BusinessRnD · draft v2 · built 2026-10-21 06:15 UTC
I2 · Prozorro tender co-pilot — deep dive
Test it
Best of the batch, but incumbents already sell AI checks for ~100 UAH. A prepaid test decides.Realistic: $1–4k/mo (my estimate) · Chance of $10k/mo: low–medium
Cards have Go deeper, Maybe and Kill buttons. Press Send feedback to save. Your marks from all tabs collect at the end of each page.
Prepared 2026-10-20 from the Prozorro public API (53 full tender records across 10 procedure types, 2,308 decided tenders from one overnight slice, all 141 translation tenders of the last 12 months), the text of Law 4888-IX, Prozorro's standards repository and competitor pages. Rates from the overnight slice are indicative, not annual statistics. Notes and data: research_notes/Owned ideas deep dives/i2-prozorro_notes.md, i2-prozorro_stats.json, i2-prozorro_sample53.csv.
Bottom line
Almost everything is public, but mostly not structured. Open tenders have a median of 13 criteria and 21 requirements in structured fields, but 89% are bare "I confirm" boxes. Of 439 recent 24-hour defect notices, 97% cite a clause in the Word documentation (6.7k–19.5k words). Rejection grounds, protocols, notices and bidders' contacts are all public (Prozorro API). A system can triple-check formal completeness, but can't honestly guarantee technical compliance or a win.
The pain is measurable. In 834 decided open tenders, 20.4% of awards were rejected and 34% of tenders got a 24-hour correction notice. One rejection in five was "didn't fix within 24 hours". Scaled up, that's roughly 40–50k rejections and 60k notices a year (estimate).
"AI reads the tender documents" already sells at about 100 UAH a tender. SmartTender's SmartCheck AI charges 89–129 UAH per analysis on promo (SmartCheck AI); Newtend has Newtend AI (Newtend). The draft report missed this. Your edge: each buyer's defect history, deterministic checks of the customer's own files, a 24-hour rescue and a written guarantee.
The MVP needs no runtime model. Requirements are the same for every bidder, so documentation is read once per tender: about 20–60 a day in 2–3 CPV niches, as operator work on your Claude Code subscription if you approve. File checks run on rules plus local OCR.
The most personalised channel is public rejection data: "Your bid in tender X got a notice for Y — here's a free check of your next one." It's legally grey (e-commerce consent versus opt-out, personal-data notice), so email companies once with an opt-out after a lawyer's review. Reach FOPs through a "Why did I lose?" lookup, buyer pages and opt-in alerts.
Sell a subscription, not arbitrage. Plans at 490 and 1,490 UAH plus 249 UAH checks. About 330 subscribers on 1,490, or about 485 on a blend, gives $10k/month profit: 0.8–1.2% of the "more than 42,000" competitive bidders in 2025 (Forbes UA). Bidding yourself fails the zero-owner-time test: in translation, 94% of tenders demand prior analogous contracts and prices land 38% below estimate.
Timing: Law 4888-IX takes effect on 24 March 2027, with machine-readable documents, one correction round per bidder and coded rejection grounds; Prozorro has already drafted 57 codes (Rada; ProzorroUKR). Launch on today's rules; have the 2027 rule set ready by March.
Is everything public?
Coverage by item (from the 53-record sample; "open" means open tenders with особливості under Resolution 1178):
What can be triple-checked. Three independent layers back each other up:
Structured criteria: the skeleton, straight from the API (criteria standard).
Documentation-derived requirements: read once per tender from the Word files. Today about 98% of the non-signature tender files are Word. The 2027 law bans scans and photos.
This buyer's history: what it flagged in past 24-hour notices and rejection protocols. If layer 2 misses something a repeat buyer always checks, layer 3 catches it.
Verifiable without judgement: a file for every requirement; formats and sizes; a valid КЕП signature; the signer being the authorised person in the EDR register; issue dates before the bid deadline (Law 4888-IX bans later documents except the bidder's own letters); "not older than N days" rules; VAT, single-tax and sanctions registries; bid validity and language. These are frequent in notices: authority documents 50%, signatures or seals 44%, dates 12%.
