Creative analysis and conversational data access are two different uses of AI. Both depend on a useful data system underneath.
In a marketing analytics platform, my work has included advertising API ingestion, BigQuery models and serving tables, analytical rules, and applied AI workflows. I contributed within a wider product team.
Start with complete, interpretable inputs
Advertising APIs return entity metadata, performance metrics, creative assets, and attribution-dependent results. Those inputs need consistent identifiers, dates, and definitions before they become reporting data.
The useful pattern is:
Source ingestion → raw data → staging/modeling → reporting outputs
Source recovery and transformation are separate concerns. A missing raw date needs collection or recovery. A historical attribution change may require rebuilding downstream outputs even when the raw data exists.
The platform case explains this engineering work and the rollout boundaries of its transformation architecture.
Applied AI adds another controlled workflow
Media classification is a job over selected assets. A practical workflow identifies eligible media, prepares useful frames or previews, sends the relevant context to a multimodal model, validates the returned structure, and stores the result for analysis.
Tags can describe people, objects, visual style, production characteristics, or messaging. They become useful when connected to the modeled creative and performance data.
A successful parse does not establish that a tag is correct. Useful review includes representative assets, unexpected outputs, and stability across repeated runs. Repeatability and accuracy are separate questions.
Creative similarity or grouping is also an analytical aid. Related characteristics can help compare creative approaches, but a grouping alone does not explain why performance differs.
Analytical rules should carry evidence
Campaign-health logic can compare CPC, CTR, CPA, ROAS, funnel stages, or recent performance against a historical baseline.
Qualification matters: spend thresholds, date windows, and available history affect whether a comparison is meaningful.
Return that evidence with the classification before generating an explanation. An AI-written recommendation should remain tied to the observations and their limits. See why a clearer data model can matter more than prompt changes.
Data access through MCP is a separate interface
A conversational client requests data when a person asks a question. It needs stable references, current entity state, defined outputs, and access control.
For this platform, I designed the BigQuery serving layer behind MCP. A colleague implemented the MCP transport, authentication, and server scaffolding.
That distinction matters: preparing data for an AI interface is substantive engineering work, while the wider integration also depends on other team contributions.
The reporting layer can serve both a product interface and conversational exploration. The BigQuery-to-Claude worked example shows how metric definitions and freshness can travel with a result.
Keep failures and lineage visible
For an AI-assisted workflow, retain enough context to investigate:
- The asset or entity and the source data used.
- The model/prompt version and relevant input context.
- Parsing and validation outcomes.
- Whether the result reached its intended table or application output.
- Refresh state and known data gaps.
These checks make problems diagnosable. They do not eliminate model errors or guarantee that a recommendation is correct.
Build the useful layer first
A focused first delivery might be a reporting table, one defined analytical workflow, or authenticated access to a few existing application operations. Expand when the questions and source coverage justify it.
Explore business data access for Claude and ChatGPT, or read the full marketing platform case.