Why I moved from enterprise delivery to AI agents
After 25 years building order management, payments and regulatory reporting platforms, I re-tooled around agentic AI for accounting and taxation. Here is the reasoning.
For most of my career the hard part of enterprise software was never the algorithm. It was the plumbing: getting an order from a website into a warehouse, a payment through Swift, a regulatory report out of the door before a deadline, and doing it at a scale where a bad afternoon costs real money.
That work taught me something that I think a lot of AI discussion currently misses. Enterprise systems are not short of intelligence. They are short of reliable connective tissue.
What changed in 2023
I stepped back deliberately. I wanted to understand generative AI properly rather than bolt a chatbot onto something and call it transformation. That meant certifications in Oracle and Databricks Generative AI and Oracle AI Vector Search, hands-on time with agentic workflow platforms like n8n and KNIME, and enrolling in the BS Data Science programme at IIT Madras.
It also meant going deep into a domain rather than staying horizontal. I chose accounting and taxation — ERPNext, Zoho Books, Tally — because it is a field where:
- the rules are explicit and auditable
- the cost of a mistake is immediate and measurable
- and an enormous amount of skilled human time still goes into moving data between systems
That last point is the opportunity.
Where agents actually help
The useful framing is not “replace the accountant”. It is closer to: what does a competent junior do all day that is entirely deterministic once you know the rules?
Reconciling a bank feed against ledger entries. Chasing a missing GST number. Normalising a supplier invoice into the right chart-of-accounts code. Flagging the three transactions in ten thousand that do not look like the other nine thousand nine hundred and ninety-seven.
An agent is good at these because each one is a bounded task with a checkable result. That is the key property. If you cannot check the output cheaply, you should not be automating it yet.
What I am building toward
The architecture I keep coming back to looks unglamorous:
- A real system of record — an actual ERP, not a spreadsheet
- A narrow, well-instrumented interface into it
- Agents that do bounded tasks against that interface
- A human review path for anything that fails a check
Most of the engineering effort goes into steps 1 and 2, which is exactly what 25 years of integration work prepares you for. The agent layer is the smallest part.
I will be writing here about what works and what does not — including the parts that fail, because those are more useful to read about.