Still judgement: technical-spec compliance (31% of notices), whether a past contract is "analogous" (30%), free-form letters. A third of bid PDFs are scans, so checks need OCR.
What the guarantee can honestly be:"If your bid gets a 24-hour notice or a rejection for a formal item our check marked green, we refund that check and give you a month free." Liability is capped at fees paid, formal items only, never technical compliance or winning. Backtest against historical notices first and publish the miss rate. Under a 5% miss rate (target), refunds cost under 5% of check revenue (estimate). Have a lawyer check the terms.
Alerts by CPV, region and value, with deadlines and bid-security amount (email first, Telegram optional).
Tender card with the three-layer checklist; each item shows its source (field, clause, or "flagged by this buyer N times").
Document vault with expiry tracking, reused across bids.
Bid-pack check: green/amber/red report with the fix for each problem.
24-hour rescue: alert within minutes, notice turned into a to-do, templates, countdown.
"Why did I lose?": your public rejections and notices, explained.
Letter templates (consent, sanctions declaration, authority), filled from your data.
flowchart TD
A["Free alert or why I lost lookup"] --> B["Pick CPV niches and region"]
B --> C["New tender alert"]
C --> D["Tender card and 3-layer checklist"]
D --> E{"Bid on it?"}
E -->|"no"| C
E -->|"yes"| F["Attach files from vault"]
F --> G["Bid-pack check"]
G --> H{"All green?"}
H -->|"no"| I["Fix list and templates"]
I --> G
H -->|"yes"| J["You sign with КЕП and submit on your майданчик"]
J --> K{"24h notice posted?"}
K -->|"yes"| L["24h rescue to-do and countdown"]
K -->|"no"| M["Result and lessons saved"]
L --> M
Submission stays on the customer's platform under their КЕП; you never hold their keys.
Getting to FOP owners
Channel
Reach (est.)
Cost
Customisation
Legality
Your time
Email after a public rejection or notice (companies only)
40–50k events a year, filtered to niches
~0
Very high: names the tender, defect and fix
Grey: consent versus opt-out (Art. 10); notice within 30 working days (Art. 12)
Once: lawyer review, sender identity
"Why did I lose?" lookup by EDRPOU/РНОКПП
Every bidder who searches their own code
~0
Very high, self-serve
Low risk: data is already public, nothing new is published
None
Buyer and CPV pages ("what this buyer rejects")
Long-tail search; Clarity already has per-company pages
~0
Medium
Low risk if pages don't name FOPs
None
Opt-in alert channels per CPV (email, Viber, Telegram)
Thousands per niche (est.)
~0
Medium
Clean: opt-in
Seeding posts at launch
Smaller майданчики (white-label checks)
Their supplier base
Revenue share
Low
Clean
Calls and a contract
Accountants and FOP service firms, associations
Medium
20–30% affiliate (est.)
Low
Clean
A few calls
Diia.Business marketplace, Prozorro courses
Medium, high trust
~0
Low
Clean
Application, maybe a talk
Ukrainian YouTube explainers ("439 real defect notices")
Slow, compounding
~0
Low
Clean
~0 (AI voice) or optional face
Big platforms (SmartTender, Newtend) are competitors first, not partners.
Recommended sequence:
flowchart LR
A["Pick 2-3 CPV niches"] --> B["Why I lost lookup and buyer pages"]
B --> C["Opt-in niche alerts"]
C --> D["Lawyer-approved email pilot, 300 companies"]
D --> E{"3 pct clicks and 1 pct pay?"}
E -->|"yes"| F["Scale to 1000 a month in niches"]
E -->|"no"| G["Fix offer or stop outreach"]
F --> H["Accountant affiliates and small platforms"]
H --> I["YouTube and Diia.Business"]
Subscription or arbitrage?
