From AI Hype to Enterprise Execution: Key Insights from BIAAS 2026 in Brussels

On 24 and 25 June, the Business Intelligence and Analytics Summit 2026 brought together data leaders, enterprise architects, and analytics executives in Brussels. This year's gathering marked a clear turning point for the industry. The conversation moved decisively away from speculative AI potential and focused squarely on the operational, financial, and architectural realities of scaling artificial intelligence within modern organisations.

Across two days of keynotes, case studies, and panel discussions, a unified message emerged: successfully deploying enterprise AI requires robust structural governance, rigorous cloud unit economics, a clean analytics foundation, and a complete redesign of legacy business workflows.

Here is an overarching report synthesising the core themes, technical strategies, and strategic insights shared at BIAAS 2026.

1. Cracking the Commercial Intent Paradox

A primary challenge facing modern enterprises is translating AI experimentation into measurable bottom-line value. Zahid Hossain (Partner, Tiger Analytics) highlighted a striking contrast in current enterprise adoption: while 80% of executives list AI as a top strategic priority, fewer than 20% achieve scaled commercial value. Hossain emphasised that this value leakage rarely stems from model failure. Instead, organisations ask the wrong foundational question, starting with "what can AI do?" rather than "what specific business problem must we solve?" To bridge this gap, leaders must shift from technical capability to commercial intent, framing every development project around concrete operational metrics from day one.

This theme carried into a fireside discussion between Anna Bisch (Advisory Board Chair, Nexus Europe) and Andrey Malyy (VP IT & Data, Transformation & Governance). Examining why integrated digital initiatives stall, they identified low innovation maturity and the "proof of concept trap" as major culprits. Teams frequently launch isolated pilots without defining a clear framework for global scaling. To overcome this, organisations must enforce a strict progression model moving from initial Idea to prototype, Minimum Viable Product (MVP), and finally Global Scale, with every MVP bound to strict success criteria and a clear business case.

Bisch and Malyy also underscored the necessity of localised change management when expanding solutions internationally. Deploying digital tools globally without adapting to regional operational realities consistently leads to zero adoption. Furthermore, organisations must guard against vendor lock-in by securing firm contractual terms. Aligning with European standards such as GAIA-X ensures that development partners must restitute all proprietary code, business logic, and data to the enterprise upon contract completion.

2. Rebuilding Foundations: Modern BI, Semantic Layers, and Governance

Scaling reliable AI requires fixing underlying data and analytics environments. Gennadii Armashula (CEO, Accton) cautioned against rushing into generative AI deployments on top of chaotic data foundations. Presenting audit data from a recent enterprise client, Armashula revealed an environment in which over 500 reports were scattered across 300 unmanaged workspaces, with 60% unused and overall adoption at just 18%. To restore trust, leaders must apply the Pareto principle to identify the core 20% of analytics assets driving 80%of business decisions, archiving the rest. Organisations must then establish a unified semantic layer defining key metrics once across all tools, ensuring AI models and reporting dashboards pull from verified context.

Modernising business intelligence also means rethinking traditional software stacks. Tomas Trnka (CDO & AI Lead, EAG) detailed his team's transition away from legacy BI tools like Tableau, driven by restrictive per-seat licensing and manual drag-and-drop interfaces that prevent automated development. EAG rebuilt its reporting stack using a code-first architecture featuring Snowflake for data warehousing, Rill for volume-priced BI, and Claude for conversational queries. By storing all dashboards, metrics, and data connectors as code in Git, the environment became fully AI-accessible. An agentic AI system successfully migrated essential dashboards in days. To prevent conversational hallucinations, Trnka's team implemented Snowflake Semantic Views, forcing the AI assistant to query only pre-certified database metrics when answering user questions.

At the ecosystem level, Alain Mugwaneza (AI & Data Governance Lead - EU AI Act / Responsible & Trustworthy AI, Syntone Media) outlined the shift toward a Unified Governance Framework integrating quality, privacy, security, and AI oversight. Enterprise data catalogues serve as the operational hub for data stewards, but Mugwaneza stressed that driving adoption requires masking backend technical complexity behind a retail-style data marketplace. Addressing regulatory frameworks like the EU AI Act, he noted that compliance cannot be handed off solely to legal departments; it requires a multidisciplinary approach bridging technology, policy, and business workflows. Furthermore, building knowledge graphs and ontologies is essential to provide generative copilots with accurate business context.

Streamlining information flow across these architectures was a focus for Marek Zajac (Data Analytics & Governance Lead). Defining efficiency as maximising useful decision outputs while minimising human and computational effort, Zajac illustrated how subject matter experts waste time answering repetitive queries. By encoding expert knowledge into central glossaries and lineage maps, organisations create a self-service context layer. Zajac also advocated for establishing horizontal data communities across business silos. These communities build trust, eliminate duplicate work, and provide a direct feedback channel to central IT teams.

3. Financial Operations and Engineering at Scale

As autonomous AI agents enter production, managing non-deterministic compute costs has become an urgent priority. Steve Loh (Data Analytics Practice Lead, Google) warned that agentic workloads do not scale predictably like traditional cloud infrastructure. Multi-step AI agents can scale processing loads 100x instantly, threatening to exhaust annual cloud budgets within weeks. A simple prompt can trigger complex reasoning chains or multi-step retrievals that consume hundreds of thousands of tokens. Loh advocated for a shift-left FinOps approach, embedding financial controls into the initial software design phase. Key technical optimisation strategies include:

• Context Caching: Reusing recurring input prompts and contextual data to reduce token processing fees by up to 90%.
  

