This guide is for a CFO, controller, accounts payable lead or procurement manager at a mid-sized company who has to decide where an agent stops in the buying cycle. It applies the general approach of the AI automation for business operations guide to one named process, and it is the buy-side companion to our order-to-cash AI agents guide. The build work behind it sits in the agentic AI automation service.
In this guide
- Which procure-to-pay steps can an AI agent run today?
- The procure-to-pay agent-fit matrix
- How should invoice-match exceptions be routed?
- What controls keep an agent inside segregation of duties?
- A sequencing plan for a procure-to-pay agent
- Where AI fits
- What delivery record exists, and what does not?
- Which alternatives should a mid-sized company compare?
- What are the limits of this guide?
- Frequently asked questions
- Next step
Which procure-to-pay steps can an AI agent run today?
Procure-to-pay is a chain of decisions, not one task. IBM's description of the cycle runs from identifying a need through requisition and approval, sourcing, the purchase order, receiving, invoicing and matching, to payment. Each step has its own data, its own error cost and its own answer to the question of whether an agent may act alone. A wrong line-item code is cheap to fix; a payment released to the wrong bank account usually is not.
The useful split for a mid-sized company has three levels. An agent acts alone where the rule is clear and the action is reversible. An agent drafts and a person approves where the action commits money or changes a record other teams rely on. A person decides alone where the decision is about trust in a supplier or about who gets paid.
The procure-to-pay agent-fit matrix
| P2P step | What an agent can do alone | What needs a person | Why the line sits there |
|---|---|---|---|
| Requisition | Read a request, fill item, quantity and cost centre, check it against budget rules | Approve a request above the budget owner's limit or outside the catalogue | Budget authority belongs to a named person, not to a model's reading of a policy |
| Purchase order | Draft the PO from an approved requisition and an existing supplier's agreed terms | Approve a PO to a new supplier, on new terms or above a set value | A PO is a commitment to spend; new terms are a negotiation |
| Goods receipt | Propose receipt lines from a delivery note or warehouse scan | Confirm a partial, damaged or substituted delivery | The receipt is the evidence the later match depends on |
| Invoice capture | Extract header and line fields, flag missing PO numbers and likely duplicates | Review fields the checks could not confirm | Extraction errors flow straight into the match |
| Two- and three-way match | Match invoice, PO and receipt within the tolerance finance has set, and queue matched invoices | Resolve every mismatch outside tolerance | A tolerance is a finance policy; an agent applies it, it does not set it |
| Supplier master changes | Detect a changed bank account, address or tax ID on an invoice and stop processing | Verify the change through a separate, known contact and approve it | A bank-detail change is the classic route of payment fraud |
| Approval routing | Route each invoice to the approver the rules name, chase overdue approvals | Approve the invoice | The approver signs for the spend |
| Payment release | Propose a payment run from approved, matched invoices | Release the payment run | Money that leaves the account is hard to recover |
Clean purchase orders and a stable supplier base justify a wider agent-alone column; many non-PO invoices or weak receiving discipline justify a narrower one.
How should invoice-match exceptions be routed?
Most of the risk sits in the exceptions. An agent should never resolve a mismatch by guessing which document is right; it classifies the exception, attaches the evidence and sends it to the person who owns that problem.
-
Price differs from the PO beyond tolerance
- Agent's job
- Show PO price, invoice price and the difference
- Owner who resolves it
- Buyer or procurement owner of that supplier
-
Quantity invoiced exceeds quantity received
- Agent's job
- Show receipt lines and delivery date
- Owner who resolves it
- Receiving or warehouse lead
-
No goods receipt recorded
- Agent's job
- Ask the requester to confirm receipt
- Owner who resolves it
- Requester, then accounts payable
-
No PO number on the invoice
- Agent's job
- Suggest likely POs from the same supplier
- Owner who resolves it
- Accounts payable
-
Likely duplicate invoice
- Agent's job
- Block it and show the earlier invoice
- Owner who resolves it
- Accounts payable
-
Unknown supplier or changed bank details
- Agent's job
- Stop all processing for that supplier
- Owner who resolves it
- Finance controller, through out-of-band verification
The last row deserves its own control. The FBI's public service announcement on business email compromise describes fraudsters timing a request to move payment from one bank account to another under their control, and recommends verifying account changes through a secondary channel. An agent that reads supplier emails must treat any such request as a stop signal, never as data to update.
What controls keep an agent inside segregation of duties?
An agent is a new actor in the approval chain, so it needs its own identity and its own limits, as our guide to AI agent security and human approval sets out. The OWASP guidance on excessive agency names the core risk: an AI system given more functionality, permissions or autonomy than its task needs. In procure-to-pay that becomes a short checklist:
- The agent's account can create drafts and propose matches; it cannot approve a PO, approve an invoice or release a payment.
- No single flow lets the same identity raise a PO, confirm the receipt and approve the invoice.
- The agent cannot edit the supplier master; it can only flag a change for a person.
- Tolerances, approval limits and routing rules live in configuration that finance owns, not in the prompt.
- Every proposal, match and routing decision is logged with the documents it used.
A sequencing plan for a procure-to-pay agent
-
Start with invoice capture and duplicate checks
Extraction with field validation is the step with the clearest right answer; our guide to extracting fields from invoices with vision-language models covers the validation rules.
-
Add matching inside a tolerance finance sets
Matched invoices queue for approval; every mismatch goes to the exception table above.