Model
Price (estimate)
Who pays
To reach $10k/month profit
Your time
Main risk
Verdict
Subscription tiers
Free / Старт 490 / Про 1,490 UAH a month
FOPs and small firms
~330 on Про or ~485 on a blend
~0
SmartTender bundles it
Core
Pay per check
249 UAH (SmartCheck: 89–129 promo, 480 list)
Occasional bidders
~2,000 checks a month
~0
Price war
Top-up
Success fee
1–2% of the award, with a minimum
Winners
Verifiable from public awards; collection is hard
Invoices, disputes
Collection
Pair with done-for-you only
Done-for-you bid pack
1,500–5,000 UAH per bid + success fee
Busy suppliers
~150 packs a month
Escalations
Turns into consulting; liability
Later, capped
Data/API (defect radar, leads)
Contract
Platforms, consultants, banks
Few deals
Calls
Newtend already sells bank leads
Later
Arbitrage: bid yourself
Contract value
State buyers
See below
Sign every bid, deliver, carry penalties
Experience barrier, price pressure, competing with your own customers
No, not now
Arbitrage reality check: written translation (CPV 79530000-8, 141 competitive tenders, last 12 months). 114 awards worth 40.2M UAH, median 88,556 UAH; only 23 suppliers won, the top 5 took 57%. Prices land 37.7% below the buyer's estimate. 94% require an analogous past contract, 57% qualified staff, 18 notarised or sworn translation. Payment is almost always after delivery (140 of 143 terms, median 10 days). FOPs took 45.6% of awards. So the cleanest AI-friendly category is about $1M a year, closed to newcomers and fought on price, and you'd sign every bid and carry delivery penalties. Other "AI-deliverable" codes turn out to be mostly printing, valuation, teaching equipment or software maintenance.
Recommendation: a self-serve subscription with per-check top-ups. Add a capped done-for-you pack with a success fee after 50 paying customers. Arbitrage only as a later one-off probe in direct contracts under 50k UAH, where no tender is needed (Law 922-VIII), outside your customers' categories.
I2.1
24-hour rescue
What it is: Watches every bid a subscriber submits. When the buyer posts a 24-hour defect notice, a phone alert arrives within minutes, with the notice turned into a to-do: each defect, the clause, the documents needed and the rule that new documents can't be dated after the deadline. Templates and a countdown come with it.
Who it serves: Any bidder in open or simplified tenders. 34% of decided open tenders and 26% of simplified ones got a notice.
Evidence: 344 notices across 811 open tenders. 50 of 240 rejections were "didn't fix within 24h". From March 2027 a buyer may send only one notice per bidder (Law 4888-IX).
Why we can win: Platforms send status alerts, but none found turns the notice into a fix plan. It's deterministic, works whichever platform the customer uses, and it's the most emotional moment in the whole process.
Build effort (AI): About 1 week, including a backtest on the 439 notices collected.
Compute: No model. Rules run on the public notice text.
Revenue model: Included in Старт. The first rescue is free as the hook.
First step: Build the API watcher and run a dry run on yesterday's notices in your niches.
I2.2
Buyer defect radar and three-layer checklist
What it is: Every tender in your niches gets a checklist built from three sources: structured criteria, requirements read from the documentation, and "this buyer flagged X in N past notices". Public buyer and CPV pages show aggregated patterns.
Who it serves: Suppliers who bid often with the same buyers, which is common: 63% of open tenders get a single bid.
Evidence: 97% of notices cite documentation clauses; only 12 of 181 structured evidence slots describe a document. SmartCheck sells layer 2 only (SmartCheck AI).
Why we can win: Layer 3 is a data asset built from public history that incumbents don't show, and it compounds daily.
Build effort (AI): 1–2 weeks for 2–3 niches.
Compute: Claude Code subscription — operator, once per tender, as a bounded niche batch (needs your OK). History mining needs no model.
Revenue model: Старт and Про plans. Buyer pages are free SEO.
First step: Backtest: does the extracted checklist contain the items buyers actually flagged in the past? Target recall of 90% or more.
I2.3
Bid-pack check with a formal-error guarantee
What it is: The customer uploads their pack. Deterministic checks: presence against the checklist, КЕП validity and signer versus the EDR director, issue dates and "not older than" rules, VAT/single-tax/sanctions registries, formats, sizes, language. Scans go through local OCR. Output: a red/amber/green report backed by the refund-plus-month guarantee.
