Case Study

aiZam - An AI Assistant for Public Procurement

Softwise.AI, in collaboration with Marketplanet, designed and delivered aiZam - an AI-powered assistant for public institutions preparing procurement notices. Built on a knowledge base of more than 100,000 documents from Poland's four largest public procurement data platforms, the assistant answers natural-language questions about real procurement procedures with cited sources, helps officials write better inquiries, and turns a process that previously required hours of document review into a few seconds of conversation.

CLIENT

Marketplanet (now Simplifae)

YEAR

2025

INDUSTIRES

Public Procurement B2B Technology

DELIVERABLES

RAG-based AI Assistant, Vector Knowledge Base, Document Processing Pipeline, Query Evaluation Model

The problem

Marketplanet (now Simplifae) is Poland's leading provider of procurement technologies, serving more than 1,500 companies and institutions including KGHM Polska Miedź, Orange Polska, Tauron, Santander Bank Polska, and hundreds of public offices.

Across its public-sector clients, one challenge kept surfacing: preparing high-quality public procurement notices requires precision, knowledge of regulations, and the ability to navigate extensive documentation - while procurement officers face mounting time pressure and inconsistent source material.

A typical procurement procedure can include dozens of attachments, only some of which are genuinely relevant. Officers were spending significant time searching for precedents, comparing criteria, and trying to formulate compliant inquiries - often without clear guidance on what good looked like. Marketplanet wanted to give them a smarter way to work: an AI assistant grounded in real procurement data, capable of answering substantive questions in seconds and educating users on how to improve their own notices.

Softwise.AI was asked to design and deliver that assistant, with one non-negotiable - every answer had to be traceable to verified source documents.

{Key Challenges}

Why generic AI couldn't answer real public-procurement questions?

The project responded to the real day-to-day reality of public procurement work:

Vast amounts of unstructured data
Procurement notices come with numerous attachments, from contract templates to detailed procurement conditions. Extracting valuable information at scale required a careful selection and processing strategy.
Low quality of procurement inquiries
Institutional employees often struggled to formulate precise, complete notices - leading to ambiguous tenders and downstream issues.
Need for real-time substantive support
Users expected immediate, expert-level help rather than having to sift through hundreds of pages of documents themselves.
Solution scalability
The system had to perform efficiently across a very large data volume while maintaining speed and answer quality.
Trust and verifiability
Generic AI tools were not an option - every answer had to point back to a real, verifiable source document.

{Process}

How we built aiZam in 6 stages, from PoC to launch?

Softwise.AI ran the engagement in clear stages, validating direction early and only investing in production-grade infrastructure once the concept was proven.

1. Discovery workshops with the client

The project began with workshops alongside the Marketplanet team, jointly analyzing procurement processes and the needs of public institutions. These sessions identified typical user problems and surfaced which parts of procurement documentation are critical for generating valuable answers.

2. Rapid Proof of Concept

A simplified prototype was built early on a narrow sample of manually prepared documents. This let both teams validate the approach quickly before committing to the full build.

3. Document selection

Across the four source platforms, two document types emerged as decisive: Special Terms of Procurement and Contract Templates. Restricting further processing to these dramatically improved answer quality and reduced noise in the system.

4. Vector knowledge base construction

The selected documents were processed, segmented into logical fragments, tagged, and transformed into a vector knowledge base - enabling semantic search that understands query context rather than relying on keyword matching.

5. User interaction design

The interface was designed so users can ask questions in natural language and receive answers with direct links to the underlying document fragments - verification is one click away.

6. Query quality evaluation

A final AI component was added to assess the quality of user queries and suggest improvements, turning the assistant into a tool that not only answers but also teaches.

Solution

At the heart of the solution is the AI-powered Public Procurement Assistant - a tool that supports the creation of high-quality procurement notices based solely on verified data from real procedures.

A knowledge base of more than 100,000 documents

The system integrates and processes data from the four largest public procurement platforms in Poland, with metadata enrichment covering number of submitted bids, procedure type, CPV codes, procurement budget, and contracting authority.

Quality-first data extraction

AI models read the content of each document, but only the two decisive document types - Special Terms of Procurement and Contract Templates - are passed downstream. Lower-value attachments are deliberately filtered out to protect answer quality.

Semantic search and generative answers

Natural language processing matches user questions to the right document fragments even when wording differs from the source. A generative layer then produces short, precise answers tailored to the query, drawn strictly from the verified knowledge base.

Filters, sources, and transparency

Users can narrow results by procedure type or CPV code, and every generated answer includes a direct link to the original document - so trust and verifiability are built in by default.

Query evaluation that educates

A complementary language model analyzes user queries and suggests improvements, helping officials learn how to write more precise procurement notices over time.

{Results}

Over 100,000 procurement documents distilled into one searchable knowledge base

Over 100,000 source documents consolidated into a single knowledge base covering Poland's main public procurement platforms.
Immediate AI-powered substantive support for public-sector users on criteria, budgets, and procedures.
High data quality through deliberate filtering down to the most relevant document types - Special Terms of Procurement and Contract Templates.
Semantic search with rich filtering by procedure type and CPV codes.
An inquiry-evaluation model that helps users formulate better procurement notices.
Full source traceability - every answer is one click away from the original document.

The interface was designed so users can ask questions in natural language and receive answers with direct links to the underlying document fragments - verification is one click away

{Business impact}

From hours of document review to seconds of natural-language Q&A

For Marketplanet and its public-sector clients, aiZam changes how procurement officers work day to day:

Faster preparation of notices. Hours of manual document review collapse into seconds of natural-language conversation.
Higher-quality, regulation-aligned procurement. Answers are drawn only from real procedures, raising the baseline for what good looks like.
Transparent, verifiable AI. Source-linked answers make the system suitable for a sector where every decision must be defensible.
A learning loop for users. Query evaluation gradually improves how officials phrase inquiries - a competence uplift, not just a productivity gain.
A scalable foundation. The architecture handles a large and growing document base while preserving speed and answer quality.

{Key recommendations}

5 rules for deploying AI on top of large document estates

A few principles emerged from the build that apply to any organization deploying AI on top of large, unstructured document estates:

Filter ruthlessly on document type. Quality of answers tracks quality of source selection - including everything is rarely the right answer.
Pair generative answers with citations. In regulated or scrutinized domains, source-linked output is what makes AI usable in practice.
Validate early with a narrow PoC. A small, focused prototype is the fastest way to confirm direction before investing in production infrastructure.
Treat semantic search as the backbone. Keyword search is not enough when users phrase the same question in many different ways.
Use AI to improve users, not just speed them up. A query-evaluation layer turns a tool into a teacher and compounds value over time.

SUMMARY

Softwise.AI and Marketplanet delivered aiZam, a production AI assistant that directly addresses the challenges faced by public institutions in creating procurement notices. Through deep document analysis, deliberate data selection, and natural language processing, the solution accelerates everyday administrative work, increases the quality and transparency of procurement processes, ensures regulatory compliance, and builds user competence through guidance. aiZam is a clear example of AI used not as a novelty but as a real tool that increases the effectiveness and value of public-sector operations.