How we built a multi-agent AI system on Databricks that answers any business question — in seconds, for the right person, from any channel.

In the modern enterprise, data is everywhere — yet insights often remain out of reach. Organizations have successfully migrated to the cloud and centralized their data in Lakehouses, but the "last mile" of data consumption is still broken. Getting real-time answers to cross-functional questions remains a manual, siloed process.

Imagine a VP asking: "What is our total revenue this week, and how many of those deliveries are currently delayed?" Today, that simple question triggers a chain of emails, manual SQL queries across departments, and spreadsheet consolidation. By the time the answer arrives, the morning is gone — and the data is already stale.

At CI&T, we believe the future of business intelligence lies in Agentic Orchestration: systems where specialized AI agents collaborate to answer complex, cross-domain questions in real time, from any channel, delivering the right answer to the right person.

That belief led us to build MASBI — Multi-Agent Supervisor Business Intelligence. Built on Databricks, MASBI turns natural language questions into real-time insights across multiple business domains, for multiple users — through a single API call, with proper access governance, and even via WhatsApp.

One Interface, the Right Answer, for the Right Person

One of the most underappreciated challenges in enterprise data access isn't just where data lives — it's who should be allowed to see it, and how to enforce that without forcing every user into a different system.

In large organizations, a regional sales manager, a logistics coordinator, a CFO, and a warehouse operator may all need data-driven answers daily — but from entirely different domains, with entirely different access privileges. Traditionally, this means maintaining multiple BI tools, multiple access control policies, and multiple training efforts across the organization.

MASBI collapses that complexity into a single conversational interface. A user simply asks their question — regardless of domain — and the Supervisor Agent handles everything: routing the request to the right specialist agent, enforcing data access permissions aligned with Unity Catalog's governance layer, and returning only the information that user is authorized to see.
Someone without access to financial data will not see financial data — not because of a locked dashboard, but because the system simply won't surface it.

The result is a unified experience where fewer systems mean less friction, onboarding is simpler, and governance is stronger — all without the user ever needing to know which agent answered, which table was queried, or which access policy was applied.

The Challenge: Breaking the Silos of the Modern Data Stack

Most organizations suffer from "Insight Silos." Sales has its dashboards, Logistics its tracking tools, Finance its proprietary reports. The data may live in a unified Lakehouse, but the people who need answers are scattered across functions — each locked into their own BI tool and their own way of accessing data.

Consider a large retail operation with thousands of stores, dozens of distribution centers, and millions of daily transactions. When a regional manager wants to understand why a promotion is underperforming, they need sales velocity from Commercial, stock levels from Supply Chain, and campaign metrics from Marketing — all correlated. In practice, they get three separate reports, delivered at three different times, in three different formats.

The insight gap isn't technical. It's organizational. And it costs real money — in delayed decisions, missed replenishment windows, and reactive operations instead of proactive ones.

Our goal was to build a system that could:

- Automatically understand the domain of any question
- Route it to a specialized AI agent
- Query live data directly from the Lakehouse
- Scale efficiently while maintaining cost-effectiveness and security
- Work across any channel — API, browser, or WhatsApp

The Solution: One Question, Many Minds

Built on the Databricks Platform, MASBI orchestrates six specialized AI agents and multiple Genie Spaces — natural language-to-SQL engines — spanning seven business domains: Sales, Logistics, Finance, Customers, Inventory, Marketing, and Documentation.

Rather than one "jack-of-all-trades" bot, MASBI uses a Supervisor-specialist pattern. When a question arrives — via REST API, browser, or WhatsApp — an intelligent routing layer determines which domain it belongs to and hands it off to the right specialist. That agent queries live data and returns a natural language answer. The user never needs to know which agent responded. They just get the answer.

For cross-domain questions, the system detects that multiple domains are involved and orchestrates responses from multiple agents simultaneously. The complexity is hidden; the simplicity is what the user experiences.


The Genie Challenge: When Six SQL Engines Need to Work as One

Each domain agent is powered by a Databricks Genie Space — a natural language-to-SQL engine that translates human questions into precise queries against Unity Catalog tables. Genie Spaces are remarkably powerful on their own. But orchestrating six of them behind a unified interface introduces real engineering challenges:

Routing accuracy: How do you determine which Genie Space should handle a given question — especially when it's ambiguous? "What is our margin on delayed orders?" touches Finance, Logistics, and Sales simultaneously.

Conversation continuity: If a user asks about revenue and then says "break that down by store," how does the system maintain context without re-classifying every follow-up?

API throughput at scale: Genie Spaces have rate limits. When multiple users are asking questions simultaneously across different domains, how do you prevent bottlenecks?

Cost awareness: Routing every question through a large language model for classification is expensive. At enterprise volumes, those token costs compound quickly.

Solving these challenges is what transforms MASBI from a demo into an engineering blueprint for production-grade, multi-domain AI systems.

Smart Routing: Efficiency at Scale

The most critical component of MASBI is something the user never sees: a multi-layered semantic routing pipeline that classifies every question and directs it to the right agent with minimal latency and minimal cost.

Instead of sending every request to a Large Language Model for classification — which is slow and expensive — MASBI uses a cascading approach. Permission checks and session context resolve a significant portion of requests with zero token cost. Semantic embeddings handle the majority of remaining questions in milliseconds. Only genuinely ambiguous queries reach the LLM fallback layer — and even then, a cost-efficient open model is used rather than a premium API.

The result: the vast majority of requests are classified without consuming a single LLM token. At enterprise scale, this is the difference between a system that is economically viable and one that is not.

Lakebase: The Quiet Engine That Makes It All Work

If the routing pipeline is the brain of MASBI, Lakebase is its memory — and one of the most consequential architectural decisions in the project.