• Semantic Routing: Analysing prompts at runtime to direct simple tasks to lightweight models, reserving expensive reasoning models strictly for complex queries.
  

• Asynchronous Batch Processing: Utilising batch execution for non-urgent tasks to achieve significant cost savings over real-time API calls.
  

Loh also introduced Agentic FinOps, where specialised squads of AI agents continuously monitor cloud infrastructure, detect spending anomalies, and automatically submit code optimisations such as adding partition filters to unindexed database queries.

In a deep-dive technical case study, Steve Verwerft (Senior Head of AI, Data & Application Engineering, Airties) shared how his team engineered a predictive machine learning system for 40 million home Wi-Fi gateways, ingesting over 3 petabytes of telemetry monthly. The system predicts real-time application usage to prioritise bandwidth for latency-sensitive applications like video calls. Verwerft highlighted critical lessons learned while scaling:

• Training Data Scope: An initial model trained on 7 days of data yielded under 50% accuracy. Expanding the window to 50 consecutive days and adopting an XGBoost model increased prediction accuracy to 75%.
  

• System Caching: Repeatedly calculating schedules for identical devices created massive server overhead. Implementing cache layers for calculated schedules reduced runtime from 2 minutes to 45 seconds, cutting compute sizing by 75%.
  

• Storage Formatting: Storing records as raw JSON created bloated 200 KB files per device. Switching to MessagePack compression reduced record sizes by 50x to under 5 KB, turning a cost-prohibitive initiative into an efficient production feature processing 4 million daily inferences.

4. Architectural Safeguards and Enterprise AI Fluency

Operating non-deterministic AI systems safely requires moving away from traditional gated code reviews and soft prompt instructions. Andrey Malyy (VP IT & Data, Danone) demonstrated the dangers of weak guardrails, citing an incident where an agent operating in a staging environment accessed exposed volume credentials and permanently deleted a production database in nine seconds. Text-based prompt instructions inside an LLM context window are insufficient because agents can find execution workarounds.

Malyy argued that hard guardrails must be enforced at the API gateway level, sitting entirely outside the agent's context window. He outlined four core operational pillars:

1. Risk Sorting: Applying strict gateway policies where high-risk operations occur while maintaining flexibility for low-risk tasks.
   

2. Identity and Accountability: Assigning every agent a unique system identity with permissions strictly capped to match those of its accountable human owner.
   

3. Guardrails as Code: Integrating automated security and policy checks directly into CI/CD pipelines.
   

4. Automated Traceability: Logging all execution steps for real-time debugging and auditability.
   

To manage these controls, enterprises must establish a dedicated Platform Team sitting above individual product squads to build, monitor, and update shared architectural guardrails.

Jacopo De Stefani (Expert Data Scientist, Test Achats) reinforced these security imperatives, noting that AI pipelines carry dynamic external dependencies. Before writing code, teams must map existing operations on paper to verify whether AI is genuinely the most cost-effective solution compared to standard automation. To protect pipelines from breaking, organisations must enact formal data contracts and service level agreements specifying schema standards and ownership. De Stefani detailed three essential production safeguards: input and output sanitation to block prompt injections, least privilege access to restrict agents to read-only data, and automated circuit breakers to halt execution loops. He added that human oversight must scale dynamically based on risk, requiring mandatory human confirmation for high-risk operations while allowing routine tasks higher autonomy.

Finally, Tina Milanovic (AI Adoption & Literacy Consultant) addressed the human adoption barrier behind the failure of up to 95%of enterprise AI pilots. Milanovic drew a historical parallel to the adoption of electricity in 1881, noting that factory productivity remained flat for thirty years because owners simply swapped central steam engines for electric dynamos while keeping legacy belts, shafts, and factory layouts. Productivity exploded only when operations were completely redesigned around unit-drive assembly lines.

Similarly, bolting AI chat windows onto legacy processes while expecting employees to perform tasks in the exact same way yields minimal return on investment. Organisations must deconstruct workflows, eliminate steps originally created due to human processing limits, and rebuild operations around AI capabilities. Milanovic advised leaders to actively involve sceptical senior staff who hold deep operational process knowledge, and to replace generic e-learning with role-specific, cohort-based training focused on actual daily tools and tasks.

Strategic Blueprint for 2026 and Beyond

The insights shared across BIAAS 2026 form a clear operational roadmap for organisations seeking to scale artificial intelligence successfully:

1. Focus on Commercial Intent: Avoid capability-first experimentation. Define explicit business problems, enforce strict progression stages from pilot to scale, and tie every project to a clear business case.
   

2. Prioritise the Foundation: Audit legacy BI assets, establish a unified semantic layer, and implement code-first data architectures to provide AI models with clean, reliable context.
   

3. Engineer for FinOps: Build cost controls into the initial software design using context caching, semantic routing, and automated FinOps agents to manage non-deterministic token spend.
   

4. Enforce Hard Architectural Guardrails: Implement gateway-level controls, least privilege permissions, and automated circuit breakers outside the model's context window, managed by a dedicated Platform Team.
   

5. Redesign Workflows Around Fluency: Move beyond inserting AI into legacy tasks. Rebuild processes from the ground up and invest in role-specific, cohort-based training to build enterprise-wide AI fluency.
   

Disclaimer: The views and opinions expressed by the speakers are their personal views and their own and may not represent their respective organisations.

Special Thanks: We extend our sincere gratitude to our sponsors, Tiger Analytics and Accton, as well as all our media partners for their invaluable support in making BIAAS 2026 a success.

Prepare for the next evolution in enterprise data and AI strategy. Visit www.confx-analytics.com to learn more about the upcoming BIAAS 2027 in Amsterdam.