We showed that one Tableau workbook can be rebuilt into Power BI with Microsoft’s agentic stack: Skills for Fabric, Modeling MCP, PBIP/PBIR, and an AI coding agent. The harder enterprise question is different. What changes when the estate is hundreds or thousands of interconnected workbooks?
A single marketing dashboard is a tutorial. An estate of 2,500 workbooks is a portfolio problem: discovery, shared sources, edge-case semantics, verification, and who owns the retry loop when the agent is wrong.
One workbook vs 2,500 workbooks
| One workbook | ~2,500 workbooks | |
|---|---|---|
| Unit of work | One PBIP, one page set, one prompt sequence | Waves by shared source, owner, or criticality |
| Context for the agent | Fits in one repo the agent can see | Must be sliced. The agent cannot hold the whole estate |
| Source of truth | Often embedded Custom SQL or one extract | Published sources, duplicates, and orphaned files |
| Repeatability | You can babysit prompts by hand | Need frozen prompts, versioned sources, retries, and QA gates |
| Semantic risk | A few calcs and one parameter | LODs, blends, RLS, parameters, and metrics reused everywhere |
| Done means | Desktop opens and looks right | Values and filter contexts match, then the next wave starts |
| Failure mode | Fix the file | Silent drift across hundreds of “successful” rebuilds |
Repeatability comes before clever prompts
At one workbook, you iterate in chat. At estate scale, treat the agent like a junior developer on a factory line.
- Prompts: freeze templates for measures, date tables, metric selectors, and page layouts. Change them on purpose, not mid-wave.
- Agent context: one shared semantic model or one subject-area repo per run. Do not dump 2,500 workbooks into a single session.
- Source versioning: Tableau packages and destination PBIP/PBIR under Git so you can diff and revert.
- Retries: schema failures and blank visuals are expected. Define when to re-prompt, when to fix by hand, and when to stop the wave.
- Human QA: every promoted report needs a named reviewer. “The file opened” is not a sign-off.
Portfolio discovery is the real first prompt
Before any agent rebuilds visuals, inventory the estate:
- Workbooks: owner, project, last viewed, last modified
- Published data sources and which workbooks depend on them
- Embedded Custom SQL and extracts that quietly duplicate a shared source
- Duplicates and retirement candidates
- Ownership gaps (orphaned assets nobody will validate)
That map decides waves. Shared sources become shared Power BI semantic models. Embedded one-offs stay local until reuse appears. Unused workbooks should not consume agent time.
Semantic edge cases multiply with fan-out
A demo workbook usually has SUM, DIVIDE, and a parameter. An estate adds:
- LOD expressions and table calculations that do not paste into DAX
- Blends that become relationships or upstream joins, not “another table in the PBIX”
- Parameters that must become field parameters, slicers, or what-if parameters on purpose
- Row-level security with a real identity path
- Metrics that look identical across workbooks but are not
Sample those patterns before you scale the one-workbook prompt loop. If the agent cannot explain the LOD or the blend on three critical workbooks, do not run it on three hundred.
Verification is matching values, not opening files
Agree reconciliation totals and filter contexts before rebuild. Compare the same date window, the same campaign or region filter, and the same grain. A report that loads with the wrong CTR or spend is not migrated. It is a new dashboard that happens to use similar fields.
How to combine AI, automation, and people
- AI-assisted agent work: bounded workbooks with clear sources, after discovery, with frozen prompts and Git.
- Automated migration platform: high volume, shared sources, and repetitive layouts where inventory and conversion are productized.
- Human expertise: metric redesign, security, ambiguous LODs/blends, and adoption after cutover.
Use them together. The one-workbook tutorial proves the agent path exists. Estate scale decides where that path is enough, where a platform carries volume, and where a person must own the definition of “correct.”
Start from the Tableau migration checklist
Before you scale the one-workbook agent loop, walk the Tableau migration checklist. That post is the full planning list. Here is the short version of what it covers:
- Who has access today: user roles and permissions
- How data arrives: pipelines, connections, and integration requirements
- How the move will be delivered: process and delivery model
- Data governance and quality rules you must keep
- Which models and workbooks are actually used
- A catalog of workbooks, sheets, and dashboards
- Infrastructure and tool fit, plus budget
- Adoption and what success looks like after go-live
At 2,500 workbooks, that catalog and the shared-source map come first. The agent rebuild is a later wave, and only for the workbooks the checklist says are worth moving.
If you need the inventory first, start with a migration assessment. The point of that step is the dependency map, not a commitment to any single rebuild method.
