AI agents: Now BI Can Finally Deliver Measurable ROI

AI is undoubtedly causing the biggest platform shift in the era of BI. For the first time, we can move on from building and selling decision interfaces to delivering decision intelligence. This shift will unlock new revenue streams and finally enable the industry to attach a clear ROI tag to BI.

Just imagine how much more convenient it would be to have a well-designed co-pilot embedded in your dashboard, automatically summarizing and distributing tens of thousands of QBR reports to all your customers. Or how much time a production-planning agent could save by deciding on ad-hoc demands within minutes instead of weeks. Or how much money a root-cause analysis autopilot could save by cross-checking millions of data points in the blink of an eye as soon as a disruption in your business operations occurs.

But how do you ensure these innovations don’t end up being shelved as stalled pilots (as 95% of them do), and instead become production-safe and easily commercializable?

Four Essential Pillars of Customer-Facing Agents

To avoid falling into that 95%, there are four essential features your customer-facing agents must have to be successful:

1. Embedded in Workflows

Your agents have to be embedded in your workflows or applications. Agents living in isolated AI apps or chats, waiting for input, have the same chance of succeeding as a colleague working from home while everyone else is buzzing in the office. Ensuring that your agents are well-integrated and able to intercept relevant data as it flows, while having access to the tools they need to perform actions, is a must.

2. Tailored to Unique Business Processes

Agents need to be tailored to your platform and your customers’ unique business processes. Black boxes work well in demos but rarely deliver above-average results. Imagine hiring a new team member for a critical role without giving them proper instructions or any onboarding. When selecting a platform for your agents, always consider the freedom it gives you to swap core building blocks, and beware of vendor lock-ins. Especially with novel and fast-evolving AI applications, where winning practices and modules have not been decided yet, you want to make sure you always have the option to rebuild and recompose.

3. Performant and Scalable

Agents place unprecedented requirements on the performance and scalability of infrastructure and platforms. Most existing systems are designed around large human workloads, but agents introduce a completely new scale. Have you ever seen an analyst performing tens of thousands of analyses in parallel? Make sure your agents have access to tools and services that can operate effectively at high speed and high volumes.

4. Unquestionably Reliable

Agents running in production and working with your customer data must be absolutely reliable. One mistake, one wrong answer, one innocent data leak can shatter trust in your agents forever. Always prioritize agents where security is a core philosophy.

Reliability goes beyond data privacy, though. It’s about ensuring your agents never hallucinate. Instead, they must always ground their answers in data and facts that are traceable and auditable.  New agentic platforms are emerging every day, many with impressive capabilities. Keep in mind that while they may perform magically on synthetic data, safety and trustworthiness in real-life operations are what are truly needed.

Foundations for Building Production-Safe Agents

The question then becomes: How do you go about building and offering agents that adhere to these four essential features?

1. Strong Knowledge Foundations

Ensure agents have access to comprehensive and well-structured knowledge about your business. Knowledge is the foundation that gives agents the context they need to reason and ground their answers.

Introduce ontologies into your business domains. Ontologies are formal specifications of knowledge — covering both structured datasets and unstructured documents.

LLMs play a crucial role in enabling agents to work with this knowledge. By understanding written text, they unlock new ways to extract valuable information from documents, transform it into knowledge graphs, and connect it with your business’s existing semantic model. This allows agents to better understand the business context and find every “needle in the haystack”.

Well-designed agentic platforms should also recognize differences across business domains and help you effectively build industry-specific ontologies that your agents can tap into.

2. Transparent Actions

Make sure agent actions are well defined. Introduce agentic workflows that orchestrate planning, escalation rules, and tool usage. This provides the transparency and control needed to run production-safe agents.

Eventually, think about the guardrails you incorporate into the agent’s digital soul. A robust control tower should clearly define what agents are allowed and forbidden to do, when they should ask a human for input or approval, and how they should cooperate to achieve a given goal in a way that aligns with your business values and intentions.

3. Be Aware of AI Strengths and Weaknesses

Be cognizant of LLMs’ strengths and weaknesses. Decide what tasks to delegate to LLMs — bearing the risks of sharing sensitive data — and what can instead be handled by cheaper and fully deterministic solutions.

LLMs are excellent at identifying intent, creating summarizations, or writing elegant code. But they aren’t appropriate for basic logic or tasks requiring strict determinism.

Build on these foundations and you will be able to offer assistants that help with portfolio rebalancing without exposing customer holdings to LLMs; inventory optimization agents that can distinguish between overstocks and expirations without causing more waste; production planning agents that can navigate complex directives, guidelines, and procedures; or root cause analysis automations that won’t hallucinate about the true causes of your challenges.

GoodData’s Philosophy and Path Forward

We hope you find these essentials and foundations helpful when creating and offering your embedded agents. At GoodData, we have spent countless hours refining our philosophy of production-safe and commercializable agents — agents that are able to work reliably from raw data and documents, through their inner workings and decision-making, all the way to the final interfaces your customers interact with.

After pioneering and dominating the embedded analytics space, we have accumulated enough experience to translate our knowledge from building decision interfaces to delivering decision intelligence.

GoodData’s code-driven approach to building and accelerating AI development, combined with a platform that embodies the essentials of being production-safe, well-integrated, transparent, performant, secure, and knowledgeable, will serve as a strong foundation for your agentic future. To better understand the technical shifts in BI that enable this new era of agents, why not read “Why AI Changed the Way We View BI”?

Let’s turn your data into intelligence together. Schedule a demo and talk with our team about AI opportunities for your business.