Any ERPUse Case

AR collections + private AI

AI for ERP Accounts Receivable and Collections, Grounded in Your Open Items

Short answer

AI for ERP accounts receivable reads open customer items, payment history, disputes and order status from the ERP, then ranks who to call first, drafts the follow-up with the real invoice facts, and explains why DSO moved. It does not post cash or release holds on its own: a collector approves every action, and every number traces back to a query against the ledger.

ERP: NetSuite, SAP S/4HANA, Infor SyteLine, Dynamics 365 Finance, Epicor Kinetic
Industries: Manufacturing, Aerospace, Industrial Distribution
Written for: Controller

Discuss your NetSuite environment See use cases

Most manufacturers run collections out of an aging report and a spreadsheet. The aging report says who is late; it does not say why. A past-due invoice may be late because the customer disputes a price, because the shipment was short, because a certificate of conformance never reached their receiving dock, or simply because their payment run is monthly. The collector has to open the ERP, the shipping record, the email thread and the dispute log to find out, and that detective work is where the day goes.

ERP-native dunning is deliberately mechanical. SAP's dunning run (F150), NetSuite's dunning letters, Dynamics 365 collections and Business Central reminders all escalate by days overdue and balance. They cannot tell the difference between a customer who always pays on day 45 and one who has gone quiet, and they cannot read a dispute note. So collectors either override the automation by hand or ignore it and work from their own list.

An AI layer closes that gap by combining the structured ledger (open items, payment terms, promise-to-pay dates, credit memos, unapplied cash) with the unstructured context around it (dispute notes, delivery confirmations, email threads). The goal is not to automate the relationship. It is to hand each collector a ranked worklist with the reason for each account and a draft message that already cites the right invoice numbers, amounts and ship dates.

This page describes what that looks like on the major ERPs, what it should never do, and what to ask a vendor before connecting an AI agent to customer financial data. For controllers at defense and aerospace suppliers, there is also the question of where the data runs, which is why private and on-prem deployment is a first-class option.

What usually gets in the way

The problems we hear most from controller teams running NetSuite.

Aging buckets hide the reason for lateness

A 60-day bucket mixes genuinely slow payers, disputed invoices, unapplied cash and invoices the customer never received. Collectors spend most of their time classifying, not collecting.

Cash application leaves unapplied balances

Remittances arrive with partial references, short pays and bundled invoices. Unapplied cash sits on account, so aging looks worse than reality and collectors chase customers who already paid.

Disputes live in email, not the ledger

Pricing, freight and quantity disputes are discussed by email and phone. The ERP shows only an overdue balance, so the next collector repeats the same call.

Dunning ignores customer behavior

Rule-based dunning escalates every account on the same schedule. It annoys reliable payers and under-reacts to the account whose pattern has quietly changed.

DSO moves and nobody can say why

When DSO rises two days, finance rebuilds the story by hand: mix of large customers, a billing delay, one big dispute. The answer arrives after the month-end meeting.

Where AI earns its place in NetSuite

Each use case names the ERP objects it reads or writes, so your ERP team can judge the integration effort before anyone commits budget.

Ranked collections worklist with reasons

Each morning the agent scores open items by amount, age, customer payment pattern, dispute status and promise-to-pay history, and lists the top accounts with a one-line reason such as 'pays on day 42 historically, now day 58, no dispute logged'.

Touches: NetSuite customer and invoice records, SAP BSID/BSAD open and cleared items, SyteLine customer ledger, D365 CustTrans

Outcome: Collectors start on the accounts where a call changes the outcome; fewer calls to customers who are about to pay anyway.

Drafted follow-ups with verified facts

The agent drafts the email or call script using the actual invoice numbers, amounts, ship dates, PO references and proof-of-delivery, so the collector edits instead of researching.

Touches: Invoice header and lines, shipment and delivery records, customer PO, contact roles

Outcome: Shorter prep per account; fewer messages with the wrong amount or missing PO.

Cash application suggestions

For remittance advices and bank lines with partial references, the agent proposes invoice matches and flags short pays with the likely cause (discount taken, freight deduction, tax difference). A person confirms before anything is applied.

Touches: Customer payment, payment application, credit memo, bank statement lines, SAP FF67 / F-28 equivalents

Outcome: Lower unapplied cash balance and fewer false past-due flags, with the final posting still a human action.

Dispute triage and routing

The agent classifies dispute notes (price, quantity, quality, freight, paperwork), attaches the order and shipment evidence, and routes to sales, logistics or quality with the question already framed.

Touches: Sales order, shipment, return authorization, quality notification, dispute case tables

Outcome: Disputes reach the right owner sooner; aged disputed balances shrink over a few quarters.