-
Add approval routing and reminders
The agent routes and chases; the approver still approves.
-
Add PO drafting for existing suppliers on agreed terms
New suppliers and new terms stay with procurement.
-
Leave supplier changes and payment release with people
until the earlier steps have produced a track record, then revisit the matrix.
-
Measure before and after each step
, using the baseline method in our guide to measuring AI automation ROI after launch.
Where AI fits
Netbase works with the major commercial and open-source AI tools and models, chosen per project. Each item below states how mature it is at Netbase.
- In delivered work
Product recommendation engine
Built for 4over4's online printing store from browsing and purchase history; not a procure-to-pay tool, but the named delivered-AI proof this cluster cites.
- In delivered work
Document AI platform
Field extraction and classification delivered for a client that is not named; the same extract, validate and review pattern applies to supplier invoices.
- Available capability
Procure-to-pay agents with approval gates
Matching, exception routing and PO drafting as one agent workflow is a capability Netbase is growing; it is not yet tied to a published procure-to-pay case.
What delivery record exists, and what does not?
- RevenueFlow Assistant. Netbase provides this Chrome extension, which reads payment emails with read-only access, processes them locally, flags duplicates and writes only user-approved rows to Google Sheets. It handles incoming payments, not supplier invoices, but ships the rule this guide recommends: the tool proposes, a person approves, a duplicate stops the write.
- Cloud ERP for a US client (not named). Netbase works as offshore development partner on a multi-tenant cloud ERP delivered as SaaS to small and mid-sized businesses; its planned second phase covers purchasing and vendor management, financial management and inventory.
- ERP consultancy. Odoo consultancy and customisation, e-commerce-to-ERP integrations that sync orders and inventory by API, and ERP modules for finance and accounting and inventory. None of these engagements is named, and none publishes a figure.
What does not exist. No published Netbase record describes a full procure-to-pay agent running from requisition to payment, and no accuracy, touchless rate, time-saved or cost figure is published for any procure-to-pay work. The matrix and tables above are a delivery view built from these engagements and the cited guidance, not a measured result.
Which alternatives should a mid-sized company compare?
| Route | Strength | Choose it when |
|---|---|---|
| Your ERP's built-in matching and workflow rules | Already connected to your POs and ledger | Most invoices are PO-backed and the default rules match your approval policy |
| An accounts payable or procurement platform with AI features | Fast start, vendor-maintained extraction and matching | Your process is close to standard and you accept the platform's default agent line |
| A custom agent layer over your ERP with your own gates | The matrix, tolerances and routing are yours | Your suppliers, documents or systems are specific enough that a default over- or under-automates a step |
When NOT to add an agent: if invoice volume is low enough that one person clears it weekly, if most spend has no purchase order, or if the supplier master is out of date. Fix PO discipline and supplier data first; an agent will only automate the mess.
In Netbase custom development most projects run on fixed-price contracts agreed after discovery. Approvals, audit logs and access control for agents are delivered through the enterprise AI agent platform, relevant to professional services firms and any mid-sized company buying on purchase orders.
What are the limits of this guide?
- The matrix is a starting position from delivery practice and public security guidance, not an audit standard; your auditors and your own control framework set the final line.
- No touchless rate, accuracy, duration or cost figure is claimed; real numbers depend on your data quality and volume.
Plan the next step with a Netbase consultant
Frequently asked questions
It can queue invoices that match the purchase order and receipt within the tolerance finance has set, but the approval itself should stay with the person who owns the spend, so the agent never signs for money it helped prepare.
The agent should classify the mismatch, attach the purchase order, receipt and invoice, and route it to the owner of that exception type, such as the buyer for a price difference or the warehouse lead for a quantity difference.
No. A changed bank account on an invoice or email should stop processing for that supplier until a person verifies the change through a separate, known contact.
Start with invoice capture, field validation and duplicate checks, then add matching within tolerance, because those steps have clear right answers and leave approvals and payments with people.
Next step
Tell us which procure-to-pay steps you run today and where you want the approval line, and we will book a solution review to scope the agent boundary. You can also see the agentic AI automation service.
Related services and solutions
AI automation and agents that keep people in charge Netbase provides AI automation and agent development for operations teams that want repetitive, multi-step work done by software while people keep approval over decisions that matter. We combine rule-based automation with AI where it adds value, build in human approval, and measure return against a baseline.
Learn More
Enterprise AI agent platform: run AI agents with approvals, audit logs and cost limits An enterprise AI agent platform is a governed runtime where AI agents plan and carry out multi-step work through approved tools, while approvals, audit logs, access rules and cost limits keep every action accountable. Netbase offers it as a forward-looking solution, built inside your own environment rather than licensed as a product.
Learn More
Work with Netbase remotely
Every Netbase engagement is delivered remotely, so clients in any country and any industry can work with us; Netbase has worked with clients in 80 countries. Contracts and invoices are issued by Netbase JSC, and payment is made by bank transfer to Netbase JSC's company account or PayPal Business. Talk to the CEO, Huy Nguyen (David), on WhatsApp or by email.
Discuss a project
Netbase JSC helps organizations design, build, modernize, and operate digital products and AI-enabled business systems.project@netbasejsc.com
+84 937 869 689
91 Nguyen Chi Thanh, Dong Da, Hanoi, Vietnam
Get in touch
Tell us what you want to build, modernize, or operate.