A look at what artificial intelligence actually does in public procurement, and where it still falls short.

Arthur C. Clarke once wrote that any sufficiently advanced technology is indistinguishable from magic. He was right. And he was also describing exactly the problem with how most companies approach AI today.

The European public procurement market is no exception. Everyone is talking about AI. Fewer people are being honest about what it can and cannot do. At Hermix, we have been building AI into our platform since day one, using it to help companies like Capgemini, Accenture, Unisys, Fujitsu, Indra, and dozens of others win public contracts. We have learned a lot about where AI delivers real value, where it struggles, and how to close the gap.

The problem AI is actually solving

Public sector sales is expensive. Contractors spend 2-6% of contract value on sales and bidding. Presales managers dedicate up to 50% of their time to finding, reading, and qualifying tenders. Proposals are still largely written by hand. And somewhere in the middle of all that manual effort, good opportunities get missed and bad ones get chased.

The market is also fragmented. Relevant tenders are spread across hundreds of portals across Europe. Finding them takes hours. Reading the specifications takes more hours. Qualifying the opportunity commercially, understanding who has won similar contracts before, what budgets look like, who the competition is, can take days.

That is the problem AI is solving. Not magic. Compression of effort.

What we actually use AI for at Hermix

There is a lot of vague talk about “AI-powered” platforms. We prefer to be specific.

AI and machine learning handle data quality, cleanup, interpretation, scoring, and recommendations. Public procurement data is messy. Company names appear in dozens of variants across different portals. Authorities are duplicated, misclassified, or inconsistently tagged. ML models clean and normalize this at scale, automatically, so that when you search for a buyer or contractor profile, you are looking at accurate consolidated data rather than a fragmented mess.

GPT models analyze tender specifications, extract technical requirements, and support compliance checking. When a tender document runs to 200 or 300 pages, a GPT-powered summarization can produce a structured one-page brief covering objectives, budget, award criteria, required experts, and eligibility conditions in minutes. Our users can open any tender in the platform, click the Summary tab, and read a structured analysis without touching the original document.

AI Chat goes further. Rather than reading a fixed summary, users can ask their own questions directly against the tender document. Which are the experts needed? What is the minimum turnover requirement? How many CVs must be submitted in the technical proposal? The answers come back in seconds, drawn from the actual document, cited and traceable.

SearchGPT powers our authority and contractor profile building. Rather than manually assembling information about a buyer, SearchGPT browses, cleans, and structures publicly available information about buyers and contractors, combining it with our own procurement database to produce deep profiles. For a tender from Frontex, for example, users can instantly see 241 open opportunities worth €3.2 billion, 659 awarded contracts worth €5.6 billion, top preferred suppliers, and 11 contact points, all in one view.

GraphRAG and Deep Reasoning handle the harder problem of large, complex documents and analysis chains that require intermediate reasoning steps. Standard GPT models struggle with very long documents and multi-hop reasoning. GraphRAG addresses this by understanding relationships between entities in a document, not just surface text. Deep Reasoning generates intermediate artefacts and multiple answer paths before arriving at a conclusion. These techniques are what allow Hermix to analyze framework contracts, multi-lot structures, and consortium requirements accurately.

AI classifiers analyze and categorize clarification questions from procurement portals, helping users track and understand the Q&A landscape around active tenders.

What works well, and what does not

This is the part most AI vendors skip. We will not.

AI handles certain tasks with high reliability. Grammar, spelling, and style correction. Summarizing and extracting structured content from documents. Translation across European languages. Compliance checking against stated requirements. Classifying and tagging content at scale. For these tasks, accuracy in our systems runs at 98-99%.

AI struggles with other tasks, and it is important to be direct about this. Generating original ideas is not something current models do well. They recombine existing patterns, they do not invent. Understanding emotional context, cultural subtext, or the unspoken dynamics in a client relationship is beyond current capability. Non-supervised AI running without human review produces unreliable output at a rate that would be unacceptable in a bid process.

There are also technical limitations that affect procurement specifically. Very large documents with dense legal and technical content push the boundaries of context windows. Some less-resourced European languages have less training data and produce less reliable outputs. Scanned or encrypted documents that require OCR before analysis introduce additional error rates. Explainability, being able to trace exactly why an AI produced a particular output, remains a challenge, which matters in regulated procurement environments.

The honest conclusion: AI is a tool. A powerful one when used correctly. A liability when treated as a black box.

The adoption barriers that nobody talks about

Beyond capability, there are practical barriers to using AI in procurement that deserve attention.

Security and privacy concerns are real. Tender documents can contain commercially sensitive information. Regulations around data handling in public procurement are strict, and vary by jurisdiction. Any AI tool used in this context needs to be assessed against these requirements, not assumed to be compliant.

Upfront setup and integration costs are often underestimated. A platform that uses AI well has invested significantly in data pipelines, model selection, tuning, and quality control. That investment is embedded in the product. Building it from scratch in-house is expensive and slow.

User resistance and skills gaps are significant. AI tools require prompt engineering skills to use effectively, and teams that have spent years working with Excel and manual research processes do not automatically adapt. Change management matters as much as the technology.

The question worth asking

Here is the scenario that concerns us most, and it is worth raising openly. AI is now capable of writing procurement documents. AI is capable of writing proposals in response to those documents. AI is capable of evaluating those proposals.

If all three happen simultaneously, what are we actually measuring? The quality of the underlying offer, or the sophistication of the AI used to generate it?

For standard goods and services, price-based evaluation may remain dominant regardless of AI involvement. For complex, creative, or strategic projects, that dynamic is different. The human judgment, the original architecture, the insight about what a buyer actually needs beyond what they wrote in the specification, these remain genuinely human contributions that AI cannot replicate.

The companies that win consistently in public sector do not outsource their thinking to AI. They use AI to compress the effort of finding opportunities, understanding buyers, qualifying commercially, and managing documentation. The strategic judgment remains theirs.

What this means in practice

At Hermix, our AI tools support four stages of the procurement process: discovery, qualification, strategy, and bid orchestration.

Discovery means finding the right tenders faster, across TED, BOSA, Doffin, Germany, the Netherlands, Romania, and EU institution portals, with automated alerts and renewal forecasting so users never miss a relevant opportunity.

Qualification means understanding an opportunity commercially before committing resources to it. Who has won similar contracts? What did those contracts pay? What is the buyer’s history and spending pattern? What does the competitive landscape look like? Our platform answers these questions in minutes from a database covering over 250,000 authorities.

Strategy means market intelligence: buyer profiles, partnership analysis, competition graphs, EC spending data. Understanding where the money is and who the key players are before you decide where to invest your business development effort.

Bid orchestration means AI summarization, AI Chat, document analysis, and compliance checking to reduce the manual burden of the bid process itself.

The result, based on case studies with our users, is an 87% reduction in time spent on market research, a 75% reduction in time for tender analysis, and a 60% reduction in time needed for proposal writing. A four-person team delivers the output of six.

That is not magic. That is what well-implemented AI actually looks like in a complex B2G environment.

Hermix helps companies win public contracts with AI-powered tender monitoring, market intelligence, and bid analysis. Used by Capgemini, Accenture, Unisys, Fujitsu, Indra, Microsoft, and 50+ other companies across Europe. Create your free account at hermix.com/sign-up