Building Agentic AI solutions in enterprise environments✦Training for IronMan 2027✦Building Aether Ops · AI-era governance for small, mid & large enterprise✦Ultra Marathon 2028 · because ERP wasn't hard enough✦Leading system integration post-$13B IPG acquisition at Omnicom✦Finance nerd by day. AI builder by night. Endurance athlete on weekends.✦8+ years · $17B+ in transactions · 150+ global business units✦Currently shipping: Aether Ops · Onyx · and things I can't talk about yet✦Stevens Institute · MS Business Intelligence & Data Analytics✦Ask my agent anything 👇✦Building Agentic AI solutions in enterprise environments✦Training for IronMan 2027✦Building Aether Ops · AI-era governance for small, mid & large enterprise✦Ultra Marathon 2028 · because ERP wasn't hard enough✦Leading system integration post-$13B IPG acquisition at Omnicom✦Finance nerd by day. AI builder by night. Endurance athlete on weekends.✦8+ years · $17B+ in transactions · 150+ global business units✦Currently shipping: Aether Ops · Onyx · and things I can't talk about yet✦Stevens Institute · MS Business Intelligence & Data Analytics✦Ask my agent anything 👇✦
HomeAboutNotesBuildsStack
Schedule↗

Govind Waghmare

Agentic AI for ERP. Cutting through red tape. Building in public from NYC.

govindwaghmare@icloud.comcal.com/govindw

Pages

HomeNotesBuildsAboutStackPartners

Social

XLinkedInGitHub
© 2026 · New York City·RSS·llms.txt

Every Agent Is a SOX Risk

Jun 18, 2026

Here is the sentence that should be keeping finance leaders up at night: the moment an AI agent can post a journal entry, change vendor master data, or approve a transaction, it is no longer automation. It is a control. And controls get audited.

Most companies are not treating it that way yet. They're stacking agents on top of ERP the way they stacked macros on top of spreadsheets, fast and ungoverned, and they will find out the hard way that an auditor does not care whether the thing that touched the general ledger was a person or a process.


The non-human identity problem

The scale is the part people miss. In the average enterprise, non-human identities already outnumber human users by roughly 80 to 1. One large financial institution reportedly had over 4.2 million non-human identities against about 50,000 human accounts. Nearly two-thirds of organizations apply weaker controls to those identities than to their employees.

That is a segregation-of-duties failure waiting to be written up. If an agent can both create a vendor and approve a payment to it, you have a problem, and it does not matter that the agent had no intent.

Every agent that touches a financial system needs the same discipline a human user gets: a real identity, least-privilege scope, and a joiner-mover-leaver lifecycle. Provisioned when it's deployed. Re-scoped when its job changes. Revoked when it's retired.


The clock is real

By August 2026, most obligations for high-risk AI systems under the EU AI Act become enforceable: risk management, technical documentation, logging, and post-market monitoring. The SEC's cyber disclosure rules already force material incidents onto an 8-K within four business days. The regulatory environment is not waiting for enterprises to figure out agent governance.

The honest guidance for this year is unglamorous: agentic AI belongs in human-supervised roles for control work, not autonomous agent-led testing. Let agents draft, collect evidence, and reconcile. Keep a human on the sign-off until the audit trail proves itself.

The companies that win the next finance cycle won't be the ones with the most agents. They'll be the ones who can prove, on demand, exactly what every agent did and why it was allowed to. That proof is the product. Everything else is a demo waiting to fail an audit.

Ask a follow-up about this note
$