What to keep in your ERP when you add AI-native tools
The ledger, inventory valuation and job costing need a system of record. Order entry, purchasing and follow-up often do not. A practical way to draw the line.
Contents
Adding AI to a mid-sized company's back office raises an architectural question before a commercial one: which system holds which record. Get it wrong and the company ends up with two versions of its stock, its orders or, worst of all, its books.
The case for a single system of record
An ERP exists so that every transaction is recorded once and flows to the financial statements. Microsoft's manufacturing documentation for Business Central shows how tightly linked those flows are: production orders consume materials, post output and scrap, update inventory and calculate the cost of finished items "for financial reconciliation"1. Pull any one of those steps into a separate system without a defined write-back, and the reconciliation breaks.
The vendors that sell AI operations layers acknowledge this. OpenLazarus's manufacturing page says it reads jobs, routings and inventory from the ERP and writes back what changes "so there is still one record", and that most shops keep the ERP for job costing and the books2. DOSS says it adds control over operations without replacing the accounting system, and connects to QuickBooks Online, Xero or NetSuite as the ledger3.
What to keep, and what can move
| Record or task | Keep in the ERP or ledger | Can move to AI agents | Why |
|---|---|---|---|
| General ledger, close, consolidation | Yes | Preparation only | Auditors need one trail from entry to statement |
| Revenue recognition | Yes | No | Rules-driven and audited |
| Inventory valuation and job costing | Yes | No | Feeds cost of goods sold |
| Stock counts and reorder signals | Master stays | Monitoring, alerts, PO drafts | Agents act on the count; the ERP values it |
| Order entry from email, phone and fax | Order record stays | Reading, matching, drafting | The slow, error-prone step is re-keying |
| Purchase orders and supplier follow-up | PO record stays | Drafting, sending, chasing | Most of the time goes on correspondence |
| Bank matching | Posting stays | Proposing matches | A person or rule confirms the posting |
| Collections reminders | Ledger stays | Reminders and payment matching | Low risk, high volume |
The pattern holds across vendors. Business Central's own agent creates a sales quote before any order, because a quote has no effect on planning, reservations or cash-flow forecasts4. OpenLazarus's order-desk teammate builds the order "for a person to confirm", and anything it is unsure of waits for a person5.
Where OpenLazarus draws the line
Because OpenLazarus is the top pick in our main guide, its boundary is worth stating exactly. Its bookkeeping article describes agents that categorise transactions, reconcile bank accounts and produce statements from a ledger kept as structured files in a shared workspace6. That suits a small company's books. For a mid-sized company with audited accounts, we read its pages as a layer that prepares and feeds the ledger, consistent with its own advice to keep the ERP for the books2. Its pages do not describe consolidation, revenue recognition or inventory costing.
A four-question test for each process
Run each candidate process through four questions before moving it.
- Does it post to the financial statements? If yes, the posting stays in the ledger; only preparation moves.
- Is there one master for every field it touches? Write down which system owns the customer, the price, the stock count and the order status.
- What is the approval limit? OpenLazarus's examples are purchase orders above $15,000 and credits above $5005; set your own in writing.
- Can the auditor trace it? Every agent action needs a log your staff can read; OpenLazarus logs every action where people can see it5.
A sound split for most mid-sized companies
- The ledger, close, revenue recognition and inventory valuation stay in one audited system.
- Agents take the correspondence and data entry that feed those records.
- Every agent has written limits, and every action is logged.
- Write-back to the ERP is tested before go-live, not after.
The migration guide shows how to sequence the move once the line is drawn.
Questions buyers ask
Should we replace our ERP with AI agents?
For most mid-sized companies, not the ledger. The general ledger, consolidation, inventory valuation and job costing need controls and an audit trail that a dedicated ERP or AI-native ledger provides. The daily work that feeds those records, such as order entry, purchase orders, supplier chasing and bank matching, can move to agents that write back to the ERP.
Can an AI layer and an ERP stay in sync?
Vendors that sell operations layers say so. OpenLazarus says it reads jobs, routings and inventory from the ERP and writes back what changes, so there is still one record. DOSS integrates with QuickBooks Online, Xero or NetSuite as the general ledger. Ask any vendor which system is the master for each field and how conflicts are resolved.
What should never be automated without approval?
Anything that commits money or changes a promise to a customer. Vendors' own examples include material orders above a set value, credits above a threshold and moving a promised ship date. Those limits should be written down per agent before go-live.
Sources
- Manufacturing overview: Business Central. Microsoft Learn, 15 Jul 2026. Accessed 8 Nov 2026.
- Manufacturing. OpenLazarus. Accessed 8 Nov 2026.
- DOSS: the AI-native operations platform for consumer goods. DOSS. Accessed 8 Nov 2026.
- Sales Order Agent overview: Business Central. Microsoft Learn, 2 Oct 2026. Accessed 8 Nov 2026.
- Wholesale and distribution. OpenLazarus. Accessed 8 Nov 2026.
- AI Bookkeeper: replace QuickBooks, Xero and FreshBooks. OpenLazarus, 21 Dec 2025. Accessed 8 Nov 2026.
Updated 8 Nov 2026: First published.