AI agents and document automation inside your ERP
Assistants that answer from live ERP data and draft records for human approval, invoice and PDF extraction with review, bounded by permissions and cost ceilings.
What this covers
- An agent in the Desk that answers from live records — aggregations, search, link tracing and saved reports — with no vector index to keep in step
- Digital and scanned PDFs, photos and e-invoices read into draft Purchase Invoices and Expense Claims
- Supplier and line-item matching, totals checks and a confidence score on every extraction
- Every write is a draft a person approves, denies or edits; nothing posts on its own
- Several LLM providers behind one interface, with failover, so one outage is not your outage
- Bounded by the same permission rules as the user behind it, and by a cost ceiling you set
How an engagement runs
- 01
A first conversation
You describe what is slow, manual or fragile. I ask questions, not for a budget, and say plainly whether it is work I should do.
- 02
Scope in writing
What changes, what does not, how we will know it worked, and what I need from you: access, a copy of the site, a person who knows the process.
- 03
Build on a copy
Work happens on a staging copy of your site. You review it there before anything touches production.
- 04
Go-live with a gate
Releases and migrations follow a rehearsed plan with a tested rollback. Nothing writes to production without approval.
- 05
After go-live
Fixes, monitoring and upgrades under support and maintenance, if you want them.
Related work
Questions I get asked
- Will it change our data on its own?
- No. Anything that writes stops at a draft with a review card: what it read, what it matched, what it will post. Approve, edit or deny — the agent does not decide.
- Does it need our data sent somewhere and indexed?
- No index and no copy. The agent queries the live database through tools at question time, which is also why its answers cannot go stale.
- What stops it from reading records the person may not see?
- It runs as that user. The permission rules that apply in the Desk apply to the agent's tools, so it cannot read or write anything the person behind it could not.
- Is this only for ERPNext?
- The deepest work is inside Frappe and ERPNext, but the same pattern has been built into Odoo — a GPT vendor-bill reader prototyped on v15, and an agentic automation engine on v19.