Check delivery notes, orders and invoices with AI before recording them

The DDT check, orders and invoices with AI prepares a comparison between the documents of the same supply and makes visible the differences to be verified. Quantity ordered, delivered and invoiced must remain separate; prices, missing references and partial deliveries require explicit rules. The result is a queue of anomalies, not an accounting certification or a payment authorization.

Reading an invoice well is not enough to know if it corresponds to the supply. You need to link it to the relevant order and delivery documents. A difference can indicate an error, an incomplete delivery or an agreed condition: the flow must bring the evidence to the referent instead of automatically deciding which explanation is true.

Technical sources are verified at 9 October 2026 and the examples are hypothetical. The guide to invoice automation describes the general cycle; here we address the documentary comparison to be included in the framework of AI for businesses.

What documents and references are used for comparison?

Collect the relevant order, the available DDTs and the invoice to check, keeping references and versions. The order may have a revision; the delivery may be split; the invoice may include multiple documents. The project must describe these relationships before comparing the totals. If a document is missing, the status of the comparison must show the absence.

Use identifiers and references in sources. Supplier name and similar amount may suggest a candidate, but they are not enough for a certain association. A comparison table must distinguish verified link, candidate to be reviewed and absent reference. The model can help gather evidence, without hiding the association passage.

n8n documents the extraction of structured information with attributes or schema. It is a primitive for preparing document data; it does not prove that the relationship between order, DDT and invoice is correct in your system. Official Information Extractor.

When you have data already structured in management software, use that path where appropriate. Don't pass every value through a generative reading just because the project contains AI. The document part can serve for unstructured notes or formats, while calculations and comparisons follow verifiable rules. Separating roles makes it easier to find the cause of an anomaly.

How to handle partial deliveries and more DDT?

A partial delivery may be expected. It shows quantity of the order, quantity of the connected deliveries and residue, distinguishing the given from its interpretation. The presence of a residue does not prove a supplier error: it may require a check on the state of supply. The contact person must see the documents included in the calculation and any missing ones.

A purely demonstrative example: the order requires ten units and a DDT reports four. The comparison shows six units not represented by that document, without concluding that the supply is in default. If a second DDT arrives, the stream must link it correctly and update the comparison. The numbers illustrate the structure, not a tolerance or a real result.

Avoid adding twice the same document received through different channels. A forwarding or a copy is not a new delivery. Keep a stable reference and keep the reason why the document enters the group. If the identity is not rebuildable, the case remains under review. The Guide to the warehouse and inventory with AI expands the theme of downstream operating states.

The invoice may cover part of the supply or group deliveries. The comparison must follow the agreed process, without imposing the idea that each order has only one DDT and one invoice. This is a choice of data model, to be verified with those who know the management software. If the relationship is ambiguous, the system must ask for verification instead of choosing the most plausible group.

How to compare items, prices and units?

Keep the original values before normalizing them. A different code can be an approved alias or a different article. A unit price may refer to packaging, piece or other unit. Conversion must have a source and a rule; a similar description does not justify implicit equivalence. The useful comparison makes visible which passage led to the normalized data.

The price should be compared with the relevant condition, keeping separate additional items and discounts. If the document confirming an agreement is missing, indicate the limit. Do not infer that an equal total means all the correct lines: differences on different articles can compensate. The referent must be able to control the entries, not just the aggregate result.

The Guide to AI for purchases and suppliers distinguishes comparison of offers and control of conditions over time. To make comparable offers before purchase read compare supplier quotes with AI. Here the need concerns an already documented supply and the differences that emerge between its elements.

If the data comes from a scan, keep uncertain fields visible. A misread character in a code can change the link. An incomplete amount must not become zero. Missing data, unreadable data and an actually null value must have different statuses. This lets the person responsible understand whether a better source or a decision about the anomaly is needed.

How to define tolerances that do not hide errors?

Tolerance must be an explicit business rule, related to field and context. The model does not decide how much deviation is acceptable. A monetary rounding rule does not automatically apply to quantities. A condition allowed for a family of products must not become the rule of every supply.

Write the reason, the field, and the intended action. The comparison may mark a difference as falling under the rule or require a verification. It must still maintain the original value and reference, so the auditor can reconstruct the path. “Accepted” should not mean that the difference has been erased from view.

Anomaly card to be sent to referent

It presents the linked documents, the affected line, the original values and the difference. It indicates the possible rule applied and what remains uncertain. Ask the manager if the case requires document integration, correction or other verification. The decision must remain separate from registration and payment.