Who it serves: FOPs and small firms without a tender specialist.
Evidence: Signatory-authority documents appear in 50% of notices, signatures or seals in 44%, dates in 12%. A third of bid PDFs are scans.
Why we can win: No competitor found checks the supplier's own files or offers a guarantee.
Build effort (AI): 2 weeks; КЕП verification needs the certified Ukrainian signature library.
Compute: No model (rules plus local OCR). Optional tech-spec matching runs customer-side.
Revenue model: 249 UAH per check, or 10 a month in Про.
First step: Verify 20 public bid packs end to end (signatures, dates, registries) to measure false alarms.
I2.4
"Why did I lose?" lookup and rejection-triggered outreach
What it is: A free lookup (type your EDRPOU or РНОКПП, see every notice and rejection with fixes), plus, for companies only, one email after a public notice or rejection offering a free check of the next bid.
Who it serves: The roughly 40–50k rejected bids a year and everyone who got a notice (estimate).
Evidence: Contacts were public in every sampled bid; rejection grounds always, protocols in 221 of 240 rejections.
Why we can win: Nobody else turns rejections into personal, timely offers.
Build effort (AI): 1 week.
Compute: No model.
Revenue model: Lead generation for the subscription.
First step: A lawyer reviews the email text, the privacy notice (Art. 12) and opt-out; then a 300-company pilot.
I2.5
Niche alert channels and buyer pages
What it is: Opt-in alert feeds per CPV niche (email, Viber, Telegram) and daily-refreshed pages per buyer and CPV.
Who it serves: Suppliers searching for tenders or researching a buyer.
Evidence: About 115k open tenders a year (Prozorro); Clarity already has company pages, so differentiate with defect data.
Why we can win: Unique "what this buyer rejects" content at near-zero cost.
Build effort (AI): 1 week.
Compute: No model for the data. Claude Code subscription — operator for page copy.
Revenue model: Free funnel into the Старт plan.
First step: Launch pages for the top 200 buyers in your niches.
I2.6
Partners: smaller platforms and accountants
What it is: A white-label check for platforms without AI, a 20–30% affiliate link for accountants and FOP service firms, a Diia.Business marketplace listing.
Who it serves: Suppliers who trust these intermediaries.
Evidence: Platform fees are regulated and identical (Newtend tariffs), so platforms compete on extras; the big two already have AI.
Why we can win: A neutral check small platforms can't build themselves.
Build effort (AI): 1 week after I2.3.
Compute: Same as I2.3 (no model).
Revenue model: Revenue share or affiliate commission.
First step: After day 30, a one-page offer to 5 small platforms and 20 accountants.
I2.7
Done-for-you bid pack with success fee
What it is: The AI assembles the pack from the vault, fills letters and runs I2.3; the customer reviews, signs and submits. 1,500 UAH plus 1% of the award on a win (estimates); public awards make the fee verifiable.
Who it serves: Busy suppliers bidding occasionally on larger tenders.
Evidence: Platforms' legal teams sell bid preparation without published prices (Newtend); SmartTender's advertorial puts reading the documentation alone at 30 minutes to 3 hours for a specialist.
Why we can win: AI does the hours of work; the guarantee and the defect radar back it.
Build effort (AI): 1 week on top of I2.3.
Compute: No model, plus customer-side for anything judgement-heavy.
Revenue model: A fee per pack plus a success fee, capped at a few packs per week at first.
First step: After 50 paying subscribers; scope excludes legal advice and appeals.
I2.8
Arbitrage probe: bid as a supplier yourself
What it is: Your FOP bids where AI can do the work, starting with written translation.
Who it serves: You; state buyers pay.
Evidence: See the translation numbers above. FOP group 3 can sell to anyone up to 1,167 minimum wages a year (Tax Code).
Why we can win: Only on price and speed in direct contracts under 50k UAH.
Build effort (AI): Low to build; high recurring load (your КЕП on every bid, delivery quality, penalties).
Compute: Claude Code subscription — operator (your own production, not a customer feature).
