ERP chatbots in 2026 — from answering to doing

"ERP chatbot" is the phrase people search; what they find spans everything from a FAQ widget bolted to a login page to full agents that move business processes. The category term has outlived the category — so this piece sorts the 2026 landscape by what actually matters: what each tier does, where users report the pain, and the questions that separate a demo toy from something your operations can lean on. Disclosure: we build at the acting end of this spectrum; the sorting below is written to be useful wherever you land.

#The three tiers hiding under one word

Tier one: scripted responders. Intent-matching bots answering "how do I reset my password" from a knowledge base. Cheap, fine for deflection, nothing to do with your ERP data.

Tier two: conversational query. The real step: chat that reads live ERP records — "what's the stock on X?", "which invoices are overdue?" — turning natural language into data access. Many products marketed as ERP chatbots sit here, and the complaints concentrate here too: answers without sources, context lost between questions, and the demo-versus-production gap when a bot trained on clean demo data meets a real database's ambiguity.

Tier three: acting assistants. Chat that doesn't stop at the answer — it prepares the next step: the reminder drafted, the quotation built, the transfer staged. This tier crosses the agentic line, which is exactly why it must arrive governed: whose permissions, what approval, what log.

#What users actually complain about

Common complaints in published reviews and user forums — collected in this piece's research record — are strikingly consistent, and worth using as your evaluation lens. Unsourced answers: a number with no record behind it is a guess wearing a suit — click-through-to-the-record is the fix, and its absence is disqualifying for finance use. Goldfish memory: bots that lose the thread between related questions make multi-step work slower than the screens they replaced. The confidence problem: wrong answers delivered fluently — worse than no answer, because fluency gets trusted. The demo gap: everything works on the vendor's data; the pilot on yours is where truth lives.

#The questions that sort the market

Six, in the order that eliminates fastest. Does every answer cite the record it came from, with a click-through? Does it read your live database or a synced copy — and how stale? Whose permissions does a question run under — the asker's, or a service account that sees everything? When it acts (tier three), what stands between the draft and the commit? Where do questions containing your data go, and does anything train on them? And the log: can you see who asked what, including the refusals? Products at every tier fail some of these; knowing which failures you can live with is the actual selection decision. Our answers are on the Ask Odient page — sourced answers, live reads under the asker's own permissions, actions held for approval — and role-scoped variants on the assistants page.

#Frequently asked questions