DSO and aging explanations

Ask 'why did DSO go up in September' and get a decomposition by customer, billing lag, terms mix and disputed amounts, with the SQL or API query shown so finance can verify it.

Touches: Invoice dates, terms, payment dates, revenue by customer, credit memos

Outcome: Month-end variance commentary prepared from data in minutes instead of a day of spreadsheet work.

Promise-to-pay tracking

The agent reads logged promises, compares them with actual receipts, and surfaces broken promises so follow-up happens on the day, not at the next aging review.

Touches: Collection notes, activity log, customer payment records

Outcome: More consistent follow-up on commitments without a separate tracking sheet.

Statement and invoice-copy requests

Customer AP teams ask for copies, statements and remittance details. An internal assistant retrieves them from the ERP and drafts the reply for approval.

Touches: Invoice PDF archive, statement of account, customer contact records

Outcome: Routine requests are answered same day with less collector time.

Reference architecture

A read-only retrieval and reasoning layer sits between the AR ledger and the collector. Structured data comes from supported ERP interfaces; documents and emails are indexed separately; an open-weight model runs inside your boundary; any action proposal passes through a human approval step.

1
ERP data connectors
Read-only service account using the ERP's own interfaces: SuiteQL or SuiteTalk for NetSuite, OData/CDS views for S/4HANA, IDOs or SQL views for SyteLine, data entities for D365. Row-level access mirrors the collector's ERP role.
2
Document and email index
Dispute emails, proofs of delivery and remittance PDFs are chunked and embedded into a private vector store, tagged by customer and invoice so retrieval stays scoped.
3
Model serving
An open-weight model served with vLLM or Ollama on your GPUs or in a private cloud. No customer balances or contact data go to a public API.
4
Scoring and grounding
Prioritization is a transparent scoring function over ledger features; the model explains it rather than inventing it. Every number in an answer shows the query that produced it.
5
Approval and audit
Draft messages, match suggestions and hold-release requests are queued for a named approver. Prompts, retrieved records and decisions are logged for audit.

Integration notes for your ERP team

Use a dedicated read-only integration user per ERP; do not reuse a collector's login. Map it to a role limited to AR, customer master and shipment views.
NetSuite: SuiteQL against the transaction, customer and transactionline tables, or saved searches exposed through RESTlets; respect concurrency governance on the account.
SAP S/4HANA: prefer released CDS views and OData APIs for open items over direct table reads; on ECC, read BSID/BSAD/KNA1/KNB1 through RFC or a replication layer.
Infor SyteLine: read customer ledger and invoice data through IDOs or a read replica and keep any write-back, such as hold release, behind an approved workflow rather than direct SQL.
Dynamics 365 and Business Central: use data entities or OData web services for customer ledger entries, and keep reminders and collections journals as the posting path.
Index email by invoice and PO references so a dispute thread can be joined to the ledger row it concerns.
Never allow the model to post cash applications, credit memos or hold releases directly; it proposes, a person posts.

Deployment options

On-premises GPU server

Defense and aerospace suppliers, or customers whose contracts restrict financial data movement

Model, vector store and connectors run inside your network next to the ERP. Nothing leaves the boundary.

Private cloud or VPC

Mid-market manufacturers on cloud ERP with a controlled cloud tenancy

Dedicated model serving in your cloud account with private networking to the ERP API; no shared multi-tenant inference.

Hybrid with local retrieval

Teams that want a faster start with sensitive data kept local

Ledger queries and document retrieval stay local; only de-identified aggregate prompts, if you choose, go to an approved external model.

Compliance and data control

How the architecture supports your obligations. Certification and accountability stay with your organisation; the design keeps the evidence straightforward.

SOX internal controls

The assistant is read-only against the ledger, so it does not alter the control environment. Approvals, prompts and outputs are logged, which supports evidence for access and change reviews.

GDPR and local privacy law

Customer contact data stays in your environment. Retrieval is scoped by role, and retention for prompts and indexed emails follows your policy.

FDCPA and collection practice rules (consumer contexts)

Most manufacturer AR is business to business, but where consumer debt exists, message templates and contact frequency rules are set by your compliance team, not the model.

Export control and CUI handling

For defense suppliers, running the model on-prem lets the customer keep order and contract data inside the boundary they already control for ITAR or CUI obligations. Netray does not hold these registrations on your behalf.

Where Netray fits

ERPray

Natural-language questions, aging dashboards and approval-gated agents over live ERP ledger data, read-only by default, with the query shown.

DataRay

Useful when dispute notes, remittance PDFs and bank files live outside the ERP and need to be searched alongside it on-prem.

Custom build

For bespoke scoring models, EDI remittance parsing or integration with a specialist collections tool.

