Home · Business · Finance & Accounting · Invoicing & credit control

PARTLY

As of 13 August 2026, AI can only partly approve supplier invoices.

This still needs a person who signs their name to it.

Can you do it?

15 minutesto a draft.

30 minutesto something you’d act on.

Cost, all in£0

Skill neededchat-fluent

Who has to check ita colleague

What the alternative costsA purpose-built invoice-processing tool such as Nanonets is an alternative to a general chatbot; the supplied tool data gives no price.

If this goes wrong, an incorrect or duplicate invoice can be paid and recovering the money may require supplier disputes, bank action and internal investigation.

What to actually do

  1. Hand it to a person

    The route this page recommends

    A person who owns the outcome does this end to end, worth it when the failure is dear.

  2. Use a tool built for this

    Second choice
  3. Do it yourself

    The distant third

    A chat interface, chat-fluent skill, and roughly 30 minutes until you can act on the result.

    How to actually do it

    1. Open the accounts payable or accounting system and record the invoice as pending rather than approved or paid.
    2. Gather the original invoice, matching purchase order, goods-received note or service confirmation, supplier record and the relevant approval-limit and cost-centre rules.
    3. Remove unnecessary personal or bank information, then paste the invoice and supporting records into the prompt in clearly labelled sections.
    4. Ask the model for a recommendation using only the supplied evidence, and require it to list every mismatch, missing document and duplicate-payment warning.
    5. Compare each stated amount, supplier, invoice number, VAT figure and delivery detail against the original documents and the accounting system, correcting any extraction errors.
    6. Send the recommendation and the source documents to the authorised budget holder or finance colleague, who makes the approval decision and records it in the accounting system.

    Prompt

    You are reviewing a supplier invoice for a UK business. Do not approve, reject, send, schedule or authorise any payment. Produce a recommendation only: APPROVE FOR HUMAN AUTHORISATION, HOLD FOR REVIEW, or REJECT, followed by a short reason.
    
    Use only the evidence supplied below. Do not invent missing facts, VAT treatment, purchase orders, delivery records, supplier details or approval rules. Check for:
    1. Invoice number, invoice date, supplier name, bank details, net amount, VAT amount, gross amount and currency.
    2. Agreement between the invoice and the purchase order, including supplier, quantities, prices, VAT and total.
    3. Evidence that the goods or services were received, if provided.
    4. Duplicate invoice numbers, duplicate amounts, changed bank details, unusual urgency, unexplained charges and missing evidence.
    5. Whether the stated approval limit and cost-centre rules are met, if those rules are supplied.
    
    Return this structure:
    - Recommendation:
    - Evidence supporting it:
    - Exceptions or mismatches:
    - Missing information:
    - Checks a human must complete:
    - Confidence: low, medium or high, with a reason.
    
    An invoice may be recommended for human authorisation only when all supplied evidence matches and no material exception is present. If evidence is missing or ambiguous, recommend HOLD FOR REVIEW. Flag any bank-detail change for independent verification through a known supplier contact, not through contact details shown only on the invoice.
    
    Invoice:
    [PASTE INVOICE TEXT OR EXTRACTED DATA]
    
    Purchase order:
    [PASTE PURCHASE ORDER OR WRITE NONE]
    
    Receipt or delivery evidence:
    [PASTE RECEIPT, GOODS-RECEIVED NOTE OR WRITE NONE]
    
    Supplier and approval rules:
    [PASTE RELEVANT RULES OR WRITE NONE]

    Open it prefilled in ChatGPT or Claude, or copy it into Gemini, which takes no prefill link.

What it gets wrong

  • AI cannot know whether goods or services were genuinely received unless you provide reliable delivery or service evidence.
  • AI cannot take responsibility for authorising a payment, even when its recommendation appears correct.
  • AI cannot resolve an unexplained price change, conflicted supplier relationship or unusual payment request without a human investigation.
  • AI cannot independently verify changed bank details through a trusted supplier contact.
  • AI can misread invoice layouts, VAT details or duplicate documents, so its extracted figures still need comparison with the originals.

What caps this at PARTLY: legal accountability, judgement under ambiguity and stakes of error.

How we scored this

Five axes, each scored nought to two by hand: ten means AI carries the task cleanly, and the thresholds that turn a total into YES, PARTLY or NO are published in the methodology. Each axis name links to its definition.

AxisScore (0–2)
Output1
Inputs1
Verification1
Liability0
Effort delta1
Total4 / 10

FAQ

Can ChatGPT approve supplier invoices?
It can compare an invoice with a purchase order and delivery evidence and recommend approval, holding or rejection. It should not be the final approver or send the payment, because the authorised person remains responsible.
Can AI match invoices to purchase orders?
Yes, AI can extract invoice fields and compare suppliers, quantities, prices and totals with a purchase order. It cannot reliably resolve missing receipts, unusual charges or disputed prices without a human decision.
Is it safe to let AI approve invoices automatically?
Automatic approval is only suitable for tightly controlled, low-risk cases with reliable system rules and exception handling. Bank-detail changes, duplicates, mismatches and missing delivery evidence should be held for independent human checks.
What information does AI need to check an invoice?
Give it the invoice, purchase order, goods-received note or service confirmation, supplier record and applicable approval rules. It also needs enough context to identify duplicates, changed bank details and unexplained differences.

Nearby answers

Assessed by gpt-5.6-luna (gpt-5.6-luna) on 2026-08-13, second-checked by an independent model. Wrong somewhere? Email [email protected] and it gets re-checked.

The newsletter

AI news, new answers and product picks, straight to your inbox.