What a language model does with documents
A language model works with text the way rules work with numbers. A procurement table says how much was bought and from whom, and the documents say what exactly was bought and on what terms. The model finds the right places in a document, moves them into a table and compares the wording.
The main operations of the model in audit work:
1. Extract party details, sums, deadlines, volumes and requirements from a contract, a tender specification or an act.
2. Compare the documents of one procurement with each other and with the documents of other purchases of the same goods.
3. Find passages that look like a risk sign and give a page reference and a quote.
4. Prepare a draft note that the auditor edits and signs.
The list of 73 red flags from the Open Contracting Partnership (OCP) shows where this is needed. Many flags live in documents, not in numbers: unjustified technical requirements, physical similarity of documents from different bidders, a complaint about a tender, and a gap between the work done and the contract terms.
The page on how AI finds violations describes how documents fit into the common flow of signals together with tables and registers.
A specification written for one supplier
A specification written for one supplier shows itself through requirements that narrow the circle of bidders with no link to the task: an exact brand without the words “or equivalent”, a rare mix of features, experience requirements that no other market player has. OCP counts unjustified technical and qualification requirements as signs of bid rigging.
What the model does: it compares the specification with other buyers’ specifications for the same goods, marks requirements that others do not have, and checks whether the description matches the product card of one specific model.
A sample signal: “Buyer V, specification for laptops: some requirements appear only in the description of one model and are absent from other purchases of the same goods; links to the specification items and to the model description are attached”.
The second sign is found in bids. If documents from different bidders match in layout, typos and structure, OCP calls this physical similarity of documents and counts it as a sign of collusion between bidders. The model compares bids in pairs and shows the matching passages.
A requirement for one supplier can be lawful: new equipment must work with what was already bought. So a signal holds both the sign and a place for the buyer’s justification, and a person makes the assessment.
Matching the contract, the act and the payment
Matching three documents answers a simple question: did the buyer pay for what it ordered and received? The model compares the subject, quantity, price and deadlines in the contract, the acceptance act and the payment order.
The OCP list has risk signs for this check: payments under a contract exceed its sum, contract changes raise the price, and the work done differs from the contract terms.
The match also finds simple gaps: an act signed before the work started, other recipient details in the payment order, a volume in the act larger than in the contract. For a person, this means hours of leafing through folders; for a model, one operation per set of documents.
The match is especially useful where documents are many and sums are small. A person checks dozens of small acts under one contract by sampling, while the model checks them all and shows only the gaps.
A match gives a signal, not a conclusion. A gap in dates can be a typo, and an extra volume may be covered by an amendment that has not yet reached the system. The model shows the gap and both documents, and the auditor finds the reason.
The page on AI for control and audit tells more about such projects in audit work.
Complaints and appeals as a source of signals
Complaints from bidders and appeals from citizens describe a problem in their own words, often before it shows in the numbers. OCP includes a complaint about a tender in its list of red flags: the complaint itself draws more attention to the procedure.
The model helps sort the flow of appeals. It links each one to a purchase or a contract, pulls out the core of the claim, finds repeats and connects appeals from different people about the same thing.
Appeals are unstructured data, and audit bodies see the main benefit of generative AI in handling exactly this: 45 of the 59 OECD survey participants put work efficiency and analysis of unstructured data first.
A signal from an appeal is checked the same way as a signal from a table. The model shows which purchases are mentioned, which documents exist for them and whether the claim matches the data, and a person draws the conclusion about a violation.
Citizens’ appeals contain personal data, so the model for them runs on the organization’s own servers and gets access only to its task. The page on 243-FZ requirements and security covers the rules for such systems.
Experience of the Accounts Chamber and the OECD
The Accounts Chamber of Russia announced the launch of a state digital audit system in 2026, where AI serves risk-based audit. According to its deputy chair, the system that reads acts and payment documents already saves tens of thousands of working hours.
In March 2024, the OECD published a report on generative AI for anti-corruption bodies, supreme audit institutions and internal audit. 59 organizations from 39 countries took part in the survey.
The picture is calm. About half of the participants (30 organizations) do not use generative AI yet and only study use cases, and another 24% (14) are at the development stage. Supreme audit institutions have gone furthest: 47% (8) of them are at least at the development stage.
On tasks: 37% of participants (22) put evidence gathering and document review first. In Brazil, language models search large data sets for signs of fraud. In Finland, France, Greece and the UK, they prepare drafts, sort tables and write summaries.
The main obstacles named by participants are a lack of skills and IT limits, and data protection comes up almost as often. The OECD advises building models into low-risk processes first, and notes that the survey results apply only to the organizations that answered.
Where the model makes mistakes
The OECD names the main risks of language models in audit: hallucinations, opaque conclusions and bias. For an auditor, each looks concrete, and each has a defense.
1. Hallucination: the model writes a contract clause that is not in the document. Defense: every sentence of the note carries a page reference and a quote.
2. Opacity: the model marked a specification but did not explain why. Defense: the signal holds the sign that triggered it and a passage of text.
3. Bias: the model marks documents of one type or one region more often, because the training examples were like that. Defense: regular comparison of signals with a random sample.
A setup error happens too. OCP warns that if a flag fires on 90% of procedures, it most likely has many false positives, and the indicator needs fixing.
Another source of errors lies in the data, not the model. If the system has no amendment or an act scan cannot be read, the model will see a gap where there is none. So the signal card states which documents the model saw and which it did not find.
A red flag means a risk, not proof of a violation. The model prepares the material, and the auditor, the investigator or the court draws the conclusion about a violation.