AI in Public Procurement: A Guide for Bidders and Contracting Authorities

Using AI in public procurement means applying large language model and machine learning tools across the procurement process: monitoring and searching for tenders, analysing requests for tenders, drafting bids, conducting market research and preparing procurement documents. Finland's public sector buys goods and services for roughly EUR 38 billion a year, so even small efficiency gains are significant in euro terms — research estimates suggest that 50–80% of the manual work related to procurement could be automated.

AI concerns both sides of procurement. For a bidder, it is above all a way to find relevant tenders in time, to digest large tender documents quickly and to produce better bids with less effort. For a contracting authority, AI supports market research, improving the language and structure of tender documents, and information retrieval. The Finnish Act on Public Procurement and Concession Contracts (1397/2016, the Procurement Act) is technology-neutral: it neither prohibits nor specifically regulates the use of AI on either side.

AI is already part of the Finnish procurement landscape. The national procurement portal Hilma runs an AI assistant called Hankintavälkky that answers questions about public procurement, and Hansel's Tutkihankintoja.fi service uses AI to open up procurement data. The Ministry of Finance's Procurement Finland programme has made AI a key development theme. On the supplier side, specialised tools such as Haavi use AI to review tens of thousands of procurement notices every week.

This guide covers the use of AI first from the bidder's and then from the contracting authority's perspective. Finally, it reviews the regulatory framework: what the Procurement Act, the Act on the Openness of Government Activities and the EU AI Act require, and how general-purpose AI tools differ from procurement-specific ones.

The procurement process runs from planning and market research to the contract notice, bid preparation, comparison and the contract period. AI is useful at almost every stage, but in different ways for different parties. The table below summarises typical use cases.

The biggest gains come in stages that involve large volumes of text: screening contract notices, analysing tender documents and drafting bid responses. Decision-making, by contrast, stays with humans — the go/no-go decision and pricing on the bidder's side, and the comparison and award decision on the contracting authority's side, the latter also for legal reasons.

Note that the benefits of AI compound when the stages are connected. When the same system knows the company's capabilities, past bids and references, it can both pick out suitable tenders and lay the groundwork for the bid — with disconnected tools, the same information has to be re-entered every time.

The bidder's first and often biggest AI benefit is finding tenders. Around 19,000–20,000 procurement notices are published in Hilma every year, with further opportunities in the TED portal, municipal small-value procurement channels and specialised portals. Traditional keyword search based on CPV codes easily misses relevant notices, because contracting authorities classify and name procurements inconsistently. AI-based semantic search understands what a company actually sells and identifies suitable tenders regardless of wording.

The second major use case is RFP analysis. The tender documents of a public procurement can run to hundreds of pages: the request for tenders, the draft contract, service descriptions and annexes. AI extracts the suitability requirements, mandatory minimum requirements, award criteria with weightings, contract risks and deadlines in minutes. This speeds up the go/no-go decision — the assessment of whether it is worth participating at all.

In bid writing, AI works best as a drafter and checker. It helps structure the response according to the RFP, shape the text towards the award criteria and verify that every requirement has been addressed. The experience, evidence and winning content must still come from the company itself: AI does not know the customer, the reference deliveries or the competitive situation unless it is given them.

The fourth use case is managing bid assets. Keeping ESPD responses, references, certificates and standard answers in a structured knowledge bank means AI can pull the right ones into each bid. In Lassila & Tikanoja's experience, procurement-specific AI saves about 20% of the working time spent on bids.

A bid may be drafted with the help of AI. The Procurement Act contains no provision prohibiting the use of AI in bid preparation or requiring it to be disclosed, and a contracting authority cannot, as a rule, reject a bid merely because of how it was drafted. Tenders are assessed on their content: whether the bid meets the requirements of the request for tenders and how it scores against the pre-announced award criteria.

Responsibility, however, rests entirely with the bidder. A bid binds its submitter regardless of whether the text was written by a human or a language model. Language models can produce convincing but false content — invented references, wrong figures or non-existent certificates. This is not merely a quality risk: under Section 81(1)(10) of the Procurement Act, a contracting authority may exclude a bidder guilty of providing materially false information, for example in details concerning suitability, references or certificates. Every AI-generated claim must therefore be verified before the bid is submitted.

The second key rule concerns the data fed into AI services. Published tender documents are generally public, but a bidder's own material — pricing logic, customer information, technical solutions — often contains trade secrets. Free consumer AI services may use inputs to train models, which means the information can end up with the service provider. Business-grade services where inputs are not used for training and data processing is contractually agreed are the right choice for confidential material. For comparison: trade secrets in tenders are also protected within the procurement procedure, as they are confidential under Section 24(1)(20) of the Act on the Openness of Government Activities (621/1999), and even a competing bidder with party status has no right of access to them — though the total price used in the comparison must always be disclosed (Section 11(2)(6)).

In pricing, AI can be used for background analysis such as researching realised procurement prices. Bear in mind, however, that if a bid appears abnormally low, the contracting authority must under Section 96 of the Procurement Act require the bidder to explain its prices or costs. The bidder must be able to justify its price itself — 'the AI calculated it this way' is not an explanation.