If there is no approved tolerance, the first pilot can show all the relevant differences and collect decisions. Only after the team defines which rules are appropriate. Don't turn a choice made about a case into a universal authorization. The sample must help to know the exceptions, not hide them to improve an indicator.

How to organize the anomaly queue and the decision?

Each anomaly should have a referent, a state, and the next step. "Difference detected" is not a decision. The contact person can confirm a partial delivery, request a document or correct a correspondence. The system retains the relevant source and motivation in the project path, making the initial data distinguishable from the subsequent decision.

n8n documents human review for tool actions. A useful review must receive proposal and necessary data, allowing for a controllable outcome. It is a primitive design, not a guarantee that the accounting process is already covered. Official Human Review.

Do not link the comparison directly to a payment without a separate process. Registration also requires appropriate fields and permissions. For the required-responsible passage read approval of purchases with AI. The result of the comparison prepares evidence for those who decide, without assuming the qualification of a complete professional verification.

Two hypothetical scenarios for manufacturing and services

A textile manufacturer receives a supply in multiple deliveries

Imagine a Lombard manufacturer receiving materials in multiple deliveries. The comparison preserves linked units, codes, and documents. An invoice referring to a delivery is not compared with the entire order without making the perimeter visible. The referent sees quantities represented, residues and missing references.

A note on the variant of the material requires verification with the relevant agreement. The system does not declare a description equivalent just because it looks close. The person in charge controls the source and decides on the operational treatment. The scenario shows the type of useful evidence, without reporting economic results or measured accuracy.

A plant company controls the materials of a clerk

Imagine an Emilian company that has an order, two DDTs and an invoice with a transport item. The stream prepares the difference table and indicates that the additional entry is not reconciled with the available documents. It does not conclude that it is wrong: the referent must verify the agreed conditions.

If an updated document arrives, the comparison keeps the relevant version and reviews the anomaly. The upstream phase is described in import PDF orders into management software. Properly acquired data helps the next comparison, but does not prove on their own that a supply is complete or that the invoice is ready for all operations.

What changes when a document arrives after review?

The comparison has a version. If a new DDT or updated invoice arrives, the contact must see what data has been added and what decisions could be passed. Do not silently reuse a previous outcome on a changed group of documents. The system maintains the sources and prepares the new difference in the agreed perimeter.

A decision on a partial delivery may still be relevant, but the residue changes when other documents arrive. The flow must distinguish already considered quantities and new quantities, avoiding duplication of deliveries. A copy of the DDT may look like a new entry without representing a new event. The revision checks the references, not the order in which the files were uploaded.

If the updated source corrects an error, it retains the reason for the replacement in the intended path. Do not leave both versions in the calculation as if they were two cumulative elements. The referent must know which document was used and which remained historical. The project should not infer this role from the file date alone.

Review closed anomalies only for what the change is about. A price change does not require any control over quantities to be reopened, but it must make its impact visible. Conversely, an incomplete source can prevent the old outcome from being maintained. The route must help the person in charge to distinguish these conditions instead of assigning an all-correct generic state.

In the pilot prepares an updated case and the expected comparison. Check that the system shows the new version, does not add up copies and does not lose the referent's decision. It is a design test to be performed, not an already observed result. This control makes the discussion of the continuity of the process more concrete: the question is not just how many bills it reads, but how it keeps recognizable what has been verified when sources change. If the result is not rebuildable, the board remains open and the team knows which dependency to solve.

How to evaluate the pilot and decide the budget?

Prepare cases with partial deliveries, repeated documents, different units and absent references. A referent defines the expected comparison and controls links, lines, and differences. It measures found anomalies, missed anomalies, and false alarms. It also assesses how much work it takes to review each card. The number of documents drawn up is not a sufficient measure of quality.

If the system links the wrong group, improving the summary form is not enough. If the units are incomplete in the catalog, the problem is the source. If the queue stays still, the problem is responsibility. The pilot must distinguish these causes before increasing autonomy. We do not set here a universal threshold of deviation or accuracy: it depends on the consequences and rules of the process.

The cost includes integration, extraction where needed, comparison rules, review and maintenance. Agree a spending limit for the trial and the evidence needed to extend it. Avoid matching by resemblance, invented tolerances and totals that hide differences between individual lines. An anonymised set containing an order, a delivery note (DDT) and an invoice lets Giallo assess the automation of business processes with the contact person who can confirm the expected result.