The vocabulary we use.

Salesforce's AI agent platform, successor to Einstein. Creates agents that consume Data Cloud as their context source and act over Sales, Service, and the remaining clouds. It's not a chatbot — it's an execution layer.

Salesforce's operational context layer. Materializes profiles in real time, feeds Agentforce, Marketing, and Sales. It's not just a CDP — it's a foundation. Treating it as "just a CDP" leaves value on the table.

Salesforce's sales core. Pipeline, opportunities, accounts, contacts, commercial process automation. It was the platform's original product; today it works best when integrated with Data Cloud and Agentforce.

Salesforce's customer service product. Cases, omnichannel, SLAs, knowledge base. Well implemented, it becomes the common language between operations, quality, and product.

Platform to unify customer identity, segment, and activate in marketing channels. Product category that Data Cloud transcends — still useful as a concept, but insufficient as a definition of the data layer.

Both Data Cloud's harmonized data model (canonical objects like orders, cases, contacts) and Salesforce's strategic thesis — every interaction happens over the same unified profile.

Mandatory step before any Salesforce configuration. Documents steps, exceptions, rules, and SLAs in a format readable by both non-expert humans and LLMs. Skipping it is the largest source of rework in CRM projects.

Data transformation tool that became the de facto standard in analytical modeling. Uses SQL + Jinja + tests + documentation as code. The real win isn't the model — it's the versioned documentation.

ETL extracts, transforms, then loads — bottlenecking on intermediate servers. ELT loads raw into the warehouse and transforms there — using the engine that already exists. The inversion is structural, not aesthetic.

Typical stack today: ingestion (Fivetran/Airbyte) → warehouse (Snowflake/BigQuery/Databricks) → transformation (dbt) → BI (Tableau/Looker). Not a prescription — common vocabulary among data engineers in 2026.

Warehouse: structured data, SQL queries, analytics focus. Lake: raw data in varied formats, cheap-storage focus. Lakehouse tries both — works when the problem justifies it, hurts when it's the default answer.

Formal agreement between producer and consumer of a data asset: schema, quality guarantees, change policy. Reduces the "pipeline broke because someone renamed a column" — the most common problem in data teams.

Modeling approach (Kimball) with fact and dimension tables. Still valid in 2026 — not as the only architecture, but as the pattern for the analytical consumption layer.

Culture where business people create their own analyses from trustworthy models. Fails when the model is bad — each department creates its "final draft" and the company loses single source of truth.

Visualization and BI platform (Salesforce). Excellent for rapid visual exploration over well-modeled data. It doesn't replace modeling or business decision — it visualizes, doesn't think.

Pattern where an LLM retrieves relevant snippets from an external base before generating a response. Reduces hallucination and enables use of up-to-date knowledge. The hard part is "retrieving well", not generating.

Language model trained on vast text corpora (GPT, Claude, Gemini). Useful as a reasoning layer in AI pipelines; dangerous as the sole source of truth. Treat as a very well-informed junior human.

Numerical representation (vector) of text, image, or other data. Enables semantic similarity calculation. Foundation of vector search and RAG. The quality of the embedding determines the quality of retrieval.

Database optimized for vector similarity search (Pinecone, Weaviate, pgvector). Typical RAG component. Choice depends on scale, latency, and existing tooling — not hype.

Training a ready-made LLM further with domain-specific data. Expensive, slow, rarely the first move. RAG and prompt engineering usually deliver 80% of the value at 10% of the cost.

Discipline of writing instructions for LLMs that produce reliable outputs. Versioned, tested, and measured — not improvised. Almost always the first experiment before RAG or fine-tuning.

Process of measuring agent performance on real cases and edge cases. Includes accuracy, error types, escalation rate, cost per interaction. Without rigorous evaluation, an agent in production becomes silent liability.

LLM response that sounds plausible but is factually wrong. RAG and prompt engineering reduce; never eliminate. That's why governance and human-in-the-loop are part of the design, not overhead.

Set of practices for operating AI responsibly: logs of every interaction, audit of decisions, kill switch, retention policy, incident process. It's not overhead — it's what separates a project that survives a new board.

Kliente 360 engagement model for AI projects — four weeks of Map, Prototype, Validate, and Decide. Low-commitment entry to validate the thesis before a long project.

Kliente 360's delivery method. Five verbs applied across the three practices: Map, Prototype, Validate, Deploy, Sustain. A hybrid of strategic consulting, technology project, and rapid AI deployment.