One working day, from zero to a daily revenue digest in Slack and three dashboards you actually trust. Tier 1 (CSV drop), so you can do the whole thing without procurement, a warehouse, or a new contract.
This is the order I'd run if you handed me a laptop, a Claude seat, and your Shopify login at 9 AM and wanted something real by 5 PM. It's Tier 1 on purpose: you export CSVs, Claude reads them. No MCP connectors, no BigQuery, no data engineer.
Tier 1 is the least automated path and the least trustworthy on arithmetic, so every step below carries a verify clause. That part is not optional. A model can mis-add a 200k-row CSV, reference a column that isn't there, or quietly compute against last month's export. What you build today is a fast analyst, not a system of record. Shopify and your bank are still the system of record.
One honesty note before you start. The "$20 Claude seat replaces a four-figure SaaS bill" line only holds at Tier 1, where you do the exporting by hand. The moment you want this to pull data on its own, you're into paid third-party MCP middleware or a warehouse, each with its own bill (per-brand connector and warehouse cost varies by vendor and data volume and isn't consistently published, as of 2026, verify). Today's build is the cheap version deliberately. It's also the version most likely to keep a human in the loop, which at the start is a feature.
A realistic expectation: by end of day you have the digest plus three dashboards solid (MER, CAC, contribution margin). The other three (new vs returning, cohort LTV, inventory cash) are built and sanity-checked, but you'll keep refining their assumptions for a week. That's normal. Don't let the cohort table hold up shipping the digest.
Three things open so hour 1 isn't a hunt:
You also need a paid Claude plan. Pro at $20/mo covers Tier 1 with lean CSVs (pricing as of 2026, verify at platform.claude.com/docs/en/about-claude/pricing). If your exports are large or you want headroom, Max. The Slack digest in hour 3 needs the Slack connector, which needs a paid plan and Claude Desktop.
Goal: every CSV the day depends on, on disk, filtered lean.
Lean matters more than it sounds. A model's arithmetic accuracy degrades as the file grows, and you pay tokens for every row, so the same export filtered to the columns and date range you need is both cheaper and more accurate than the raw dump. Filter before you export, not after.
Pull the last 13 months. You want a full year plus the current partial month so year-over-year and trailing-12 both work: