What happened
An IT hardware salesperson got tired of doing detective work before every cold call, so he handed the job to an AI agent. The project is called AI Tender Radar, and it does exactly one thing well: it watches public and commercial tenders for servers, storage systems and components, reads the attached paperwork, hunts down a phone number, and drops a one-screen briefing card into Telegram. In the author's own framing, it finds clients in government procurement many times faster than doing the same research by hand.
He works in IT equipment sales and runs systems integration projects, so his ideal customer is a company that needs servers or a storage system right now. The good news is that those companies announce themselves. They publish a tender — a public procurement notice — that says what they want, how much of it, and by when.
A tender is a signal, not a conversation
Here's the catch. A tender tells you someone has a need; it doesn't tell you who to call. To make a phone call worth making, a rep has to read the technical specification (the document that spells out the requirements), work out what kind of company the buyer is, check whether they actually have money, and find a human being with a phone number. That's a serious chunk of unpaid work per lead, and plenty of tenders turn out to be a bad fit only after all that reading.
So he automated the homework, not the call.
What actually lands in the rep's Telegram
Each card fits on one screen and has four blocks:
• The need. What equipment, how many units, and the deadline — pulled from the technical specification rather than the tender title, because titles say almost nothing.
• The customer. Revenue, profit and capital for the most recently published year, taken from public tax-service filings.
• The contact. A phone number, and if the agent found a specific person, their name and role.
• The opening. How to start the call and which questions to ask, based on what's in the documents.
Two buttons sit under the card — “In work” and “Pass” — and a rep can reply with a note. New leads and in-progress leads export to Excel. Cards arrive on weekdays at 9:00, 12:00 and 15:00 Moscow time, and tenders whose submission window has closed never show up at all. The agent doesn't call or email anyone. It prepares; a human decides.
Why version one collapsed
The first build was a classic pipeline: read the tender, look for a contact, if there's nothing check the customer's website, and so on. Think of it as a train on fixed rails — fine until the rails run out. Tenders are wildly inconsistent. A contact might sit inside the notice, in an attachment, on the company website, or somewhere else entirely. Every new pattern meant another branch of code, and the branches multiplied.
Version two swapped the rails for a map. The agent gets a set of tools and a written brief, and decides for itself which files and pages to open and when to stop. All of its tools are read-only: read the tender and its documents, search the web, open a web page or a PDF, and pull company financials from the tax service. The brief sets the order — the tender and its documents first, then the customer's website, then open sources.
The fences around the agent
Giving a language model freedom sounds reckless until you see the guardrails. Every research run has a cap on steps and a cap on model spend. Every link the agent follows gets validated so it can't wander onto internal addresses. Anything it reads on the internet is labelled untrusted, so some random web page can't smuggle instructions into the agent's head. If a model response can't be parsed, the run repeats once; after that the tender is flagged as failed instead of quietly disappearing. The documents it can handle are PDF, DOCX, DOC, XLSX, TXT, and ZIP or 7z archives. Scans with no text layer, RAR files and old XLS files aren't supported yet — those are marked as a data gap, and the agent goes looking for the same facts elsewhere.
What it means for you
You probably don't sell servers. The pattern still travels, because “someone published a need, go read it and figure out who to call” is most of B2B sales, and a fair bit of recruiting, grant-hunting and flat-hunting too.
If you sell to businesses
A five-person systems integrator loses one person's morning to tender paperwork. An agent that pre-reads and pre-qualifies ten notices before 9 a.m. changes how many calls the team gets through, without a single new hire.
For a sales manager, the real value isn't raw speed — it's sequencing. Cards land at 9, 12 and 15, so the vague “I'll get to the research later” becomes a routine you can actually keep. Nobody has to remember to do it, because it's already in the chat.
If you're a one-person operation
Public procurement is basically a published list of organisations with budget and a deadline. That's about as close to a free lead list as you'll get, and the barrier was never the finding — it was the reading.
