Why a fast answer can still be wrong

Brazilian marketing agency Convert Human Digital Business faced a problem that was not a shortage of charts, but conflicting answers to the same client question. Advertising, sales, and CRM data were available, yet metric definitions, segmentation rules, and familiar shorthand often lived in analysts' heads. If one employee calculated revenue after returns and another before returns, a generative model could confidently accelerate both incompatible answers.

In a published Google Cloud case study, Convert says it first tried agents that queried SQL directly and interpreted the results. They handled simple questions, but more complex ones lacked the meaning behind business measures. That distinction matters to any company hoping to “ask its database in plain language”: access to a table is not knowledge of what a column means or which exceptions should be counted.

What the company built

According to the project description, client marketing and transactional data are collected in BigQuery. A Looker/LookML semantic layer sits above the warehouse and documents metrics, dimensions, synonyms, and client-specific rules. Five specialized agents then cover media performance, advertising spend pacing, search, creative performance, and ecommerce. These are roles within one analytical system, not five independent sources of truth.

A practical request path is:

  • a person asks a question about a campaign or sales;
  • the system identifies the domain and uses an agreed metric definition;
  • a query runs against the authorized data for the relevant client;
  • an analyst reviews the sourced calculation before it reaches the client if an error could affect spending or reporting.

The final step is our editorial recommendation for safe use, not a claim that Convert implemented exactly that approval stage. The case study does not describe permissions for every dataset, how disputed answers are reviewed, or the error rate. Its design cannot be transplanted without making those decisions.

What is documented and what remains a claim

Google Cloud quotes Convert's team as reporting 65% less time to deliver reports, 8.4 times faster complex analysis, and 32% lower operating costs. These are results for one company in a publication by its technology supplier. The baseline period, workload mix, cost components, license expenses, and independent verification method are not disclosed. Treating those percentages as a guaranteed saving for another agency would be wrong.

The mechanism is more transferable. A semantic layer defines a calculation once and makes it available consistently to BI and agents. Independent research on natural-language-to-SQL tasks, such as BIRD, shows why real database values and domain context complicate the path from a user's phrase to a correct query. A dictionary is not magic either: a wrong metric definition will produce consistently wrong answers. A process owner, not a model, must approve the meaning of revenue, lead, or margin.

What a small or midsize business needs

Do not start with five agents and a new warehouse. For a first pilot, choose one painful group of questions, such as a weekly report on advertising spend and confirmed orders. You need accessible exports from the ad platform, CRM, and accounting system; a shared customer or order key; a dictionary of 10–20 measures with formula, period, currency, exclusions, and owner; and examples of previously checked answers.

If the business already has BI and a SQL warehouse, define the semantics there first. A local language model may parse questions and draft explanations within the company's own environment, but its location does not reconcile inconsistent data. RAG can retrieve textual rules, policies, and notes about measures; numerical totals should come from an auditable query, not a retrieved text passage. An agent allowed to change campaign budgets is a separate phase requiring constrained permissions and human approval.

How to measure value without borrowing percentages

Before deployment, collect 30–50 typical questions from the previous month and record current preparation time, correction rates, and review cost. For two weeks, compare system answers with analyst-approved ones: are filters and periods correct, do totals reconcile with reports, and is the source traceable? Count only answers accepted without substantial rework.

The cost model should include more than model calls: data preparation, licenses, storage, metric-dictionary maintenance, review time, and error correction. Compare cost per accepted answer or completed report, not price per million tokens. If freed hours cannot be redirected to paid or genuinely useful work, the result is increased team capacity rather than an automatic cash saving. That pilot will answer the payback question for your own business without purchasing Convert's entire architecture.

Primary source: https://cloud.google.com/customers/convert-human-digital-business