AI Stack 2026: What You Actually Need
Vendors will sell you everything. Here's what it actually takes to run working AI in your company.
"You need an MLOps platform, a data lake, a feature store, a model registry, a vector database, LLMOps..." That's how a typical conversation with a vendor goes. In reality, most companies need far less.
Minimum Viable AI Stack
This is about the internal stack, not the sales tooling. For the GTM side I laid out a separate map of the autonomous funnel, 138 tools across seven stages.
Before you buy yet another tool, check whether you have the fundamentals:
LAYER 1: DATA
- Data Warehouse: Snowflake, BigQuery, or PostgreSQL for small companies.
- ETL: Airbyte, Fivetran, or dbt for transformations.
- Vector DB (if RAG): Pinecone, Weaviate, pgvector.
LAYER 2: MODELS
- LLM API: OpenAI, Anthropic, Google.
- Orchestration: LangChain, LlamaIndex, or your own code.
- Prompts: Version them in Git, not in a spreadsheet.
LAYER 3: APPLICATION
- Backend: FastAPI, Node.js, whatever you know.
- Frontend: React, Vue, or integration with your existing app.
- Auth & Security: Don't reinvent the wheel.
Before you pick an orchestrator, check whether you actually need several agents. I wrote separately about when MAS makes sense and when one agent is enough.
What you DON'T need at the start
Seriously, don't buy these before your first working pilot:
- ✗ MLOps platform: Git plus simple CI/CD is enough.
- ✗ Feature store: needed at scale, not at the start.
- ✗ Model registry: at the beginning you have one or two models.
- ✗ Enterprise AI platform: 6-figure cost, 12-month rollout.
"The best stack is the one you know. Don't learn Kubernetes just to ship a chatbot."
This is what I do hands-on: advising on AI strategy and building agents that survive the demo.
Build vs Buy: A decision framework
Every component of the stack is a build vs buy decision:
BUY (almost always)
- • LLM: don't train your own language model.
- • Infrastructure: cloud, not your own servers.
- • Auth: Auth0, Clerk, not a homegrown solution.
BUILD (usually)
- • Domain logic: nobody knows your business better.
- • Legacy integrations: vendors don't support your 2005 ERP.
- • Prompts and workflow: this is your IP.
DEPENDS
- • Orchestration: LangChain vs your own code (depends on how complex your flow is).
- • Vector DB: managed vs self-hosted (depends on scale).
- • Monitoring: custom vs off-the-shelf tool (depends on team maturity).
That last row is more important than it looks: whatever you pick for monitoring has to count retries and fallbacks. Otherwise you are buying yourself silent failures.
Stack evolution
An AI stack grows along with your needs. Here's the typical path:
- Phase 1 (Pilot): OpenAI API + a Python script + a spreadsheet to track results.
- Phase 2 (MVP): you add a database, a simple frontend and basic monitoring.
- Phase 3 (Production): CI/CD, proper logging, alerting, backups.
- Phase 4 (Scale): now consider an MLOps platform, a feature store and the rest.
Real costs
For a mid-sized company (100 to 500 employees), a realistic AI stack can cost:
MONTHLY INFRASTRUCTURE COSTS
- LLM API: 500 to 5,000 PLN (depends on volume).
- Cloud (compute, storage): 1,000 to 3,000 PLN.
- SaaS tools: 500 to 2,000 PLN.
- Total: 2,000 to 10,000 PLN/month.
This is not a 6-figure budget. But it does require a 6-figure budget for the people who build and maintain it. What that second half of the bill is actually made of is broken down in the post on a quote split into six line items.
Where to start
The most expensive tool is the one you don't use. Start with the OpenAI API and a single Python script. Add the next tool only once a concrete problem actually starts to hurt.
I've deployed AI at mid-sized companies. That's where this comes from.