How an engagement runs

Phase 1 . 1 to 2 weeks

Discovery

-AR data inventory across ERP, email and bank files
-Current worklist and dunning process map
-Access and role design for the read-only user
-Success measures agreed with the controller

Phase 2 . 2 to 4 weeks

Prototype

-Ranked worklist on a historical snapshot
-Drafted follow-ups reviewed by collectors
-Query-shown answers for DSO and aging questions

Phase 3 . 4 to 6 weeks

Pilot

-Live read-only deployment for one collector group
-Approval queue for drafts and match suggestions
-Weekly review of accepted versus edited suggestions
-Audit log review with internal audit

Phase 4 . Ongoing

Scale

-Rollout to additional entities and currencies
-Monitoring of suggestion quality
-Runbook and ownership handover

Questions to ask any vendor, including us

A short list that separates real NetSuite AI work from a chatbot demo.

Does the assistant ever write to the ledger, and if so, through which approved ERP workflow?
Where does the model run, and does any customer balance or contact data reach a third party?
How are prioritization scores computed, and can my team inspect and change the weights?
How does retrieval respect ERP roles and entity-level segregation?
What is logged for each draft, match suggestion and approval?
How are multi-currency, tax and credit memo effects handled in explanations?
What happens when the model is unsure of a match: does it say so or guess?
What does my own ERP vendor's collections feature already do, and where does this add something?

Frequently asked questions

Can AI post cash application or release credit holds on its own?

It should not. A safe design proposes matches and release requests, shows the evidence, and waits for a named approver. Posting stays in the ERP's own workflow so segregation of duties and audit trails are unchanged. Auto-posting is a policy decision for your controller, not a default setting.

Is my ERP's built-in collections module enough?

Often yes for rule-based dunning and standard aging. Add an AI layer when the limiting factor is context: dispute notes, delivery evidence, customer-specific payment patterns, and ad hoc questions like why DSO moved. If your collectors rarely override the dunning schedule, you probably do not need it.

How accurate are payment-pattern predictions?

They are only as good as your history. Customers with regular payment runs are predictable; one-off buyers are not. Expect useful ranking rather than precise dates, and measure suggestion acceptance by collectors during the pilot before relying on any score.

Can it run without sending customer data to a public cloud?

Yes. Open-weight models served with vLLM or Ollama can run on your own GPUs or in a private cloud, with retrieval over your ERP and documents inside the same boundary. That is the default option for defense and aerospace suppliers.

Which ERPs does this work with?

Any ERP with a supported read interface: NetSuite, SAP S/4HANA and ECC, Infor SyteLine and LN, Dynamics 365 and Business Central, Epicor, and others. The ledger objects differ, but the pattern of read-only retrieval plus approval-gated actions is the same.

Will it write emails that sound like us?

Drafts can follow your templates and tone guidance, and collectors edit before sending. Treat the first weeks as calibration: track which drafts are sent unchanged and which are rewritten, then tighten templates accordingly.

What about customer-facing portals?

Keep the first phase internal. Exposing an assistant to customers raises authentication, data-scoping and liability questions that are better tackled once the internal worklist is trusted.

Related guides

AP automation + on-prem AI

AI for accounts payable invoice matching in your ERP

On-prem AI reads vendor invoices, runs 3-way match against your ERP PO and receipt, and routes only real exceptions to your AP team. No invoice data leaves your network.

CFO + finance close + on-prem AI

AI for the ERP Month-End Close: Variance Commentary Without the All-Nighter

AI for the ERP month-end close: draft variance commentary, flag journal entry anomalies, and speed audit support, grounded in the general ledger, reviewed by finance before anything posts.

Ask your ERP anything

Natural Language Query for ERP Data: Ask SAP, Infor, or Oracle a Question in Plain English

See how natural language query over SAP, Infor, Oracle, and NetSuite data works: grounded text-to-SQL, role-based permissions, and a visible audit trail.

RFQ-to-quote + on-prem AI

AI for quote-to-cash in manufacturing and job shop ERPs

AI reads incoming RFQs, drafts quotes from your ERP's own cost and pricing history, and flags parts your shop has never quoted before, keeping the estimator in control.

RAG + SQL + permissions

A Private LLM Grounded on Your ERP Data

How a private LLM answers questions on your ERP data: RAG plus text-to-SQL, role-based permissions inherited from the ERP, and where each fits.

Agents + approval gates

AI Agents for ERP, Running On-Prem

A practical guide to on-prem AI agents for ERP: what they can safely automate, where human approval belongs, and how to design the guardrails.

Plan it with numbers

Free Tool

Talk it through with an engineer who knows NetSuite

Bring one real question your team cannot answer from the ERP today. We will map the data path, the model, and where it runs, and tell you honestly if AI is the wrong tool for it.

Start the conversation See ERPray

All AI for ERP guides