Lakebase, Databricks' managed PostgreSQL service, solves three critical problems with a single service:

Conversation state. Every user session is persisted with a configurable time-to-live. When a user asks "and by region?" after a revenue question, the system retrieves context and routes the follow-up to the same agent — with zero classification cost. This alone eliminates an estimated 30% of redundant routing calls.

Intelligent caching. Every Genie answer is cached using a normalized hash of the question. When the same question is asked again — by any user, in any channel — the response is served from cache in roughly 16 milliseconds, compared to 25+ seconds for a live Genie call. For organizations where dozens of managers ask the same morning briefing questions, the impact is immediate: instant answers at zero API cost.

Graceful queuing. When Genie API rate limits are reached, requests are queued with automatic retry and exponential backoff. The system absorbs traffic spikes silently — no user-visible errors, no lost queries.

The economics scale naturally as well. Lakebase Autoscaling supports scale-to-zero: when no one is using the system — evenings, weekends, holidays — compute cost drops to zero. Session data, cache, and queue all persist on storage, ready to serve instantly when demand returns. For a proof of concept, you're not burning budget overnight. For production, cost tracks actual usage.

Beyond the Desktop: The Power of Multi-Channel Delivery

For AI to be truly transformative, it must meet users where they are. A field manager checking stock levels from a warehouse floor isn't going to open a BI dashboard — they're going to reach for their phone.

By integrating with Twilio, MASBI delivers the same level of intelligence via WhatsApp — the most widely used messaging platform in Latin America. A store manager can text "how many deliveries are delayed to my region?" and receive a precise, data-backed answer within seconds. The same routing, the same agents, the same live data. The channel changes; the intelligence does not.
Whether a field manager is checking inventory on their phone or a director is querying sales figures via a REST API, the intelligence layer remains consistent, secure, and fully governed.

Full Observability: Tracing Every Decision in the Chain

In a multi-agent system, the hardest question isn't "what went wrong?" — it's "where in the chain did it go wrong?" When a user receives an unexpected answer, was it a routing misclassification? A poorly generated SQL query? A stale cache hit? Without end-to-end observability, debugging becomes guesswork.

MASBI integrates MLflow 3 Tracing across the entire request lifecycle. Every interaction produces a complete trace that records:

Routing decisions — which classification layer resolved the request (permission check, semantic embedding, or LLM fallback), the confidence score, and the domain selected.

Agent activations — which Genie Space was invoked, the natural language-to-SQL translation generated, and the execution time of the underlying query.

Infrastructure interactions — cache hits or misses in Lakebase, queue events during rate-limit scenarios, and session context retrievals.
Token consumption — exactly how many LLM tokens were spent per request, enabling precise cost attribution by domain, user, or channel.

This isn't just a debugging tool — it's a continuous improvement engine. By analyzing trace patterns at scale, teams can identify which domains have the lowest routing accuracy, which questions trigger expensive LLM fallbacks most often, and where caching strategies can be optimized. In production, observability is the difference between a system you operate and a system that operates you.

Why This Matters for the Enterprise

The challenges MASBI addresses are universal to any data-intensive organization where decisions move faster than traditional BI can respond.

Retail and consumer goods operate on razor-thin margins and split-second decisions. A delayed replenishment order means empty shelves. A misread demand signal means excess inventory. A late response to a logistics bottleneck means missed SLAs and unhappy customers.

Financial services can't afford to wait for overnight batch reports to understand portfolio exposure. When market conditions shift intraday, asking "What is our current credit risk concentration in Southeast Asia?" and getting a live, cross-domain answer — pulling from trading, compliance, and customer data simultaneously — can be the difference between a managed risk and a regulatory incident.

Healthcare and life sciences face an even higher stakes version of this problem. A hospital operations manager asking "Which departments are currently over capacity, and how does that correlate with our staffing levels and pending discharge approvals?" isn't asking a reporting question — it's a patient safety question. Speed and accuracy directly impact outcomes.

Manufacturing and industrial operations measure downtime in revenue lost by the minute. A plant director asking "Which production lines are underperforming this shift, and are there pending maintenance tickets linked to those assets?" needs an answer before the next shift begins — not tomorrow morning in a PowerPoint.

Telecommunications network operations teams managing millions of endpoints need instant correlation between customer complaint volumes, network fault logs, and field technician availability. Siloed data in that context doesn't just slow decisions — it degrades service quality at scale.

Across all of these industries, the pattern is the same: the data exists, the platform is capable, but the last mile of human access remains broken. By leveraging a unified platform like Databricks, organizations across any sector can:

Reduce operational friction — move from "requesting a report" to "asking a question"
Maintain governance — keep all data and AI interactions under Unity Catalog
Scale responsibly — use semantic routing, intelligent caching, and scale-to-zero infrastructure to manage costs
Democratize data access — give any user, from the C-suite to the warehouse floor, conversational access to business data in any channel

Looking Ahead

This is more than a technical demonstration — it is a blueprint for how CI&T helps clients navigate the complexities of GenAI. Our roadmap includes implementing advanced evaluation frameworks to ensure every answer is as accurate as it is fast, and exploring hybrid architectures that combine custom orchestration with Databricks Agent Bricks for even faster time-to-value.
The question is no longer "How do we build a better dashboard?"
It is: "How do we make the data answer the question itself?"

Want to see MASBI in action? Reach out to CI&T to schedule a live demo and discover how agentic orchestration can transform your organization's relationship with data.

Thomaz Antonio Rossito Neto Specialist Data Engineer & AI Solutions Architect at CI&T Databricks Certified Professional

Thomaz Antonio Rossito Neto

Thomaz Antonio Rossito Neto