How to adapt your company to an information- and collaboration-centred culture

December 1, 2025 · Baoss
BAOSS IT consulting: data-driven culture with AI agents in the enterprise

Data-driven culture 2026: generative AI and autonomous agents

In June 2015 we published a pioneering article on this blog. It was about how to adapt the company to a culture centred on information and collaboration. We talked about Big Data, traditional Business Intelligence and ETL tools. We also talked about the Apache Spark promise: processing data 100 times faster than Hadoop MapReduce. The reference startup was ClearStory Data —today absorbed by Alation— and its ‘Interactive Storyboards’ concept for breaking silos. Ten years later, the BAOSS IT consulting team returns to that same ground with a new thesis.

Ten years have passed. The problem hasn’t changed: organizations are still drowning in data yet thirsting for decisions. However, the available technology and deployment speed have changed radically. So has the profile of who can analyze. In 2025, the bottleneck is no longer processing, because Spark solved it. Nor is it visualization, which Tableau, Power BI and Looker democratized. Today’s bottleneck is the cognitive last mile. That is, turning data into actionable context before the decision window closes.

The real problem in 2026-2027: it’s not a lack of data, it’s an excess of cognitive friction

Take the Data and AI Leadership Executive Survey 2025 by Wavestone (a survey of 125 Fortune 1000 CDOs). There, 78% of organizations report having invested in the ‘modern data stack’ (Snowflake, Databricks, dbt, Fivetran). Yet only 23% consider that their non-technical employees make data-based decisions routinely. The gap is not technological: it is operational and cultural.

  • Semantic fragmentation: 64% of analysts spend more than 50% of their time ‘translating’ metric definitions between departments (Finance vs Marketing vs Operations).
  • Decision-data latency: the average time from a business question arising to a validated answer is 11.3 days (Forrester, The State of Data-Driven Decision Making 2026).
  • Dependence on scarce profiles: only 12% of complex analytical queries are solved by a data scientist; the rest queue up or get answered by intuition.
  • Reactive governance: data catalogs (DataHub, Amundsen, Purview) document what existed yesterday, not what the business needs today.

Indeed, at BAOSS we see it in every engagement: companies have mature data lakes, governed lakehouses and impeccable CI/CD pipelines. Even so, there are executives who keep asking for ‘an Excel with yesterday’s number’. The reason is that they don’t trust the dashboard or don’t know how to phrase the question in SQL.

Why traditional BI (and classic ‘self-service’) has failed

The classic model —IT models, business consumes— assumes business questions are predictable and stable. In 2025, they are not. For example, a supply chain director must decide now whether to reroute a container. To do so, they need to cross IoT sensor data (frequency: seconds), weather forecasts (update: hourly) and spot freight rates (volatility: minute). Also, they must consider supplier behaviour (history: years). Consequently, no prebuilt dashboard covers that ad hoc combination.

The ‘self-service’ of 2015-2020 (drag-and-drop in Tableau/Power BI) moved the bottleneck. Before, you waited for IT. Now you wait for the analyst who knows DAX/MDX and understands the semantic model. Thus, the business user remains out of the loop. The cost of that gap? AI productivity: how to close the 9 gap in your company measures it.

The 2026 paradigm shift proposed by BAOSS IT consulting: autonomous AI agents + semantic RAG + MCP

As ambitious as it sounds, the architecture that is working in production —not in PoCs— combines three layers:

  • Unified semantic layer (Semantic Layer 2.0): not a passive catalog. It is a living knowledge graph mapping metrics, dimensions, lineage, owners, freshness SLAs and —key— business definitions in natural language. Tools: dbt Mesh + DataHub + MCP (Model Context Protocol) to expose context to LLMs.
  • Reasoning layer: multi-step autonomous agents (LangGraph, CrewAI, AutoGen, or OpenAI Assistants API v2 / Claude 4 Tool Use). They don’t just generate SQL; they also plan: decompose the question, identify sources, validate quality, execute, interpret, propose next best action and cite sources.
  • Hybrid execution layer: vLLM / Ollama for local inference of open models (Llama 3.1 70B, Nemotron 3 Ultra, Qwen 2.5 72B) when latency or data sovereignty demand it; GPT-4o / Claude 4 Sonnet API for complex reasoning and executive narrative generation.

The result is a ‘senior virtual analyst’. It understands ‘why did margin fall in Mexico last week?’ and answers in 47 seconds. The answer includes a breakdown by SKU/channel/supplier and correlation with the exchange rate and the port strike. It also includes an impact simulation had an alternative supplier been activated. And a link to the contingency runbook. Everything is auditable, versioned and governed.

