Decision-Making: Intuition or Data?

April 11, 2025 · Baoss

Decision-making 2025: Intuition vs AI with real data

In 2016, when we published the original version of this article, the debate was “intuition or data”. In 2025, however, the question has changed radically. It is now: is your organisation making decisions with AI-curated data, or still leaving millions on the table for fear of letting go of control? The gap is no longer technological: it is cultural and architectural. At BAOSS IT consulting, in fact, we see it every week with our clients.

BAOSS IT consulting: from intuition to AI-driven, data-based decision-making

The real 2025-2026 problem: There is data, but not decisions

According to the McKinsey Global Survey 2024 (published Q1 2025), 72% of companies have already adopted generative AI in at least one business function. Even so, only 23% report a measurable EBIT impact. In Spain, the AMETIC 2025 Report puts real adoption of data-driven decision making at 31% of mid-sized companies (50-500 employees), versus 68% for large corporations. The gap is notable. To choose well who to take that step with, read how to choose an AI technology partner in 2026. Then review whether your company is exposed to the disruption shock.

The bottleneck? These are the three factors we see every week at BAOSS:

  • Fragmented architecture: for example, data in silos (ERP, CRM, Data Lake, SaaS) with no unified semantic layer. LLMs hallucinate because context grounding is non-existent.
  • A “last report” culture: moreover, executive committees keep requesting static PDFs instead of conversational dashboards with natural-language drill-down.
  • Lack of evaluation loops: they deploy AI agents without quality metrics (faithfulness, answer relevance, hallucination rate). They also fail to add human-in-the-loop for critical decisions.

What the data says in 2025 (not 2016)

First of all, let us update the figures that matter:

  • First, average ROI in well-executed GenAI projects: 3.2x at 6 months (Deloitte State of GenAI Q4 2024, n=2,770).
  • Also, reduction in time-to-insight: 68% using RAG + autonomous agents vs traditional BI (Gartner 2025).
  • For example, the Asia-Pacific/Europe gap has reversed: EMEA grows 23% YoY in GenAI spend vs 18% APAC (IDC Worldwide AI Spending Guide 2025).
  • Likewise, trust in internal data: only 41% of CDOs rate their data quality as “high” (MIT CDOIQ 2024).
  • Finally, hybrid talent shortage: 4.2M global vacancies in AI engineering + domain expertise (World Economic Forum 2025).

Ultimately, the conclusion is stark: technology is no longer the limiting factor. What is, is the ability to integrate it into real decision-making processes, with governance and metrics.

How we solve it at BAOSS IT consulting: three anonymised cases (real metrics)

Case 1: Industrial B2B — From “gut feeling” to dynamic pricing with RAG + LangGraph

Context: An automotive components manufacturer (€180M revenue, 4 plants). The pricing committee decided using historical data + commercial intuition. As a result, margin eroded by 3.2 pp over 24 months.

BAOSS solution (14 weeks):

  • First, a hybrid RAG architecture: Vector DB (Qdrant) + Knowledge Graph (Neo4j) for customer-product-plant-regulation relationships.
  • Then, a multi-step LangGraph agent: retrieve → reason → simulate → recommend with human-in-the-loop at thresholds above 2% margin.
  • Also, MCP (Model Context Protocol) to connect the ERP (SAP S/4HANA), CRM (Salesforce) and market data (ICIS, SteelBenchmarker) without data duplication.
  • Finally, continuous evaluation: faithfulness >0.92, answer_relevance >0.89 (RAGAS + a curated dataset of 2,400 historical queries).

Results at 9 months:

  • +4.7 pp EBITDA margin (€8.4M per year).
  • 73% reduction in pricing decision time (48h → 3.2h on average).
  • Also, executive committee adoption: 94% (after 3 prompt engineering for executives workshops).
  • Finally, ROI 4.1x (investment of €320K incl. licences, infrastructure, change management).

Case 2: Omnichannel retail — CrewAI autonomous agents for predictive replenishment

Context: 140 stores + e-commerce (€95M). Stockouts reached 12.3% and overstock 18% on slow-moving SKUs. Moreover, the planning team was saturated (Excel + legacy SAP APO).

