The client’s marketing team had reached a structural inflection point. Marketing data lived in disconnected silos: paid media performance in Meta Ads Manager, lead and pipeline data in HubSpot, web behavior and conversion attribution in Google Analytics, and customer information in CRM. Each platform held a slice of the truth, but no system held the whole picture. Senior marketing leaders could not answer their most important questions, Where should we spend the next ad dollar, which audiences are converting, where is creative fatigue setting in, without first stitching the data together by hand.
The cost of this fragmentation showed up in time and accuracy. A single marketing analyst was spending four to five hours every week pulling Meta lead exports, downloading HubSpot reports, reconciling Google Analytics attribution, and assembling monthly tabulations for the leadership review. The work was not only slow, but it was also error-prone, because manual addition across thousands of rows introduced discrepancies that nobody could fully audit. Worse, the strategic conversations that depended on this data were locked to a monthly cadence. If a senior leader wanted to ask “with $10,000 in additional budget, where should I spend it?” mid-month, the answer was “wait for the next L1 review.”
The risk of inaction was strategic, not just operational. Well-resourced competitors were pumping out campaigns capturing share-of-voice in the dental market. The client’s marketing team could not scale by adding headcount in a lean operating environment. They needed to make a small team faster, sharper, and more strategically effective without growing it.
When Kerr’s marketing leadership came to Zimetrics, the request could easily have been read as “build us a reporting dashboard.” Zimetrics reframed it. A dashboard would have shown the same data faster, but it would not have answered the strategic questions. The real problem was twofold: the data was fragmented across systems that did not talk to each other, and the strategic reasoning that should have followed the data was bottlenecked by the time it took a human to reconcile it. Both problems had to be solved together.
Zimetrics proposed a Campaign Intelligence Assistant, built on AWS; an agentic AI system with two interlocking jobs. First, unify paid-media data from Meta Ads Manager, HubSpot, Google Analytics 4, and CRM into a single queryable foundation. Second, layer an agentic reasoning engine on top of that foundation that could compute performance metrics (CTR, CPC, CPM, ROAS, CVR), benchmark performance quarter-over-quarter, correlate results against audience segments and message themes, link ad spend to conversion outcomes through CRM and GA4 attribution, and produce forward-looking recommendations on budget allocation, geographic targeting, and demographic strategy.
The architectural principle was clear: data unification first, intelligence second. No amount of LLM cleverness can compensate for a data foundation that does not exist. By solving the unification problem at the platform layer, Zimetrics gave the agentic reasoning layer something coherent to reason over.
The Campaign Intelligence Assistant was deliberately positioned for senior marketing leaders making strategic decisions. A junior analyst might still pull a number out of Meta Ads Manager directly. A VP asking, “should I shift Q2 budget toward retargeting in Texas, given how Q1 prospecting performed in California?” needed the agent. This positioning shaped every downstream design decision; from the conversational interface to the structure of the recommendations the agent produced.
The first and most critical track was unifying paid-media data that had never lived in one place. Scheduled ingestion jobs orchestrated through Amazon EventBridge, AWS Step Functions, and AWS Lambda pulled campaign metadata, ad set data, ad-level creative, and performance metrics from the Meta Ads Manager API over secured REST integrations with API key authentication and structured retry logic. Parallel ingestion flows reached into HubSpot for lead and pipeline data, Google Analytics for funnel and attribution, and Kerr’s CRM for conversion and revenue mapping.
Databricks pipelines normalized structured and unstructured marketing data into a unified representation, applying knowledge distillation, schema reconciliation, and cleansing transformations along the way. Raw artifacts landed in Amazon S3, with processed and enriched outputs indexed into Amazon OpenSearch as the queryable foundation. The result was a single, continuously refreshed view of paid-media performance that no human had to assemble by hand.
On top of the unified data layer, Zimetrics built a performance analytics engine that computed the metrics senior marketing leaders actually use to make decisions. The engine aggregated CTR, CPC, CPM, ROAS, and CVR across campaigns, ad sets, and audience groups; benchmarked current-period performance against historical baselines for quarter-over-quarter comparison; correlated results against audience type (retargeting versus prospecting), demographic segments, geographic regions, and creative message themes such as trust, safety, and efficiency; and linked ad spend to downstream conversion outcomes through CRM and Google Analytics 4 attribution.
Each analytical operation was exposed as a structured tool that the agentic reasoning layer could invoke compositionally, giving the agent the ability to answer not just “what happened” questions but “why it happened” questions grounded in real attribution data.
The reasoning core was built on LangGraph, with Python-based agentic workflows orchestrating how the campaign assistant moved from a marketer’s question to a strategic answer. Two parallel environments were established: a design-time playground where prompt chains, tool definitions, and agent behaviors could be developed and iterated, and a runtime layer where validated agents were promoted into production.
The runtime workflow followed a six-step analytical pattern — data retrieval from the unified store, performance aggregation, comparative benchmarking, audience and message correlation, ROI and conversion analysis, and synthesis into a structured Performance Snapshot covering top performers, creative patterns, cost drivers, underperformers, and forward-looking recommendations for the next quarter.
When a senior leader asked the agent “with $10,000 in additional budget, where should I spend it?” or “what creative fatigue am I seeing on the Texas campaigns?”, the LangGraph workflow assembled the answer in seconds by composing the right analytical tools in the right order, returning a recommendation grounded in actual performance data rather than model speculation.
Foundation model access was provided through Amazon Bedrock over AWS PrivateLink, ensuring no LLM traffic ever traversed the public internet, with Amazon SageMaker available for fine-tuning where deeper customization was needed. Retrieval-augmented generation against the unified marketing data store was powered by Amazon OpenSearch as the vector layer, with metadata classification managed through the AWS Glue Data Catalog so that retrieval could combine semantic search across unstructured signals with structured lookups against campaign, audience, and geographic dimensions.
Conversation state, prompt history, retrieved context, and recommendation outputs were persisted in Amazon DocumentDB, giving every marketer session full traceability and giving the agent the ability to reason across multi-turn strategic conversations rather than treating each question in isolation. The conversational interface itself was delivered as an Angular front-end so senior leaders could ask strategic questions in natural language and receive structured analytical responses without ever opening Meta Ads Manager, HubSpot, or Google Analytics directly.
Because the Campaign Bot’s outputs would directly inform strategic ad-spend decisions, validation was treated as a first-class engineering concern rather than a final QA pass. Zimetrics built an evaluation harness that exercised the agent against a curated golden dataset of marketing leadership queries and expected analytical outputs, with A/B comparison between baseline and fine-tuned prompts to measure whether each iteration improved or degraded answer quality.
Automated scoring on precision, recall, and F1 metrics ran on every meaningful change to the agent. Subject-matter experts from the marketing team reviewed sampled responses for analytical correctness, tone, and strategic relevance through a structured human-in-the-loop process. Retrieval quality against the unified data store was independently validated by verifying correct context retrieval across varying query patterns and concurrent sessions.
Once in production, drift monitoring ran continuously through logging of prompt-response pairs, ensuring that response quality remained stable as the underlying data evolved and that any degradation was caught before it reached a senior leader making a real budget decision.
For the first time, our senior leaders can ask a strategic question about ad spend and get an answer in the same conversation, instead of waiting until the next monthly review. Zimetrics unified our marketing data and built an intelligence layer on top of it.