How to Use AI Agents to Win Public Tenders

Key takeaways for suppliers from a webinar by Sorsera co-founder and CEO Georgs Vardanjans.

Let's start with a disclaimer: no AI agent will guarantee that you will win a tender. No CRM guarantees that customers will buy your service, either. But an agent can increase your chances, cut preparation time, and flag risks that a human might miss. Below, we've summarized what Sorsera has learned building tools for public procurement since 2021. We've also included practical steps that you can start taking this week.

Why This Matters Right Now

Public procurement accounts for 13 to 14 percent of Latvia's GDP. Across Europe, the average is around 15 percent. The European Commission estimates that inefficiency here costs roughly 100 billion euros per year. One of the most visible indicators: 25 to 30 percent of tenders attract only a single bidder. Research shows that each extra supplier in a tender saves the contracting authority about 4 percent of the contract value. More competition means more efficient use of public funds.

Suppliers cite restrictive technical and qualification requirements as the biggest barriers. Difficulty accessing procurement information and language barriers in foreign markets compound the problem. In Latvia, foreign companies win only about 2 percent of tenders. Information asymmetry is real and expensive. This is exactly where AI can give suppliers a concrete advantage.

Three Levels of AI Automation

Before we talk about agents, we first need to define the levels of AI automation.

The first level is productivity tools. ChatGPT, Claude, Gemini. You chat, they help. All major large language models (LLMs) perform at a similar level right now. Switching one for another won't produce a massive productivity jump. The main benefit comes from using them at all.

The second level is semi-autonomous agents. Think of it as an assistant that works in sync with you. You wouldn't leave it running unsupervised, but it can handle a tangible share of the work.

The third level is autonomous agents. You start it up and close your laptop; the agent reaches the goal on its own. In procurement, this will remain the exception for a long time. We're dealing with sensitive information where mistakes are costly. Humans on both sides will need to stay involved for the foreseeable future.

Context Engineering, Not "The Perfect Prompt"

A common mistake is spending hours searching for the ideal prompt. The field has moved past that. What matters more now is what you actually feed the model. In other words, context engineering.

The context window is a limited resource. Modern models accept hundreds of thousands of tokens, but you can't upload an entire set of tender documents and expect a precise answer. A 3,200-line tender specification or a hundred-page PDF will leave the model unsure where to look. What makes the difference is selecting the right document sections, historical data, and legal references before asking the agent anything.

What an Agent Actually Is

Anthropic, the creators of Claude, defines an agent in a single sentence: a large language model that uses tools across several iterations. Three components: the model, the system prompt, and the tools. The platform where the agent lives adds memory.

Tools are the agent's muscles. Web search, sending emails, database queries, document analysis, calendar entries. At Sorsera, we build our own tools for procurement: Public Procurement Law interpretation, CPV code selection, historical contract analysis, and competitor research. The agent decides on its own which tool to call and how many times. You give the agent various skills, and it figures out the necessary steps by itself. That's the magic you notice in day-to-day use.

When thinking about agent design, focus less on the system prompt. Focus more on what tools are available and how much freedom the agent has to use them.

The Supplier Process in Four Stages

At Sorsera, we internally divide the supplier's work into four activities. Each stage has its own agent design.

1. Tender Monitoring

In an ideal world, you'd get a notification for every tender that's relevant to you and no irrelevant ones. In the real world, there are two painful scenarios. First: a tender is relevant to you, but you never got notified. That's a missed contract. Second: you get notifications, but they're not a good fit. That's noise.

What can an agent do at this stage? Analyze your company's website and past tender participation to suggest keywords and CPV codes. Go beyond the description and title to find semantic matches. That means catching tenders that are similar in substance, even when the terminology differs. In our monitoring filters, the user clicks a few buttons, and the agent fills in the rest.

2. Tender Analysis

The first practical technique you can use today is Claude Projects (or the matching feature in other tools). Create a project for one specific tender, upload the documents, write your instructions, and chat with the agent about that tender. The whole team can work in one place.

But remember context engineering. If the specification runs to 500 pages, tell the agent where to look: technical specification, qualification requirements, penalty clauses, warranty terms. Don't trust it to figure out priorities on its own.

An agent needs several tools for tender analysis: web search, document analysis, organization analysis to understand what the contracting authority has purchased before, a historical tenders tool for frequent suppliers and contract price ranges, and competitor analysis.

On the Sorsera platform, we have a data analysis tool that shows top suppliers by industry, their win rate as a percentage, the most common contracting authorities, and price ranges. Take "education and training" as an example. An agent with access to this tool can answer: "What are the largest construction tenders in Latvia over the past year?" It returns specific contracts, amounts, and dates.

3. Tender Strategy

This is the stage where you decide whether to bid at all. We recommend a simple method from statistics: expected value.

