Bringing AI into a traditional enterprise

Context

IT project manager at a fast-growing Taiwan insurance brokerage with a 700-strong sales force. Every back-office team runs on the same outsourced industry ERP — slow and buggy, its one real asset the decade of policy and customer data locked inside it — while the real workflows live in a sprawl of shared spreadsheets. My mandate grew from coordinating that vendor and modernizing internal workflows into building the arsenal a 700-strong sales force runs on. The IT unit is four people and I lead it technically, but everyone covers a different domain — so on these systems I am the whole function at once: the PM, the RD, and the one driving the AI agents. All ten are mine.

Problem

The company was locked in. The vendor holds not just the data model but the data itself — even though the raw data was the company’s property all along. Every enhancement came with an unimaginable price and lead time, and switching systems was a non-starter — migrating the data was the wall. So the ERP stayed slow and buggy while the company’s real work — from product data and incentives to policy outreach, scheduling, vendor exchange, and document management — lived in a sprawl of hand-maintained, error-riddled spreadsheets. High-frequency manual work across many teams, running on data nobody fully controlled.

Constraints

What I have to work with is the part of the database the vendor has released, read-only — I did not choose not to write to it; write access was refused. There is no documentation: the old system’s own screens are the only thing to check against, and the vendor will not explain its workflows either. The incumbent therefore stays the system of record until the tool layer we are building is ready to take over. Policy and customer records are personal data: everything stays on the intranet, never the cloud, on Linux hosts that accept connections only from designated machines, and the new systems persist no policy data at all, reading it read-only at compute time. ISO 27001 certification is in progress. And the tools had to be run day to day by non-technical staff — and to enforce compliance, not reproduce the shortcuts the old system quietly tolerated.

Approach

• Negotiate the data out first The vendor kept a tight hold on the data. With our IT team I spent over a year on it: the raw data was the company’s property to begin with, the vendor’s development capacity could not keep pace with how fast the business changed, and we were one of its largest clients. Those three together moved it — a little at a time, into read-only access to part of the database, and we are still prising out more. Without that year, nothing downstream happens.
• Reverse-engineer the flow, then optimize it No documentation, no cooperation — just some read-only tables and the old system’s screens to check against. I interviewed the staff in each department, then tested what they told me against those screens and tables, reverse-deriving field by field what the data actually meant and how it was really used. Only then did I rebuild — not copying the old process but straightening it: steps cut, shortcuts closed, compliance restored.
• Vibe coding I describe what a system needs in plain language and the AI writes it, iterating in a tight loop. That is how one person took each internal tool from idea to production in days instead of months.
• One consistent UI I fed the company's UI design guideline into that same AI workflow, so every system came out with the same look and branding automatically — ten systems, one interface, no per-project designer.

Outcome

Development output rose to 5–10× my traditional pace — enough for one person to ship a year of internal systems across many domains, cutting administrative processing time on the targeted workflows by ~60% (the investment-linked-life and A&H product master went from three working days a month to one hour; the incentive report from a week of manual calculation to real-time). That incentive report alone had been quoted at NT$200,000 by the vendor, to be built their old-school way; in-house it cost none of that and came out with access control and freely composable bonus schemes their version would never have had. Staff welcomed the automation, but the new tools enforce compliance rather than the shortcuts the old system tolerated, so adoption also meant holding people to proper process. What the year was really about is groundwork: running the brokerage’s tangled policy and administrative workflows through my own infrastructure once, cutting the total dependence on the vendor so future integration can move faster. A RAG consultation system is already underway for the far more varied P&C lines — property, auto, personal accident, pet, and more — letting an agent describe a client’s needs by voice or text and surface the products that fit. All of it points toward the migration off the vendor, and toward layering AI directly on the company’s own data and processes.

Stack

Claude (vibe coding) · React · FastAPI · Python · MySQL · RAG