# How is AI used in technical bid evaluation?

> Public sources describe AI in bid evaluation reading and sorting bid documents, flagging risky clauses and detecting irregular or collusive bidding. Published evidence on AI judging technical compliance is thin. US federal rules (48 CFR 15.305) still require an evaluation based only on the solicitation's stated factors and documented in the contract file, whatever tools are used.

- Label: Fact
- Type: Explainer
- Published: 2026-11-13T03:06:14+09:00
- By: Jun Zhong
- Researched and drafted by: Beacon's AI newsroom
- Checked by: Jun Zhong on 2026-11-13
- Canonical: https://beacon.winda.ai/library/ai-in-bid-evaluation/

## Why does it matter?

An AI system can review many bids in minutes, so a bias in its training data reaches every bid it scores, faster and more widely than one reviewer's error; the Organisation for Economic Co-operation and Development (OECD) makes this point in its 2025 review of AI in government [1]. Evaluations already draw legal challenges: in fiscal year 2025 the US Government Accountability Office (GAO) named unreasonable technical evaluation as the most common reason it sustained bid protests [5]. US federal rules require the strengths, deficiencies, weaknesses and risks behind each evaluation to be recorded in the contract file [4].

## Where in bid evaluation is AI used?

A 2023 mapping of AI solutions offered by public procurement platforms, cited by the OECD, found that 54% supported pre-tender planning, 31% the tendering stage, 4% the supply phase and 11% the whole cycle [1]. The documented uses closest to bid evaluation are these:

| Use | What the AI does | Example from public sources |
|---|---|---|
| Sorting documents | Assigns purchases to standard procurement classification codes | Ukraine's public e-procurement system uses machine learning for this [1] |
| Finding clauses | Extracts risk-bearing clauses from bidding documents | A 2021 university study on 25 engineering, procurement and construction (EPC) invitation-to-bid documents [2] |
| Screening for irregularities | Flags risks in bids, contracts and public notices for auditors | Brazil's Office of the Comptroller General: about 191,000 purchases analyzed and 203 audits triggered in 2023 [1] |
| Detecting collusion | Finds non-competitive patterns in historical bid data | Eleven machine learning algorithms tested on data from Brazil, Italy, Japan, Switzerland and the United States [3] |
| Text mining tenders | Finds wording that restricts competition | Hungary: 119,000 government tenders from 2011 to 2020 [1] |

The tools in the table support evaluators, auditors or bidders [1][2][3].

## What does the research show?

The 2021 EPC study, from Pohang University of Science and Technology, built two text-mining modules and converted 25 bidding documents from offshore, onshore plant and infrastructure projects into a database of 21,683 rows [2]. In its pilot test, the modules found risk clauses with about 92% and 88% accuracy, against about 70% and 86% for engineers reading manually, and finished the review faster [2]. The tool was built for contractors reviewing bidding documents, and was developed and tested by the same research team on 25 documents.

For collusion, a 2022 study in *Automation in Construction* reports that machine learning needs many historical auctions to calibrate but gives reasonable detection rates even with little information; the OECD reports accuracy of 81% to 95% for this work [1][3]. A Hungarian study found that adding tender text raised the accuracy of a corruption-risk model from 77% to 82% [1].

The OECD itself is cautious. It finds that research specifically on AI for public procurement is scarce and that most procurement bodies are at an early stage of AI maturity, and it quotes the Ada Lovelace Institute's finding of "a surprising lack of evidence on the effectiveness and impact of AI tools" in the public sector [1].

## What are the risks and limits?

The OECD lists these risks for AI in public procurement and government use [1]:

- **Skewed data:** a system trained on unrepresentative data may favour some bidders, and it repeats that at speed and scale. [1]
- **Low transparency:** if the buyer cannot explain how the system works and what data trained it, it cannot monitor the system on its own. [1]
- **Legal exposure:** unclear rules on AI use invite challenges from losing bidders. [1]
- **Poor data:** complete procurement data is still largely unavailable as open data, and systems are often incompatible. [1]
- **Over-reliance:** people tend to trust algorithmic outputs and may accept incorrect ones, and generative AI can state made-up facts convincingly. [1]
- **Lock-in:** skills gaps and dependence on one supplier's technology and data. [1]

## What rules apply to AI-assisted evaluation?

The rules do not change because a tool is used. In US federal procurement, the source selection authority must make sure proposals are evaluated only on the solicitation's factors, and the reasons behind ratings must be documented [4]. The FAR is cited as codified on 22 October 2026; many agencies apply a rewritten Part 15 (September 2025) as a deviation [8].

