The record-to-report problem space: verifying that an organisation's own accounting records agree with independent external evidence, that every balance can be substantiated, that every entry was properly authorised, and that the whole process completes on schedule with an audit trail. Distinct from the ERP, which records transactions but does not verify them.
The problem of organisations accumulating knowledge across dozens of systems, where an employee cannot reliably find it, trust it, or act on what it reveals. Scoped by who retrieves, not who benefits: an employee retrieving is in scope regardless of who eventually reads the output; a customer or an agent retrieving on a customer's behalf is not. Action is in scope only when it is knowledge-derived — it could not have been requested because nobody knew there was anything to request until synthesis revealed it. Not AI-scoped: the domain is organisations losing, mistrusting, and failing to act on their own knowledge, which predates AI by decades.
The problem of assembling live, structured, transactional enterprise data — a CRM opportunity, an SAP invoice, a ServiceNow ticket — correctly matched, permissioned, and formatted, into a prompt for a generative AI tool, across systems no single native platform grounding tool spans. Scoped to the cross-system, vendor-neutral layer specifically: native single-platform grounding (Salesforce Agentforce within Salesforce, Microsoft Copilot Studio within the Graph) is real and adjacent, but structurally excluded by definition, since it never reaches a system the platform doesn't already own.