What is agentic ERP? The category, explained while it's being born

Categories get named before they get defined, and "agentic ERP" is in that window right now: no single coiner, the phrase spreading through vendor guides, integrator writeups and analyst commentary through 2025–2026 (the research record collects the trail). Salesforce markets its CRM with the word front and center — "agentic CRM" headlines salesforce.com/crm — and the ERP twin of that vocabulary is being written as we speak. Since we build in this category, here is the definition we would defend in a room of skeptics — including the part most agentic marketing omits.

#The definition, and the line it draws

Agentic ERP: an ERP where AI agents execute multi-step business processes — reading context across modules, applying constraints, taking the operational actions — rather than assisting a human who executes them. The line it draws is against assistive AI, and the test is who performs the state change. A copilot that summarizes the pipeline, suggests the reorder, drafts the email while you drive: assistive. An agent told "handle the Birchwood order" that reads stock and pricing, builds the quotation, and moves the process forward: agentic. Across the 2025–2026 writeups in the record, the common thread is exactly this executes-the-process distinction — vendors differ mainly on how loudly they say the next part.

#The part the marketing omits

An agent that executes processes in your system of record has, by definition, write authority over your business state — which means the category's real questions are not intelligence questions. Whose permissions does the agent act under? What happens at the authority line — the draft-becomes-posting boundary? What record does each decision leave? The definitions that make it into vendor decks say "executes multi-step processes"; the ones written for COOs quietly add "and asks for human approval where it matters." That clause is not a limitation of agentic ERP. It is the condition under which agentic ERP is adoptable at all — an agent without governance is not a product category, it is an incident category.

#Agentic ≠ autonomous, and the maturity dial

The category's most useful clarification: agentic describes capability shape (perceive, decide, act across steps), not authority level. A fully governed agent that drafts everything and commits nothing is still agentic — the process intelligence is identical; only the dial differs. That framing turns adoption from a leap into a slope: the same agent runs read-only in month one, drafts-with-approval in month two, and earns per-workflow autonomy as your own trail accumulates evidence. Vendors who conflate agentic with autonomous are asking you to skip the slope; the ones building for the category's actual buyers are building the dial.

#What the pattern requires underneath

Strip the vocabulary and agentic ERP needs four things: cross-module context (a process is never one table — the order touches stock, pricing, credit); typed actions the agent can take, not just text it can emit; a policy layer deciding allow/hold/forbid per action; and a decision trail including refusals. That is the whole architecture, and it is what we build — governed, action-taking AI on Odoo, where the "agentic" and the "governed" are one design rather than a feature and an apology. Whether you evaluate us or anyone: buy the four requirements, not the adjective.

#Frequently asked questions