For a contracting authority, the lowest-threshold benefits of AI are in preparation. In market research, AI helps gather information on the supplier field, comparable procurements and price levels. In drafting the request for tenders, it acts as a language and structure checker: a clear, unambiguous and consistent RFP reduces bidder questions, improves the comparability of tenders and lowers the risk of Market Court proceedings. AI can also be used to 'test-read' the RFP through a bidder's eyes and spot unclear requirements before publication.

In information retrieval, the public sector is already moving. Hilma's Hankintavälkky answers procurement questions from authorities and bidders alike, and Hansel's Tutkihankintoja.fi opens up central government purchasing data. The Ministry of Finance has issued guidance on the use of generative AI in public administration, and many municipalities and agencies have drawn up their own AI policies. The basic principle is the same as on the bidder side: confidential material or personal data must not be fed into services whose data processing has not been agreed.

When a contracting authority procures an AI system for itself, it is an ordinary public procurement in which the EU AI Act's requirements for the system must additionally be taken into account. The Association of Finnish Municipalities (Kuntaliitto) advises identifying the municipality's role (deployer or developer) and the system's risk classification, and using model contractual clauses drafted for AI procurement. Market dialogue is particularly important because the market for AI systems is developing rapidly.

In internal work, AI speeds up the drafting of memos, decision proposals and reply drafts. The line runs at decision-making: the award decision and its reasoning must always be made by a human and remain the contracting authority's responsibility, as discussed in the next section.

In tender evaluation, the use of AI is bounded by the basic principles of the Procurement Act. Under Section 3, the contracting authority must treat participants and other suppliers equally and without discrimination, and act transparently and with due regard to proportionality. The comparison must be made on the pre-announced criteria: under Section 93, the tender to be selected is the most economically advantageous — the cheapest in price, the lowest in costs, or the best in price–quality ratio — and the award criteria and their weightings must be stated in the procurement documents.

It follows that AI may assist in processing tenders but cannot replace the contracting authority's judgement. Suitable AI tasks include extracting data from tenders as the basis of a comparison table, preliminary checking of tenders against the RFP requirements, and summarising long tenders to support the evaluators. Qualitative scoring and the award decision with its reasoning belong to humans: the authority must be able to justify every point and show that the comparison was made on the announced criteria — 'the model scored it this way' is not acceptable reasoning in an award decision or before the Market Court.

The known weaknesses of language models also make caution practical. Models can misread tables, weight tenders inconsistently or favour a particular writing style over substance. If AI-assisted processing leads to an erroneous or discriminatory comparison, the procedure violates the Procurement Act and the bidder can appeal to the Market Court. Contracting authorities should document where AI was used and how its outputs were verified.

A good rule of thumb: AI may do the groundwork, humans do the evaluation. This also matches the spirit of the EU AI Act, which emphasises human oversight in public-sector AI use, and the requirements of general administrative law on careful and impartial handling of matters.

The EU AI Act (Regulation (EU) 2024/1689) is the world's first comprehensive AI law. It entered into force on 1 August 2024 and applies in stages. The Act is risk-based: prohibited practices (such as social scoring) are banned outright, high-risk systems face strict requirements on risk management, documentation and human oversight, and other systems mainly face transparency obligations.

For everyday procurement work, the essential point is that ordinary AI use — searching for tenders, analysing documents, drafting text — does not fall into the Act's high-risk categories. High-risk use cases include, among others, certain systems governing access to essential public services, not tools for drafting or analysing tenders. General-purpose AI models (GPAI) have their own transparency obligations, which fall on the model providers such as OpenAI and Anthropic, not on the companies or authorities using them.

The application timeline was amended in spring 2026 as part of the EU's digital regulation simplification package: the start of the high-risk system obligations was postponed. The table below summarises the key dates. Alongside the AI Act, procurement AI use must comply with the General Data Protection Regulation (processing personal data in AI services) and the Act on the Openness of Government Activities (handling official documents and trade secrets).

The overall regulatory picture is ultimately clear for both bidders and contracting authorities: using AI in procurement is permitted, and the regulation concerns how it is used — not whether it may be used. Duties of care, trade secret protection and the principles of the Procurement Act form the boundaries within which AI can be used to full advantage.

General-purpose AI services such as ChatGPT, Claude and Microsoft Copilot are good generalist tools: they draft, summarise and polish language. In procurement work, their limits appear quickly. They do not monitor procurement channels, so finding tenders remains manual work. They do not know the company's bid history or references unless everything is fed in again each time. And because they always produce an answer, the hallucination risk is real in a field that demands precision.

Procurement-specific AI is built around this process. Haavi, for example, reviews over 60,000 procurements a week — sourced from Hilma, Tarjouspalvelu, TED, small-value procurement channels and NATO portals — and uses search profiles to identify suitable tenders for a company the moment they are published. Its AI analysis brings the RFP's requirements, award criteria and risks into a single view, and bid drafting draws on the company's own reference bank. Data stays within the service and is not used to train models.

The choice is not either–or. Many bidders use a specialised tool for monitoring, analysis and bid foundations, and general-purpose AI for complementary ideation. What matters is that confidential material is processed only in services with agreed data processing, and that a human who is responsible for the outcome always reviews the AI's output.