Leaders are tired of choosing between a bloated suite and a chaotic best-of-breed pile. The workable 2026 answer is composition: ERP remains the financial and inventory system of record, specialized SaaS owns differentiated workflows, and AI sits on top as a retrieval and drafting layer—not a second ledger.
Assign systems of record on purpose
A sample operating-system map
| Domain | System of record | AI role |
|---|---|---|
| Money & stock | ERP | Exception summaries, coding drafts |
| Pipeline & accounts | CRM | Call notes, forecast narrative |
| People | HRIS | Policy Q&A, ticket deflection |
| Product usage | Data warehouse / CDW | Anomaly narratives |
| Documents | Controlled KB | Cited answers only |
The fabric between them
iPaaS or an event bus should move canonical IDs—customer, item, invoice—not shadowy CSV dumps. Every integration needs an owner, a freshness SLO, and a visible error queue. AI retrieval should read from this fabric or from certified replicas, never from random exports sitting in a shared drive.
Interface layer rules
- One employee-facing copilot experience if possible—multiple brains, one chrome.
- Every answer cites the system of record object.
- Actions require the same approvals as the native UI.
- Prompt logs are retained like other business records when they influence money or HR.
A 180-day composition program
Days 1–30: publish the system-of-record map. Days 31–90: fix the worst identity and integration breaks. Days 91–180: launch one cross-app copilot on a single job (quote-to-cash exceptions or employee IT how-to). Expand only when override rates fall.
Conclusion
Composing AI, ERP, and SaaS into one operating system is an architectural and political act. Name systems of record, professionalize the integration fabric, and let AI draft—not dominate—decisions. Companies that do this will feel smaller and faster. Companies that skip it will keep buying tools to compensate for the last tool.