BAOSS solution (10 weeks):

  • First, CrewAI orchestration: 4 specialised agents (DemandForecaster, InventoryOptimizer, SupplierNegotiator, ExceptionHandler) with shared memory (ChromaDB).
  • Then, vLLM + Ollama models on-prem (4x H100 GPU) for <200ms latency and data sovereignty.
  • Also, Llama-3.1-70B-Instruct fine-tuning with 3 years of transactional history + promotions + weather + local events.
  • Finally, a conversational dashboard (Streamlit + LangChain) for store managers: “Why am I receiving 40 units of SKU-4422?”. They thus get a natural-language answer with causal drill-down.

    • Results at 6 months:

    • First, stockouts: 12.3% → 3.1% (target <4%).
    • Also, overstock: 18% → 6.8% (€2.1M of working capital released).
    • Also, weekly planning time: 32h → 4.5h (team reassigned to exception analysis).
    • Finally, ROI 3.8x (investment of €185K).

      • Case 3: Financial services — AutoGen for automated compliance + explainability

      Context: A credit institution (€420M in assets). Manual review covered 2,400 files/month for BdE + MiFID II regulation. The cost was €180K/month, with a 15% false-negative rate in external audits.

      BAOSS solution (12 weeks):

      • First, a multi-agent AutoGen system: DocumentParser → RiskClassifier → RegulationMapper → ReportGenerator → HumanReviewer.
      • Claude 4 Opus (via AWS Bedrock) for complex regulatory reasoning + GPT-4o for structured extraction (cost/latency optimised).
      • Also, full traceability: every decision linked to the regulatory article, source paragraph and confidence score. Immutable audit trail (AWS QLDB).
      • Finally, few-shot training with 1,200 files reviewed by senior compliance staff (golden set).

      Results at 4 months:

      • First, false negatives: 15% → 1.2% (below the 2% regulatory threshold).
      • Also, review cost: €180K → €34K/month (81% reduction).
      • Also, average time per file: 42 min → 3.8 min (automated) + 8 min (exception human review).
      • Finally, ROI 5.2x (investment of €290K). External audit 2025: “Best-in-class AI governance”.

      The tech stack that actually works in 2025 (and why yours fails)

      After 47 GenAI projects in production (2023-2025), the winning pattern at BAOSS is not “one tool”. It is, rather, an architecture with evaluation built in:

      Layer BAOSS 2025 choice Why
      Orchestration LangGraph (stateful) / CrewAI (role-based) / AutoGen (conversational) Depending on complexity: deterministic, collaborative or conversational flows. Not “one framework for everything”.
      Models Hybrid: Claude 4 Opus (reasoning) + GPT-4o (extraction) + Llama-3.1-70B/8B (on-prem via vLLM/Ollama) Cost/latency/sovereignty/quality. Automatic model routing by task type.
      Context / RAG Vector (Qdrant/Pinecone) + Graph (Neo4j/FalkorDB) + SQL (MotherDuck/DuckDB) Hybrid beats pure vector: relationships + semantics + analytics. Hybrid search is mandatory.
      Connectivity MCP (Model Context Protocol) + Airbyte/Fivetran for ELT The emerging 2025 standard. Avoids vendor lock-in and connects 300+ sources natively.
      Evaluation RAGAS + DeepEval + curated dataset + weekly human eval Without metrics there is no production. CI/CD for prompts (PromptLayer/Langfuse).
      Observability Langfuse + OpenTelemetry + Grafana Full traces: latency, tokens, cost, quality, human decisions.
      Governance Unity Catalog / OpenLineage + OPA policies Data → model → decision lineage. Complies with the EU AI Act (in force Aug 2026).

      Mistakes we see repeated: using only one vector DB (it loses relationships). Also, using a single model for everything (cost/quality) and zero automatic evaluation (flying blind). Finally, ignoring MCP (you re-invent connectors) and skipping human-in-the-loop in risky decisions (illegal under the high-risk EU AI Act).

      The human factor: From “data analyst” to “decision architect”

      The scarce profile is not the pure data scientist. It is the Decision Engineer: someone who understands the business process and designs the evaluation loop. They also curate the golden dataset, negotiate thresholds with the committee and version prompts as code. At BAOSS we train them internally (a 6-month programme) and certify them. Moreover, the market pays €65-95K gross/year in Spain.

      Ultimately, the conclusion is clear: competitive advantage no longer lies in accumulating data, but in designing decisions. The pillars are solid architecture, automatic evaluation, human-in-the-loop and talent with a Decision Engineer profile. All of them turn AI into sustainable business results.

      Ready to move from analysing data to designing decisions? At BAOSS we support you from architecture to training your team. Contact us.

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

Related articles: Decision-Making: Intuition or Data?.