another vendor) and Pinecone or Qdrant past that. Hybrid BM25 + dense retrieval + Cohere Rerank consistently beats picking a fancier embedding model. The retrieval engagement detail lives on the <a href="/services/ai-integration-services/">AI integration services pillar</a>.

"Enterprise AI platform" is the analyst-coined umbrella for everything in our 5-layer stack glued together with audit logs, RBAC, BAA paperwork, and a single vendor invoice. Real enterprise ai tools today sit in three buckets. (1) Hyperscaler stacks — AWS Bedrock, Azure OpenAI, Google Vertex — when compliance posture (HIPAA, SOC 2, FedRAMP) forces the buy. (2) Managed agent platforms — Lindy, Stack AI, Glean — when the workflow is narrow and the build doesn't earn its keep. (3) Custom stacks like ours — when the workflow is unique enough that platforms can't shape it. Most enterprises we work with end up with all three: a hyperscaler for infra, a SaaS for narrow workflows, and a custom build for the high-leverage one. The audit picks the bucket per workflow.

AI observability and llm observability is the layer most listicles skip and the layer where production AI lives or dies. Four pieces. (1) Tracing — every prompt, completion, tool call, and cost stamped with a trace ID. Langfuse for OSS-friendly self-host; Braintrust when the team wants the eval harness in the same tool. (2) Eval suites — golden-set regression run on every prompt change and every model swap. Inspect AI for serious harnesses; hand-rolled pytest for narrow workflows. (3) Drift monitoring — shadow-mode mirroring before cutover, production sampling + LLM-as-judge scoring after. (4) Cost telemetry — per-workflow, per-customer dashboards wired in at the same layer as latency. Without the eval set, you don't know when the model swap regressed quality. Without cost telemetry, a runaway prompt loop costs $9K before anyone notices.

Three engagement tiers, no surprises. A one-week ai infrastructure audit is $3,000 — workflow inventory, model + framework pick per workflow, token-cost projection, observability spec, 90-day roadmap. A pilot is $10,000–$25,000 fixed price, 4–8 weeks — one workflow shipped end-to-end on the chosen stack with eval, monitoring, and runbook. A continuous AI team is from $5,000 per month — embedded PM + AI engineer + ops analyst, monthly cost-of-ownership reporting, and quarterly re-evaluation of the tooling. Per-workflow run cost at steady state typically lands at $200–$1,500/month depending on volume and which model tier the workflow uses. Listicles avoid pricing because they don't sell the build. We do, so the numbers are on the page.

Three honest cases. (1) The workflow is under ~6 steps with no branching — plain Python and the vendor SDK is cleaner, easier to debug, and faster to ship than LangGraph. (2) An off-the-shelf product covers 80% of it. If Lindy, Stack AI, Glean, Zapier AI, or a vertical SaaS does the workflow for $50/seat/month, buy it and save the build budget for what's actually unique. (3) The problem isn't AI-shaped. Regex, rule engines, classical ML (XGBoost, scikit) outperform LLM calls on cost and quality for plenty of workloads — form parsing on a single PDF type, structured ETL, narrow classification. The <a href="/services/ai-consulting/">AI consulting audit</a> exists partly to tell you when the answer is "don't build with an AI framework at all."