Expected value = probability of winning x (financial value + strategic value)

Example: you believe you'd win 25 percent of the time and the profit would be 10,000 euros. The strategic value (references, a new client relationship) is 2,000 euros. The expected value works out to 25% of 12,000 = 3,000 euros. If the cost of participating is less than that, submit a bid.

Price is the lever the supplier can actually control. The higher the price, the lower the chance of winning. But the potential profit grows. At Sorsera, we've built a price simulator that shows how expected value shifts depending on price. It factors in your assumptions about competitors, their price ranges, and the risk of disqualification. The goal isn't to take the simulator's "optimal" number and plug it into your financial bid. The goal is to structure your assumptions and understand where you sit on the curve.

4. Bid Preparation

This is where you need to be the most careful. AI models hallucinate less than a year ago, but it still happens. All major AI providers state this about their models. In procurement, a great deal is evaluated word by word. One comma, one line can determine whether a bid gets rejected.

Anthropic's published risk-versus-autonomy chart is a useful reference here. The vertical axis is risk level. The horizontal axis is the autonomy you give the agent. A financial bid is a high-risk zone, and agent autonomy should stay low there. Error-checking in already-drafted text is low-risk and high-autonomy. The agent can work on its own there.

Our breakdown in practice looks like this.

Human only: tender strategy, the final number in the financial bid, the decision to bid, the decision on partners.

Human-led, AI assists: technical bid preparation, qualification documentation, drafting descriptions of past projects, copying and adapting.

AI-led, human supervises: finding grammar and logic errors, spotting contradictions between sections, checking compliance with requirements, and translation.

On Confidentiality

Three practical steps that are often skipped.

First, go into your AI provider's settings and turn off the option that says "use my data to improve the model." It may appear as "data training," "improve for everyone," or something similar. Without disabling this setting, your sensitive data may end up training the model.

Second. For API access, this option is usually off by default. Technology companies (including us) work through the API, and data is not used for training.

Third, keep in mind that the act of sending data to any cloud (including Google Drive) is a business decision. Risks always exist; you manage them on purpose. If your organization has an IT security specialist, talk to them.

The Subcontractor Scenario

Not everyone wants to be the party signing the contract. Many companies want to be part of the supply chain. An AI agent is especially useful in this role in two scenarios.

First: semantic monitoring. Say you're an electrical supply subcontractor. Every construction tender where an electrical component might appear is relevant to you, even if the CPV code and title don't show it. The agent analyzes the technical specification and flags the matching tenders.

Second: winner monitoring. You mark the major construction companies in Latvia. Every time one of them wins a contract, you get a notification the same day. You call, congratulate them, and offer to collaborate. You might be the first one to do it.

How to Get Your Team to Start Using AI

An observation from Sorsera's experience and client conversations: people are already using it. One large Latvian company hadn't adopted AI on paper. Then an OpenAI sales rep told them that hundreds of employees from their email domain were already using ChatGPT. The question isn't whether employees will use it, but whether they'll use it in an organized and secure way.

Practical steps.

Set up licenses at the organization level. Don't let everyone create personal accounts with their corporate email. Turn off the option for data to train the model. Create shared projects so the team doesn't duplicate the same work in parallel. Show real examples from daily work: how a developer uses Claude Code, how a sales rep prepares an email, how a procurement specialist analyzes a tender specification. At Sorsera, our 12-person team is about half technical. Developers no longer write lines of code themselves. A year ago, that would have been hard to believe.

Improving agents starts with understanding how they work. At the center is a black box that nobody, not even Anthropic themselves, understands. But everything around it is transparent: what tools are available, how often they're called, where the agent makes mistakes. If an agent is doing something wrong, start with the system prompt. Then look at tool availability. Only then consider switching the model.

Conclusions

AI won't replace supplier decisions about which tenders to enter or how to win. But it is closing the information gap that has always favored the contracting authority over the supplier. If your competitors are using agents and you're not, you risk losing the competitive edge.

Three steps you can take this week.

Set up organization-level licenses for Claude, ChatGPT, or Gemini. Turn off the data training option. Create a Claude Projects workspace (or a similar tool) for your first serious tender and put all the documentation in one place. Write a one-page document with your expected value formula, assumptions about your main competitors, and price elasticity.

If you work in public procurement and want to try specialized procurement agents, get in touch with us. We're building with pilot organizations in Latvia and abroad, including CERN and Latvenergo. We enjoy working on problems that nobody has solved yet.

Based on a webinar by Sorsera founder Georgs Vardanjans.

Sources:

• Georgs Vardanjans, Sorsera webinar "How to Use AI Agents to Win Public Tenders"

• Anthropic publications on agent design and Claude usage recommendations (anthropic.com)

• European Commission public procurement statistics on procurement's share of GDP and inefficiency costs