When you're the sales team, the researcher and the finance department all at once, a card that already tells you a buyer's revenue and profit saves you from pitching a company that can't pay. That's a different kind of speed: fewer wasted calls, not just more of them.
If you're learning or creating
The research side of AI Tender Radar is open on GitHub, which makes it an unusually good study object — a working agent with real tools, hard limits and a written brief, rather than another toy chatbot demo.
The fourth block on the card is the hardest to fake: what to say first. If you'd rather draft those lines yourself, free AI tools like the ones at MyKreaTool are a decent place to turn a dry technical specification into three questions that are actually worth asking.
How to try it right now
You don't need to be an engineer to copy the shape of this. Start free, then add pieces.
Run the free open-source version first
1. Grab the research code. The research part of AI Tender Radar is open on GitHub — look the project up by name, read how the agent reads documents and where it stops, and run it against a handful of tenders you already know the answer to. That's the cheapest way to find out whether the output is useful to you.
2. Write your filter in plain words. The author's own filter is deliberately wide: servers, storage and components, government and commercial, no minimum price. Yours should be just as plain, because that sentence is what you'll paste into the brief.
Give the agent tools and a brief
3. Pick four blocks, not ten. Need, customer, contact, opener. Don't add a fifth until the first four are reliable.
4. Keep the tools read-only. Read the notice, read the attachments, search, open a page or a PDF, pull the financials. Nothing that writes, sends or buys.
5. Write the search order into the brief. Notice and attachments, then the customer's website, then open sources. That one line replaces dozens of code branches.
6. Set the limits before the first run. A step cap, a spend cap per research, link validation, and untrusted-text labelling. This is the difference between an agent and a runaway bill.
The no-code version you can start today
7. Schedule the delivery. If you're doing this by hand, block 20 minutes at 9, 12 and 15 to process whatever came in and mark each tender “in work” or “pass”. Same rhythm, no code, and the habit survives even if you never automate.
8. Use a chatbot as the agent. Paste a tender's documents into a free chatbot, ask for the four blocks one at a time, and keep results in a spreadsheet with two status columns. Less slick than Telegram, still far faster than reading everything cold.
Upsides and what changes
The obvious win is throughput. Prepping a tender properly eats a chunk of a morning; the agent does it while you sleep, and it doesn't skip the boring parts because it's in a hurry.
The subtler win is consistency. Every card has the same four fields, so a manager can compare leads instead of re-reading a mess of notes — and a “pass” becomes an actual decision rather than something you forgot to do.
The guardrails matter as much as the speed. Step and spend caps keep costs predictable, link validation keeps the agent out of places it shouldn't go, and treating web text as untrusted is basic hygiene in a world where any page can hide instructions for a machine reader. The system also fails loudly: a tender it can't parse gets flagged, not silently lost.
And the human stays in the loop. The agent doesn't know your prices or what you can actually deliver, so it never proposes terms. It hands over context and keeps its mouth shut.
Limitations
Be clear-eyed about what this is. The agent reads PDF, DOCX, DOC, XLSX, TXT and ZIP or 7z archives, but it can't handle scanned documents without a text layer, RAR files or old XLS files — it flags those as gaps and hunts for the same facts somewhere else, which sometimes leaves the card thinner than you'd like. It has no idea what you sell or at what price, so every commercial judgement is still yours. The customer financials come from public tax-service filings, which means the card is only as rich as what a given market publishes. Research also runs into hard caps on steps and spend, so a genuinely tangled tender can end up marked as failed rather than half-finished. And it never contacts anyone — if you were hoping for a machine that books meetings, this isn't it.
Conclusion
The interesting part of AI Tender Radar isn't the speed claim. It's the shape of the job: read the notice, check the money, find the human, prepare the first sentence. Repetitive, evidence-based, and boring enough that most reps quietly skip it. Handing that to an agent with tools, limits and a written brief is a template you can copy well outside government procurement.
One action for today: pick a single tender in your field and read the attached technical specification instead of the title. That's step one of the card — and you'll immediately see how much of your pipeline work is really just reading.



Comments 0