Tesisquare
Tesisquare's multi-agent pipeline handles 97.0% of 955 supplier documents without fallback
Tesisquare, a Turin software company, built a pipeline that sorts, splits, reads and checks supplier invoices, with people reviewing only the uncertain ones. Its Head of AI reports that 97.0% of 955 production documents ran through the pipeline, and that accuracy reached 98.5% on a 100-document test.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
97.0%
Share of 955 production documents (processed up to 21 January 2026) handled end to end by the pipeline; 926 documents, while 29 went to older OCR tools or a manual web portal1
98.5%
Document-level accuracy (all required fields correct) of the full pipeline with human review, on a stratified 100-document subset of the production documents1
60.0%
Document-level accuracy of a baseline that sent each PDF straight to a language model, on the same 100 documents1
17.5
percentage points
Accuracy gain from adding the parsing step (document layout turned into structured text) before extraction, on the same 100 documents1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Companies that receive invoices from many suppliers face a steady stream of documents in different layouts and languages. Typing the data in by hand is slow and error-prone, and older OCR tools (software that turns scans into text) often need people to step in on a large share of documents. The authors also worry that a language model working alone can invent values, which is a risk with financial data.
What they built
Tesisquare, a software company in Turin, built a pipeline of five specialised steps, called agents. One recognises the document type and supplier, one splits batch files into single invoices, one converts each page into structured text, one uses a language model to pull out the fields, and one validates the result. Documents the system is unsure about go to a person on a screen that shows the document next to the extracted data. The paper says production uses Mistral-Small-3.2, an open-weight model (its files are published, so a company can run it on its own servers), with DeepSeek-OCR as a fallback for high-volume periods.
How they did it
The extraction step works from the structured text rather than raw scans, which the authors say cuts the amount of text the model has to read by 35%. The validation step checks formats and sums, and a confidence score decides which documents go to a reviewer, with thresholds that are typically set between 80% and 90%. When a reviewer corrects a field, the system turns the correction into updated instructions for the model and applies them to similar documents already waiting in the queue, without retraining the model. The production set covered more than 50 suppliers and documents in Italian, English, German and French.
Results
Of 955 production documents processed up to 21 January 2026, 926 (97.0%) ran through the pipeline without a fallback to older tools or a manual portal. On a 100-document test drawn from production, the full pipeline with human review reached 98.5% accuracy, against 60.0% for a plain model-on-PDF baseline. The structured-text step added the most, 17.5 percentage points. Reviewing a flagged document took approximately 45 seconds against around 120 seconds for fully manual handling.
What to copy
Treat people as part of the design, not as a safety net. Route only uncertain documents to them, capture every correction, and measure the share that stays on the automated path separately from accuracy.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Share of 955 production documents (processed up to 21 January 2026) handled end to end by the pipeline; 926 documents, while 29 went to older OCR tools or a manual web portal | 97.0%1 | |
| Document-level accuracy (all required fields correct) of the full pipeline with human review, on a stratified 100-document subset of the production documents | 98.5%1 | |
| Document-level accuracy of a baseline that sent each PDF straight to a language model, on the same 100 documents | 60.0%1 | |
| Accuracy gain from adding the parsing step (document layout turned into structured text) before extraction, on the same 100 documents | 17.5 percentage points1 | |
| Average human review time per flagged document in the current deployment, against fully manual processing of the same invoices | approximately 45 seconds (versus around 120 seconds)1 | |
| F1 score (a combined measure of missed and wrong fields) of Mistral-Small-3.2 in the authors' comparison of four extraction models | 92.9%1 |
Lessons for your project
Put a layout-reading step in front of the language model. In the authors' test, converting each page into structured text lifted accuracy by 17.5 percentage points, more than any other single step.
Let people correct the output and feed the corrections back as updated prompts. The authors do this without retraining the model, which keeps the cost of improving it low.
Add plain arithmetic checks, such as subtotal plus tax equals total, and send only low-confidence or failing documents to people. The authors set confidence thresholds of typically 80-90%.
Keep a non-AI fallback path. Unrecognised formats and badly degraded scans went to older OCR tools or a manual portal rather than producing unreliable output.
Expect a cold start on new suppliers. The authors advise starting with heavy human review and reducing it as corrections accumulate.
Caveats
- All figures come from a preprint by Tesisquare's Head of AI and a Politecnico di Torino academic. It was not independently audited or confirmed by a third party, and it does not say whether it has been peer reviewed.
- The paper does not name the end client whose invoices were processed, so the deployment context cannot be checked from outside.
- The paper gives no go-live date. It says documents were processed through January 2026 (data up to 21 January 2026), so the deployed date is shown as 2026 only.
- The 97.0% is a share of documents that stayed on the AI path. It is not an accuracy rate. The 29 documents that went to non-AI routes were left out of the accuracy evaluation.
- The 98.5% accuracy comes from only 100 documents (5 from each of 20 supplier and document-type groups), a small sample, and is the result with human review included. The authors' automated-only run on the same documents scored 92.5%.
- The paper also projects a reduction of about 70% in full-time staff for a scenario of 100,000 invoices a year, and savings in emissions, energy and water. These are modelled scenarios, not measured results, and are not used here.
- The 45-second and 120-second review times are stated by the authors for the current deployment without a description of how they were measured.
- The model comparison used the authors' own test set. The paper says Mistral-Small-3.2 is used in production, with DeepSeek-OCR as a fallback for high-volume periods.
Sources
- MADP: A Multi-Agent Pipeline for Sustainable Document Processing with Human-in-the-Loop
arXiv (Tesisquare Head of AI and Politecnico di Torino) · 2026-05-16 · Research paper · arxiv.org
Summarized in our own words from the sources above. Spotted an error? See corrections.