How Adservio's PerformanceAI Transforms Marketing Analytics with AI Agents

PerformanceAI: Ask a Business Question, Get a Reliable Answer in Seconds

We all remember the promises made by SaaS vendors: tools that would put critical data within reach. Most of us have been disappointed. These often cumbersome platforms force users down labyrinthine paths, stacked dashboards, filters to configure, exports to rework, to get, at best, partial answers that are already stale by the time they arrive.

Wouldn't it be simpler to just ask a question and get an answer in a flash? That is exactly what PerformanceAI delivers, the multi-agent marketing insights application built by Adservio. Its goal is simple but powerful: any marketer should be able to query performance data in natural language, for example, "show me conversion rates for our financial services campaigns in the first half of the year", and get an instant, sourced, reliable answer.

Technically, the application is built in Python and Flask on the server side, with a React front end. It runs on BigQuery and Vertex AI in Google Cloud, and orchestrates Anthropic's Claude Sonnet 5 and Google's Gemini 3.5 Flash models: the former for precision reasoning and SQL generation, the latter for fast synthesis and high-volume tasks.

In short, the ambition was to let a marketer talk directly to an agent and ask a business question, instead of going to an analyst or manually pulling data from systems like Salesforce.

Why Marketing Teams Needed an Insights Agent

Before PerformanceAI, marketing teams had performance data, but it was hard to exploit. Data remained fragmented across systems, CRM, automation platform, advertising tools, web analytics. Each functional team produced its own reports, resulting in inconsistent metrics and no single source of truth shared between marketing, sales and leadership.

Instead of spotting opportunities early, marketers waited for the numbers to be cleaned, consolidated and presented. PerformanceAI makes this experience more direct and more human: everyone gets straight to the insights and acts faster on the pipeline, spotting an under-engaged buying group at the qualification stage, a weakness in mid-funnel acceleration, or an opportunity to lift conversion during solution development.

The cost of this friction is measured in lost cycles: every question that requires a ticket to the data team adds days of delay, and many questions are simply never asked at all. At the scale of an organization, that means campaigns that keep running without any proof of effectiveness, and entire segments of the pipeline that never get the attention they deserve.

Related readData Mesh: The Architecture Revolutionizing Enterprise Data ManagementData Mesh in 2026: domain ownership, contract-backed data products, self-serve platforms, federated governance and the open lakehouse to scale enterprise data.Read the article

AI-Ready Data: The Genesis of the Project

For enterprise AI to work, it needs the right data foundations. As early as 2023, Adservio overhauled its go-to-market KPI framework and invested in a data architecture on Google Cloud: more than twenty critical sales and marketing data sources were integrated into BigQuery, freeing teams from fragmented SaaS interfaces and aligning data with real workflows. By the time the GenAI wave arrived, the data was already AI-ready.

Structuring Data for Large Language Models

The team then built custom agents on Vertex AI to access this data directly, and restructured the marketing datasets for LLMs. Early experiments delivered a decisive lesson: AI performs better when data is organized in a way that is also intuitive for humans. By relabeling fields and clarifying business semantics, the model started producing accurate, contextually relevant answers.

The first working version reached marketing leaders in mid-2025, just a few months after the initial concept. The scope grew from fifteen basic questions to the full spectrum of growth-related questions: opportunities, bookings, conversion rates, campaign impact. Since then, every iteration has broadened functional coverage while hardening the reliability of the answers.

A Multi-Agent Architecture: Routing, Prompts and Guardrails

Under the hood, PerformanceAI runs on a multi-agent architecture that has become the reference pattern for enterprise AI systems in 2026: rather than one monolithic model that does everything, a team of specialized agents hand off to each other, each with a bounded role, limited permissions and verifiable outputs.

Routing and the Prompt Architect

The user's question first enters a routing agent that identifies the type of question being asked and chooses the dataset and the agent to engage, three specialized agents cover three datasets, each with its own business rules and BigQuery fields. At the heart of the system, an autonomous prompt architect interprets the question, retrieves the relevant BigQuery schema, applies a ten-rule domain framework and builds a perfectly contextualized prompt in real time. This is passed to Claude Sonnet 5 via Vertex AI, which converts natural language into enterprise-grade SQL queries.

