A week of client reporting, done in two hours: restoring an agency’s faith in AI.

Critical client reviews, a reporting team at 200%, and an AI attempt that failed.

The agency runs high-ticket, low-frequency, high-intent campaigns for large companies, enterprises, and US government agencies, on channels like Disney+, HBO, and Twitch. Budgets run from thousands to millions of dollars, and campaigns last months or a full year. Every month or quarter, the agency proves progress against the numbers it committed to, in short but critical meetings. Behind each one: extensive, meticulous, repetitive work crossing data and turning it into a presentable deck. Two people under the VP of Operations did it by hand, running at 200% in the last week of every month. They had tried Claude on their own by feeding it massive raw CSVs with loose instructions. It failed, and the VP lost faith in automation. One of the partners, also the CFO, brought me the challenge.

Before and after.

1. Extract programmatically, clean, land in BigQuery

   Instead of pasting raw files into a chat, the data is extracted by code, cleaned, and stored in BigQuery, so the AI always starts from reliable data.

2. Turn what lives in people’s heads into SQL

   The context and business rules that only the reporting team knew were captured, documented, and translated into SQL.

3. Views built for the AI to read

   Pre-aggregated views with the right dimensions, metrics, and granularity, served to the AI through MCP. The model reads exactly what it needs, nothing more.

4. Ask for one thing at a time

   Analysis, insights, and report assembly are requested in separate steps. Each step does one job well, starts from ready data, and uses AI only where it adds value.

5. Auditable, and bulletproof for non-technical users

   Every number can be traced and verified. The team just drops what they need in Google Drive; I absorb the complexity. When something cannot be extracted or modeled properly, the system raises an alert instead of guessing.

6. Then: a sales-intelligence system

   AI agents scan selected sources and markets for high-value, high-fit opportunities. They score, digest, evaluate, and filter them, then send the best leads to each rep by region and match score, all through Slack. A person makes the final call and can add filters through a small app. The same engine now covers RFPs, public RFPs, and, starting now, recently funded companies.

1. A week of work in two hours, with fewer errors

   The pilot produced in 2 hours the reports that took the team a week, and helped find errors in the reports they had built by hand.

2. Plug and play for every new client

   Now in production. Each new client, channel, or campaign needs one context-capture session; the infrastructure is ready to scale.

3. Sales intelligence for $20 to $50 a month

   The only running cost is Gemini and Claude tokens, optimized for the highest possible return.

4. Faith in automation, restored

   After the impact of the first project, the agency started delegating more projects to me.