AI Reviews Waste Assessments: Our Contribution to the IÖB Challenge

Companies in Austria use expert reports—known as “assessment certificates”—to determine whether a type of waste may be used as a substitute fuel. At the Austrian Ministry of the Environment, staff members have traditionally reviewed these waste assessment reports manually. A recent challenge sought to determine whether AI could assist in this process.

The ministry published it on the IÖB Innovation Platform. We submitted a solution, were invited to the Innovation Dialogue, and demonstrated it live there on October 9, 2026. Here’s how we built it and what you’ll find on our test reports.

The Challenge: Expert Reports Without a Standard Format

The Federal Ministry of Agriculture and Forestry, Climate and Environmental Protection, Regions, and Water Management (BMLUK) is behind this challenge. IÖB stands for innovation-promoting public procurement: The platform brings government agencies and suppliers together before a procurement process begins. This challenge is a market exploration, not a contract.

The IÖB Challenge „AI-powered data extraction from expert reports on waste“ explains that the reports are submitted as PDF, Word, or Excel files containing continuous text, tables, scans, and handwritten notes. Each author structures them differently. For the first year, the ministry expects 200 reports; after two to three years, it anticipates about 2,000 per year. Each report is 20 to 400 pages long, and with this volume, manual review is „no longer fully feasible.“.

Three guidelines shaped our approach:

• People decide: „There are no plans to implement a fully automated assessment of the uploaded documents.“ The software scans the documents and flags any areas of concern; the decision is left to the specialist.
• No made-up values: The challenge explicitly states that „tools that distort or fabricate values“ are not the goal.
• Confident operation: The solution should „be capable of running on its own instance in a sovereign cloud“ and be transferable to the ministry’s own servers. It currently runs on a separate instance in a European data center, which can later be migrated to the ministry’s data center to ensure full data sovereignty.

Our Approach and the Test Setup

Our solution is a three-step test process:

• Read: Konfuzio reads every document and organizes the data in a structured format, whether it's a PDF, a scan, or an Excel file.
• Check: A set of rules compares the values against the thresholds and with each other. Each validation rule is a sentence in natural language that is answered with "Yes" or "No.".
• Submit: The specialist sees each value along with its source in the original document and each failed validation rule along with its reason.

The source is the safeguard against hallucinations. A value without a source cannot be verified, but a value with a source can be verified with a single click.

We did not have any concrete evidence during our market research. We therefore used the public Ministry Proposal for solid alternative fuels, an Excel file with macros. The ministry states: „These are intended to serve as a guide and may be used on a voluntary basis.“ That is precisely why a program that simply reads the fixed cells in the template is not sufficient. Some of the supporting documentation may be based on the template, while the rest can be freely formatted, scanned, or submitted with laboratory reports attached.

We created test records using the template, incorporating fictitious values and built-in errors. The sampling log in the main document was filled out by hand and scanned with a cell phone. We derived the error-free control test from the completed Excel template. Below is an overview of these test documents. With the exception of the control test, they contain intentionally incorrect information to test the system and the AI.

That’s two reports in four files. Each one was fully read and verified in 24 to 31 seconds. The limitations of this test are obvious: Our reports are clean, concise, and based on a template. Real expert reports can be up to 400 pages long and have as many different layouts as there are authors.

The Extractor: 29 fields, 14 validation rules

An extractor in Konfuzio consists of a general instruction for the entire document, one instruction per field, and the validation rules associated with the fields. The general instruction ends with the sentence that is most important for a government agency: „If a value cannot be read with certainty, leave the field blank.“ A blank field stands out during verification, whereas a guessed value does not. For the lead median, it is explicitly stated that the measured value is meant, not the threshold value from the same table.

The 29 fields cover header data, the creator and waste owner, sampling, and calorific value. In addition, there are the median values for eight metals and semimetals, the creator’s assessment, and a free-form summary titled „Points to Be Checked.“ The 14 verification rules are distributed as follows:

• Threshold values: Eight testing criteria for the median values of antimony, arsenic, lead, cadmium, chromium, cobalt, nickel, and mercury.
• Objection: The assessment must not claim compliance if a value exceeds the limit.
• Date: The sample was taken prior to the exhibition, and the date on the report matches the date the sample was taken.
• Format: The GLN (Global Location Number) has 13 digits; the key number for the waste type has five.
• Plausibility: The 80th percentile may not exceed 1.5 times the median.

