We Replaced ETL with MCP
rashidazarang.com
rashidazarang.com
I am taking screenshots of blogposts like this for a museum exhibit opening next year - lmk if you’re willing.
Here's what actually happens:
1. MCP exposes system schemas in a standardized way 2. AI analyzes the schemas and suggests mappings 3. Engineers review and validate every mapping 4. AI generates deterministic integration code (think: writing the SQL, not running it) 5. We test with real data before any production deployment
This sounds like a terrible idea, and nearly impossible to debug when it inevitably drops data.
That’s exactly why we don’t let AI run migrations. We use it to speed up the boring parts, like mapping table structures. But humans are always in control.
I'll give an example. I worked for a FTSE 100 company using a very old Product Lifecycle Management system (model manager - based actually on pre-DOS technology)....we had to upgrade it to a new fancy one.
Therefore we had to migrate all data relating to the company, and group companies engineering designs...everything to do with 2D drawings, 3D designs...any important connections etc....all electrical designs....excel sheets related to these containing lists of PCBs and their component parts in Bills Of Materials etc...There is absolutely no way in hell I would trust AI with almost any of that, to get it right....or even to attempt a load without almost immediately erroring.
But here’s what we do use AI for: • Mapping legacy schemas • Spotting patterns • Generating boilerplate ETL code fast
Then humans step in: • Validate every mapping • Write custom logic for edge cases • Test everything... every field, every BOM, every relationship • Migrate with deterministic, human-reviewed code
Meaning, correctness, completeness, etc...
Would you use it for e.g. tax information? Because if wrong, you could get fined.
Correctness: 100% schema mapping accuracy after human validation. We've never had a data type mismatch or field misalignment make it to production. The AI suggests mappings at ~85% accuracy, humans catch and correct the remaining 15%.
Completeness: Zero data loss incidents. We run reconciliation reports comparing source record counts to destination. Any discrepancy fails deployment. Most common issue: the AI initially missing compound key relationships, which we catch in testing.
Tax/Financial Data: Yes, we handle financial data for several clients, including:
QuickBooks to data warehouse pipelines (invoice/payment data)
Payroll system integrations
Revenue reconciliation between CRM and accounting
Our approach for sensitive data:
AI generates the integration logic, never sees actual records
Test with synthetic data matching production schemas
Run parallel processing for 1-2 cycles to verify accuracy
Maintain full audit logs of all transformations
Human sign-off required before production cutover