Frequently Asked Questions

What’s stopping most organizations from moving AI from pilots into production?

I see three blockers come up again and again: (1) data access and readiness—teams rush into AI tools, then realize their legacy systems, systems of record, and diverse data types aren’t easy to reach or use safely; (2) use-case redesign—many teams try to bolt AI onto old workflows instead of reimagining processes for an agentic world; and (3) adoption—AI only drives productivity when people trust it, use it, and know how it fits into their day-to-day decisions.

Why do AI pilots often look successful, but production rollouts struggle?

In pilots, people watch everything closely, fix issues manually, and fill gaps with human judgment. In production, you hit “unknown unknowns,” like broken pipelines, stale data, or low-confidence outputs. To succeed, I design for failure conditions up front—clear guardrails, monitoring, auditability, and defined escalation paths to a human review or another agent when confidence drops.

What does it mean to be “embracing AI” but not yet “AI-driven” as a business?

It means AI is present in projects and conversations, but it isn’t reliably embedded into core business processes with the right data access, governance, and adoption. I see many enterprises exploring AI (often cited in the 65–80% range), yet they haven’t built the foundation that lets AI consistently drive decisions, workflows, and measurable outcomes.

Why is data access still the first major hurdle for production AI?

Because enterprise data is spread across older warehouses, catalogs, and multiple engines, often built 5–10+ years ago. Even if it sounds simple to “point an LLM at the data,” I still need secure, governed, and reliable access across structured and unstructured sources. Without that, AI systems either can’t reach the right data or they reach it in ways that aren’t safe or scalable.

What capabilities should a production-ready data foundation provide for AI agents?

I look for four essentials: (1) strong security and access control (including role-level and row-level controls); (2) a semantic layer that adds business context so models can reason accurately; (3) a catalog that’s accessible to agents, not just humans; and (4) interoperability via open standards so data can move across engines and tools without getting locked in.

Why does a semantic layer matter so much for AI accuracy and efficiency?

Models perform best when they get data plus business meaning. I use a semantic layer to translate raw fields into business concepts, definitions, and relationships. That reduces ambiguity, improves answer quality, and avoids wasting tokens on irrelevant context.

How do data products help make AI more reliable and cost-effective?

I use data products to serve “the exact right data” for a specific purpose, with governance and context built in. Instead of dumping huge volumes into a model context window, I expose curated, relevant slices that are easier to secure, easier to audit, and cheaper to use. Action item: start by defining 1–2 high-value data products (for example, revenue and pipeline), then standardize ownership, quality checks, and access policies.

Why are open standards becoming central to AI and analytics strategies?

Open standards help data “escape” proprietary formats and reduce lock-in to a single engine. I get more portability across structured and unstructured data, better separation of compute and storage, and easier interoperability across tools. That flexibility becomes a practical requirement when multiple AI agents and analytics experiences need governed access to the same data.

Are organizations still aiming for a single, monolithic data catalog approach?

Less and less. I’m seeing a shift from the older, monolithic catalog model to a more distributed approach that’s richer and more interoperable. The goal is to make data discoverable and usable across engines and platforms, while keeping governance consistent.

How does a distributed approach reduce data replication and freshness issues?

When data formats and access patterns are interoperable, I don’t need to create massive copies just to serve different tools. That reduces duplication, lowers cost, and helps keep answers consistent because agents and analytics can query governed data closer to where it lives.

How can improving the data layer reduce AI costs and latency?

When I serve agents curated, relevant data (with context), I reduce unnecessary token usage and repeated calls to the model. In the webinar, Qlik shared an internal example where connecting the right data layer led to a 92% reduction in token consumption, about half as many LLM calls, lower latency, and higher accuracy. The takeaway is simple: better data packaging drives step-function improvements, not just incremental gains.

How should teams think about choosing models for different AI tasks?

I match the model to the job and the data. With a strong data foundation, I can often use lighter-weight, faster, less expensive models for many agent workflows instead of defaulting to the largest frontier model. Action item: define a model selection policy that considers cost, latency, and accuracy, then test it against your highest-volume agent tasks.

What trade-offs do teams face between retrieval-augmented generation (RAG) and fine-tuning?

I see teams balancing maintainability and performance. RAG can be powerful for search and grounding, but indexing and embeddings can be harder to maintain over time. Fine-tuning can improve inherent understanding, but it depends on having clean, well-governed data and clear objectives. The key point from the discussion is that you can’t make good trade-offs without a solid data foundation.

Why is streaming data becoming more relevant for AI agents now?

Earlier AI patterns didn’t always mesh well with streaming modalities, but models have become more capable. Now I can combine streaming and non-streaming data to give agents a more up-to-date view of the business—useful in areas like ad tech, manufacturing, and IoT. The practical goal is to plug agents into the pulse of the business, not yesterday’s snapshot.

What’s the recommended approach to ingestion and storage for AI-ready data?

The webinar framed it as a “four-layer cake”: ingestion (streaming and non-streaming), storage (in open, interoperable formats), a semantic layer (business meaning), and a governance layer (controls and auditability). Action item: map your current architecture to these four layers, then identify the weakest layer that’s blocking production AI.

Why is governance an enabler for AI autonomy, not a constraint?

Because no enterprise will let agents operate at scale without confidence that access is controlled, actions are auditable, and sensitive data is protected. I treat governance as the mechanism that makes autonomy possible—policies that move with the data, lineage, quality scoring, audit trails, and sandboxing where needed.

What governance capabilities matter most when agents access enterprise data?

I focus on identity and access management, auditing and logs, data quality signals, anonymization, and data sovereignty. I also want policies that aren’t trapped inside one tool, so access rules and controls remain consistent across dashboards, users, and agents.

How do you handle data quality when humans aren’t always in the loop?

I build explicit failure handling into the agent workflow. If data is stale, broken, or low-confidence, the agent should trigger a human review or route to another agent designed to remediate. I also surface data quality scores as context so users and systems can judge trustworthiness. Action item: define thresholds (for example, “below X quality score requires review”) and bake them into agent decision logic.

How do you prevent confusion when agents return different answers at different times?

I make time and freshness explicit. Two answers can both be correct if the underlying data changed between runs. In the past, a human would explain that a batch ran at 09:00. In production AI, I want agents to understand and communicate that context, and I want deterministic checks where possible to validate outputs.

What kind of team structure works best for building production AI agents?

I don’t see this succeeding in silos anymore. The effective pattern is a multidisciplinary team: AI practitioners (including prompt and context engineering), data and analytics experts, security and governance experts, and business domain experts. That mix helps you build agents that are useful, safe, and aligned to real business workflows.

How do organizations create a feedback loop to improve agents and data products over time?

I treat production as a learning system. I log and audit agent behavior, capture failure states, and feed those insights back into both the agent development lifecycle and the data product roadmap. Over time, that creates a virtuous cycle where governance, quality, and reliability compound.

How can AI help remediate data issues instead of just reporting them?

I can use agents to diagnose issues by reading logs, tracing changes, and explaining what broke and why. The webinar discussed a path toward auto-remediation where an agent can propose a fix, or even create a pull request for a data engineer to review. Action item: start with “assistive remediation” (diagnose and propose), then graduate to controlled automation once governance and testing are mature.