Execution Guardrails and Insight Synthesis

Each query then passes through a guardrail that validates and adjusts it before execution: capped costs, authorized data scopes, protection against prompt injection and against heavy or dangerous calls. Once the query runs, an insight synthesizer distills the results, full dataset statistics plus a smart sample for the interface, delivered within seconds. This strict separation between generation, validation and execution is what makes it possible to run agents in production without sacrificing governance.

RAG and Semantic Search over Marketing and Sales Data

In parallel with the SQL pipeline, PerformanceAI employs a RAG search agent, powered by Vertex AI embeddings and Gemini 3.5 Flash, to run semantic searches across the indexed datasets. When a user looks for the context of a specific opportunity or account, the system embeds the question, performs a millisecond-speed semantic search across marketing and sales data, retrieves the most relevant records and generates a contextual answer.

From Vector to Decision

The marketer thus gets a complete view of the interactions: campaigns used, contacts who responded, whether the opportunity was inbound. Combined, the two pipelines, generated SQL for quantitative questions, RAG for contextual questions, make it possible to answer in natural language with real-time precision, surfacing insights such as under-engaged buying groups or the campaigns that most influence mid-funnel progression.

The retrieval techniques matter as much as the model itself: careful indexing, metadata filtering and reranking of the results directly determine the quality of the final answer. It is a continuous improvement effort, fed by the analysis of the questions that did not get a good answer the first time around.

Related readFour Retrieval Techniques to Improve RAG: CRAG, Self-RAG, RAG-Fusion and GraphRAGCRAG, Self-RAG, RAG-fusion, Fast GraphRAG: four advanced retrieval techniques to make your RAG systems more accurate, more robust and far more scalable.Read the article

Results, Adoption and Leadership Lessons

The simplest test for any tool remains: do people use it? On that basis, PerformanceAI hit the ground running. Field marketing leaders use it daily to track new opportunities, measure marketing-influenced deals and identify engagement gaps. The first two regional testers report saving hours every week, and the CMO calls PerformanceAI a "toothbrush" product: she uses it at least twice a day. On the business side, the tool has tightened alignment between marketing and sales around shared metrics tracked live.

Four Leadership Lessons

Executive sponsorship first: involving the CMO and CRO from day one guaranteed adoption. Cross-functional collaboration next: go-to-market operations acted as a bridge between marketing, sales and IT. Iterative delivery too: combining manual and automated reporting built trust early. Finally, place AI engineers inside the business function, owning the intelligence layer, while IT provides the architecture, the guardrails and AIOps support for scaling.

Continuous Evaluation to Preserve Trust

The most structuring lesson remains evaluation: every change to the system, a new model, a new dataset, a new business rule, goes through a battery of automated evaluations that compare generated answers against references validated by analysts. Without that discipline, trust earned over weeks can be lost on a single wrong answer delivered with confidence.

This requirement applies to any AI system in production: evaluations play the role that non-regression tests played for classic software. They make it possible to switch models with confidence when a more capable generation arrives, to measure the effect of a new prompt, and to detect drift before users ever experience it.

Related readHow to Evaluate an LLM SystemEvaluating an LLM system in production: quality metrics, eval datasets, LLM-as-judge, prompt regression testing and continuous drift monitoring.Read the article

Toward Super Agents That Can Act in the CRM

The team is now training the system to cover a broader range of questions and to automate recommendations. The vision: connect PerformanceAI to Adservio's unified intelligence workspace, where the user asks a performance question and the system calls the right "super agent" in the background, one per business function. All voice-enabled, with the ability to trigger actions, update a buying group, flag a dormant opportunity, without going back through the screens of a traditional CRM.

This trajectory is part of a broader movement: with the generalization of interoperability protocols such as MCP, internal super agents will be able to talk to the agents of partners and third-party platforms. Today's conversational analytics is preparing tomorrow's agentic enterprise, where asking a question and triggering an action will be one and the same gesture.

PerformanceAI is no longer an experiment: it is a daily companion for marketers and a tangible example of how Adservio applies its own AI expertise, data architecture, agents in production, governance, to improve performance, productivity and collaboration across teams. It is also part of a deeper shift toward AI-enabled marketing, where data no longer merely informs decisions: it helps make them, and then execute them.