BAOSS cases (anonymized): from 11 days to 47 seconds

Case 1: Multinational retail — 2,400 stores, 18 countries

Challenge: country executives requested weekly ‘promotional performance’ reports that arrived 9 days late. Also, ‘promotional margin’ definitions differed across 12 markets. Stock repurchase decisions were made with stale data.

Deployed solution (14 weeks, mixed BAOSS + client team):

  • Unified semantic layer in dbt Mesh + DataHub: 347 standardized metrics, full lineage from ERP (SAP) / POS / eCommerce / loyalty.
  • LangGraph + GPT-4o agent with RAG over the semantic catalog + operational runbooks. The agent doesn’t just query Snowflake; it also retrieves the ‘urgent repurchase’ playbook and validates that the recommendation complies with the safety-stock policy.
  • Interface: Slack + Teams + web portal. Natural-language question → narrated answer + table + chart + SQL traceability + proposed action (create a PO in Ariba with one click).
  • Governance: MCP exposes metadata, permissions (RBAC/ABAC) and freshness SLAs to the agent. Full auditing in Immuta.

Metrics at 6 months post-go-live:

  • 92% reduction in response time (9.2 days → 4.1 hours on average; routine queries < 60 sec).
  • Adoption: 68% of country executives use the agent weekly (vs 12% who opened the previous dashboard).
  • Business impact: +3.7% promotional margin from agile repurchase of winners and quick cut-off of losers.
  • ROI 3.2x in 6 months (savings on obsolete stock + captured sales + 4.2 analytical FTEs freed and reassigned to predictive modeling).

Case 2: Logistics operator — fleet of 12,000 vehicles, 4 hubs

Challenge: route planners decided with static spreadsheets. Meanwhile, data lived in 7 systems with no common lineage. There coexisted telematics (GPS, fuel consumption, driving), historical traffic, delivery windows, driving-time regulation and driver availability. On-time delivery remained stuck at 89%.

Solution (10 weeks):

  • Streaming ingestion (Kafka → Iceberg on S3 → Trino) + incremental dbt for near-real-time unified views (3 min latency).
  • CrewAI agent (3 agents: Planner, Executor, Critic) + Claude 4 Sonnet for spatial-temporal reasoning. The agent proposes dynamic rerouting on incidents, simulates SLA and cost impact, and explains in natural language (‘Rerouting truck #4421 via A-2 saves 22 min but +€18 in tolls; driver meets rest requirements’).
  • Human-in-the-loop: the planner validates with one click; the action executes in the TMS (Blue Yonder) via API.
  • Local models (Ollama + Llama 3.1 70B 4-bit) for ETA and consumption inference at the edge (gateways in hubs), reducing latency and cloud cost.

Metrics at 4 months:

  • On-time delivery: 89% → 96.3%.
  • Cost km/delivery: -11.4% (fewer empty km, better consolidation).
  • Planner decision time: 45 min → 3 min per incident.
  • Edge deployment: 0 latency incidents across 18,000 decisions/day.

BAOSS 2026 reference architecture: components and key decisions

Layer Chosen technology Why (2025)
Storage / Compute Databricks (Unity Catalog) / Snowflake / Trino + Iceberg Open table formats, ACID, time travel, unified governance
Semantic Layer dbt Mesh + DataHub + own MCP Server Versioned definitions, lineage, standard LLM exposure
Agent orchestration LangGraph (stateful, cycles, native human-in-the-loop) Deterministic control + LLM flexibility; LangSmith observability
Cloud LLM GPT-4o (complex reasoning) / Claude 4 Sonnet (200k window, robust tool use) Best score on SQL + reasoning benchmarks (Spider 2.0, BIRD)
Local / Edge LLM vLLM (throughput) + Ollama (simplicity) → Llama 3.1 70B / Nemotron 3 Ultra Sovereignty, latency < 200ms, fixed cost vs per-token
RAG / Retrieval LanceDB / Qdrant + reranker (Cohere Rerank 3.5 / bge-reranker-v2) Hybrid search (vector + BM25) + rerank = accuracy > 92% on catalog QA
Evaluation / Guardrails Ragas + LangSmith + custom evals (SQL correctness, policy compliance) CI/CD for prompts: no change to prod without > 95% pass rate
Interface Slack/Teams bot + React portal (CopilotKit) + REST API Where the user works; progressive disclosure of technical detail

Technology only performs when the data-driven culture and governance operations support the architecture. It is a requirement, not an extra. That is why at BAOSS we don’t deliver just agents. We also deliver the organizational change that makes them useful. The deployed cases show it: from 11 days to 47 seconds. And with real executive adoption and 3.2x ROI in 6 months. The key piece is controlling what your company is already deploying.

Want a data-driven culture with agents that sustain it? Talk to BAOSS and we’ll design your transformation.

Prefer to talk to us directly? Contact the BAOSS team and tell us about your challenge.