Revenue model: Contract margin.
First step: None now. Revisit only after I2 has 100 customers and only outside their categories.
Where the AI compute runs
Step
Runs on
Who pays
Notes
Tender ingestion, alerts, deadlines
No model
You (server)
Public API; about 9,000 tender changes a day
Structured criteria into a checklist skeleton
No model
You
100% of open tenders
Reading tender documentation into requirements, once per tender
Claude Code subscription — operator
You (subscription)
Shared by all bidders; 20–60 a day in niches; needs your OK; beyond that, an approved provider or Azure credits (USF)
Mining notices and protocols per buyer
No model
You
Keyword rules on public text
Turning a 24-hour notice into a to-do
No model
You
Public text, structured by law
Matching customer files to requirements
No model
You
Filename rules, user tags
OCR of scanned PDFs
No model
You (server CPU)
Local tesseract
КЕП, signer authority, dates, registries
No model
You
EDR, VAT and sanctions open data
Letter templates
No model
You
Fill-in fields
Technical-spec compliance (optional)
Customer-side or Approved paid provider
Customer, or you after approval
Prompt pack for the customer's own AI; never Anthropic API credits
SEO pages, explainers, support replies
Claude Code subscription — operator
You
Content production and operation
Next steps
gantt
title I2 four-week test
dateFormat YYYY-MM-DD
section Build
Pick niches and backtest set :a1, 2026-10-21, 3d
API watcher alerts and 24h rescue :a2, after a1, 6d
Checklist extraction and backtest :a3, after a1, 8d
Bid-pack checks and OCR :a4, after a2, 7d
Web app and monobank checkout :a5, after a4, 5d
section Go to market
Lawyer review of outreach and terms :b1, 2026-10-21, 10d
Why I lost lookup and buyer pages :b2, after a3, 6d
Email pilot then niche scale :b3, after b1, 21d
section Decide
Day 30 go or kill :milestone, m1, 2026-11-20, 0d
Day 60 review :milestone, m2, 2026-12-20, 0d
I do:
Pick 2–3 CPV niches with many FOP bidders and frequent notices.
Build the watcher, checklist, bid-pack check, 24-hour rescue, lookup, buyer pages and web app.
Backtest against historical notices; publish the miss rate.
Run outreach, support and weekly metrics; prepare the 2027 rule set by March.
You do once:
monobank acquiring for your FOP.
A lawyer's review of outreach text, privacy notice, offer and guarantee.
Approve documentation reading on your Claude Code subscription, and the outreach plan.
Be the named sender (the law requires your FOP details in the email).
You do recurring:
Approve guarantee refunds and the rare escalation, about 1 hour a week (estimate).
From day 30, optional calls with partners.
Go/kill thresholds:
Week 2: the checklist backtest finds at least 90% of the items buyers flagged in your niches. Below that, narrow the niches or stop.
Day 30: at least 150 free sign-ups and at least 15 paying (or 40 paid checks), with outreach complaints under 1%. Kill at fewer than 5 paying.
Day 60: at least 50 paying, churn under 15% a month, refunds under 5% of revenue.
Kill at any point if SmartTender or Prozorro ships free buyer-defect history plus file checks.
Risks and open questions
Incumbents: SmartTender could add file checks; Prozorro already auto-checks the EDR and anti-corruption registers for winners.
Outreach legality: if the lawyer says no, GTM falls back to lookup, pages and opt-in alerts, which grow slower.
Guarantee exposure: formal items only, capped at fees; a wrong "green" on a big tender still costs reputation.
Law change on 24 March 2027: tenders started earlier finish under old rules, so you run two rule sets for a while.
Data limits: the overnight sample and keyword taxonomy are indicative; construction is under-sampled; YouControl, Opendatabot and Zakupivli.pro blocked automated access.
Ops: OCR for scans, a certified КЕП library, support drifting into consulting. Keep billing off Telegram (AIN).
Reuse: the same engine transfers to I4 grant applications.
Questions for you:
Which suppliers or niches do you know?
Do you approve reading documentation on your subscription?