According to the regulation, the median and the 80th percentile summarize the last ten test results: The median is the middle value, and the 80th percentile is the value below which approximately eight out of ten results fall. The unit mg/MJ relates the pollutant to the calorific value, i.e., milligrams per megajoule of energy. The limit values are taken from the draft and correspond to the values listed in Appendix 8 of the Waste Incineration Regulation 2024 sets limits for substitute fuels in cement production facilities, such as 20 mg/MJ for the median lead concentration.

What the validation rules find

The EBS-2027-0418 report consists of eleven pages of documentation and two pages of handwritten notes. Konfuzio reports „11 of 14 checks passed.“ The three validation rules that were not met correspond to the three test cases we had included for the demonstration.

Threshold: Lead above 20 mg/MJ

The median lead concentration is 23.57 mg/MJ. The testing guideline states the reason in plain language: „The median lead concentration is 23.57 mg/MJ and exceeds the limit of 20 mg/MJ.“ Clicking on the field highlights the cell on page 11 from which the value originates.

Objection: The Author's Assessment

On page 4, the author explains that the fuel complies with the regulation’s limit values for the median and the 80th percentile. The consistency check compares this statement with the extracted values. Its result: „The assessment claims compliance with all limit values, even though the median lead value of 23.57 mg/MJ exceeds the limit value of 20 mg/MJ.“ Without this validation rule, the creator’s statement would appear unchallenged in the result.

Date: a two-day gap in the handwriting

According to the sampling plan, sampling took place on March 12, 2027, but the handwritten report on page 13 lists March 14, 2027. The field instructions deliberately distinguish between the two dates: one comes solely from the plan, the other solely from the report. A two-day discrepancy may have a harmless explanation. The verification rule flags the discrepancy; the specialist makes the final decision.

The summary is informative. It lists the lead level, the discrepancy, and two conflicting entries in the sampling plan, but not the date. Therefore, a fixed testing rule is needed for anything that must be reliably flagged. The summary is a supplement, not a substitute.

Handwriting: Reading Confidence for Every Word

Optical character recognition (OCR) provides a confidence score for each recognized word. The „Document Layout“ view displays this information word by word. On page 12, the handwritten table of samples, 144 words were recognized: 112 with at least 95 percent confidence, and 19 with less than 80 percent.

Among the 19 uncertain words are nine units of measurement and six timestamps. „10 l“ becomes „10€“ with a 9 percent error rate. Nevertheless, the value we needed from this page came out correctly: ten qualified samples. The reading reliability indicator shows the expert where it’s worth taking a second look. You can read more about this in our article on OCR with Handwriting Recognition.

Excel and the Cross-Validation Test

Konfuzio converts each worksheet in a workbook into pages and writes the name of the worksheet at the top, such as „Compliance with Limit Values.“ The workbook EBS-2027-0418 yielded 20 pages and the same median lead value of 23.57 mg/MJ as the PDF. The validation rule for the date failed here for a different reason: „No date is entered in the sampling log.“ The log sheet for this workbook is blank.

The EBS-2027-0611 counter-test, available in both Excel and PDF formats, meets all 14 testing criteria; the median lead level is 15.34 mg/MJ. “Green” here means: measured against the limit values for cement plants that the extractor checks. Nevertheless, the summary report flags three issues for which there are no specific test rules, including a blank signature on the sampling log. So even a “green” report gets reviewed by a specialist—just more briefly.

Conclusion: Reviewing waste assessment reports means comparing them

In this test, reading the data was the easier part of the job. The difficult part is the comparison. Which date counts depends on which part of the certificate it comes from, and a “green” certificate is only as “green” as the thresholds against which it is tested. An AI system for waste assessments therefore needs a set of rules in which each testing rule cites its legal source and which the experts at the ministry can read and modify themselves. Responsibility remains with the expert, just as it does for AI Agents in the Document Process have described.

The ministry decides whether the market assessment will lead to a pilot program. A pilot program would have to use real-world data to assess how reliably each field group is read. Our test files demonstrate the process, not the accuracy in everyday use.

Anyone dealing with similar expert reports at a government agency will find information on our page for the public sector the path to a conversation using your own documents.