What forty AI inventories taught us about shadow AI

The declared number is usually three or four. The found number is usually between twenty and forty. The gap is not deceit. It is definition. Most organisations count the systems they bought on purpose. Almost none count the AI that arrived inside something else.

Departmental SaaS with embedded AI: the HR tool that started scoring résumés in a March release note nobody read. Browser extensions: summarisers and writing assistants with access to everything on the screen. Personal LLM accounts used on company data, invariably by your most productive people, which is why bans fail. And vendor systems with model components the vendor never disclosed, because the contract never asked.

Under DPDP, fiduciary duty does not accept ignorance of where personal data goes. Under SEBI 16C, responsibility does not accept it either. The inventory is not the compliance exercise; it is the precondition for every other control you will ever be asked about.

One page per system, seven fields: name, owner, data touched, vendor, model components, decision impact, last reviewed. If your register needs a workflow tool and a steering committee, it will be accurate for one quarter and fictional by the third.