For US federal agencies, Office of Management and Budget (OMB) memorandum M-25-21 of 3 April 2025 defines high-impact AI as AI whose output is a principal basis for decisions with legal, material or significant effects on areas such as rights, safety and critical infrastructure, and presumes high impact for AI used in the design, construction or testing of equipment that would pose a significant safety risk if it failed [6]. High-impact uses need pre-deployment testing, an impact assessment, ongoing monitoring, trained operators, human oversight and a way to appeal [6]. The memo does not list bid evaluation by name.

The National Institute of Standards and Technology (NIST) AI Risk Management Framework (January 2023) lists the qualities of trustworthy AI, among them validity and reliability, safety, security and resilience, accountability and transparency, explainability, privacy, and fairness with harmful bias managed [7]. Private oil and gas companies set their own policies.

## Where integrity problems show up

| Decision or step at this stage | What it affects later | In the Library |
|---|---|---|
| Material grade and heat treatment are checked against the specification, by people or with AI support | A missed deviation reaches service: at a refinery in Texas City, Texas, in 2005, a carbon steel elbow fitted where alloy steel was required failed after fewer than 3,000 hours [9]; amine and wet hydrogen sulfide cracking is found at or near welds, less often where welds were heat treated [10] | What is positive material identification (PMI)?, What is post-weld heat treatment (PWHT)? |
| The list of vendor documents is extracted from the bid | Process safety information must include materials of construction and design codes [11] | What is vendor document management? |
| AI flags or summaries are accepted without checking the source text | Over-reliance leads users to accept incorrect outputs, and errors can spread through whole systems [1] | How should engineers validate AI outputs?, What is human-in-the-loop AI for safety-critical work? |

Sources: CSB Safety Bulletin 2005-04-B, 2006 [9]; OSHA Technical Manual, Section IV, Chapter 3 [10]; 29 CFR 1910.119(d) [11]; OECD, 2025 [1].

## Sources

1. [Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions (chapters on AI in public procurement and on risks), Organisation for Economic Co-operation and Development (OECD)](https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/06/governing-with-artificial-intelligence_398fa287/795de142-en.pdf)
2. [AI and Text-Mining Applications for Analyzing Contractor's Risk in Invitation to Bid (ITB) and Contracts for Engineering Procurement and Construction (EPC) Projects (S.J. Choi, S.W. Choi, J.H. Kim, E.-B. Lee), Energies 14(15), 4632, MDPI (peer-reviewed journal); authors at Pohang University of Science and Technology (POSTECH)](https://doi.org/10.3390/en14154632), 2021-07-30
3. [Collusion detection in public procurement auctions with machine learning algorithms (M.J. García Rodríguez, V. Rodríguez-Montequín, P. Ballesteros-Pérez, P.E.D. Love, R. Signor), Automation in Construction 133, 104047 (abstract), Elsevier (peer-reviewed journal)](https://doi.org/10.1016/j.autcon.2021.104047)
4. [48 CFR Part 15, Contracting by Negotiation (FAR Part 15), sections 15.303 and 15.305, eCFR, US Office of the Federal Register](https://www.ecfr.gov/current/title-48/chapter-1/subchapter-C/part-15)
5. [GAO Bid Protest Annual Report to Congress for Fiscal Year 2025 (B-158766, GAO-26-900695), US Government Accountability Office (GAO)](https://www.gao.gov/products/gao-26-900695), 2025-12-12
6. [M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, US Office of Management and Budget (OMB)](https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf), 2025-04-03
7. [NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), US National Institute of Standards and Technology (NIST)](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf)
8. [FAR Overhaul: FAR Part Deviation Guidance, Acquisition.gov, US General Services Administration](https://www.acquisition.gov/far-overhaul/far-part-deviation-guide)
9. [Safety Bulletin No. 2005-04-B, BP Texas City, TX Refinery Fire: Positive Material Verification: Prevent Errors During Alloy Steel Systems Maintenance, US Chemical Safety and Hazard Investigation Board (CSB)](https://www.csb.gov/assets/1/20/RHUBulletin.pdf), 2006-10-12
10. [OSHA Technical Manual, Section IV, Chapter 3: Pressure Vessel Guidelines, US Occupational Safety and Health Administration (OSHA)](https://www.osha.gov/otm/section-4-safety-hazards/chapter-3)
11. [29 CFR 1910.119, Process safety management of highly hazardous chemicals, eCFR, US Office of the Federal Register](https://www.ecfr.gov/current/title-29/subtitle-B/chapter-XVII/part-1910/subpart-H/section-1910.119)

Published by Winda Beacon (https://beacon.winda.ai).
