Builder Créatif × BAT Romania · Rothmans

Response to the “Online monitoring tool” brief — §3 to §8

The brief asks for two sweeps a week.
I ran one before writing this page.

Not a demo, not a mockup: the actual sweep, on your §5 keywords, on 2026-07-25, in 12 minutes of machine time. 5 cases came out of it — including a seller who states himself that his cartons came from duty free, at 63 % below the maximum retail price. They are all below, with their URLs. Check them.

  • Urgent alert in 11 min, not 24 h
  • All four §7 deliverables, generated
  • Read-only by construction — §5 respected
  • The repeat seller finally detected

A real alert, exactly as it lands — condensed into a few seconds.

↓ The 5 cases found, one by one

The real problem

Your problem is not finding.
It's never stopping the search.

Anyone can find one isolated case. That is not what the brief asks for. It asks for a continuous presence across dozens of channels, in a language full of variants, with compliant evidence capture every time, and an alert within 24 hours. This is surveillance work, not investigation — and surveillance is exactly what humans do badly: they tire, they skip a channel on Friday, they forget the misspelling that returns the most results.

What the brief writes

  • “at least twice per week”
  • 10 listed keywords, “and equivalent misspellings, abbreviations and local slang
  • 8 families of channels in §3
  • Urgent alert “within 24 hours”
  • “repeated seller activity across multiple platforms”

What a team actually delivers

  • Two passes, Monday and Thursday, if the week is quiet
  • The 10 keywords. The variants, when there's time left
  • The 3 or 4 channels that return results
  • 24 hours — meaning the next morning
  • Nobody cross-matches 12 platforms by hand. Nobody.

The gap between those two columns is not a diligence problem. It is a cadence problem — and cadence can't be motivated, it has to be automated.

The 2026-07-25 sweep

I didn't describe the machine.
I pointed it at your brand.

8 platforms attempted, 12 minutes, zero human input. Here is what came out. Every case carries its URL: open them while we talk. Seller identifiers are deliberately absent from this page — they belong in the delivered register, not on a web page.

  • 8platforms attempted
  • 12 minof machine time, end to end
  • 5cases retained
  • 71 %noise discarded on OLX
F-01 eMAG.ro General marketplace High

Cartus Rothmans Blue

325 lei · in stock at capture time · 325 lei / carton

  • Full carton (10 packs) orderable online — “Adaugă în coș” button active, availability schema.org/InStock
  • No age check encountered before adding to cart
  • The country's largest e-commerce platform — maximum public visibility
  • Price at the reference (325 lei = 10 × 32.50): no non-duty-paid indicator here

What the machine does not settle Distance sale of a tobacco product to a consumer in Romania: for Legal to qualify (seller, status, applicable regime). The machine records the offer, it does not state the law.

www.emag.ro/cartus-rothmans-blue-5948530000105/pd/DGKF0HMBM/… ↗
F-02 Luzan.ro Dedicated online tobacco shop High

Page de marque ROTHMANS — 4 références

32,50 lei (King Size) · 24,00 lei (Classic 100S ×3) · permanent catalogue

  • Fully tobacco-dedicated online shop: cart, customer account, delivery, a “Luzan Business” section
  • Rothmans has its own brand page and promotional copy (“eleganță și satisfacție”, “rafinament”) — advertising-type content per §4
  • Rothmans King Size flagged orderable (can_be_ordered: true) at capture time
  • 3 Classic 100S SKUs shown at 24.00 lei — 26 % below the 32.50 lei reference

What the machine does not settle The 24.00 lei may be an outdated maximum price left on the product sheet, not a discount. Exactly the kind of gap a human must settle — and exactly the kind a human never spots, because it means comparing 4 product sheets to a reference on every pass.

www.luzan.ro/brand/rothmans… ↗
F-03 OLX.ro Classified ads High

Rothmans international 3 cartoane

800 lei · posted 16 July 2026 · 267 lei / carton — 18 % below reference

  • Bulk sale — 3 cartons at once, i.e. 30 packs, in a peer-to-peer listing
  • 267 lei per carton against the 325 lei reference: −18 %
  • “international” in the title — origin claim to verify against the pack photos
  • Recent listing: 9 days old at capture time

What the machine does not settle Bulk volume + below-reference price = the profile §8 rates High. Pack verification from photos (language, health warning, fiscal stamp) required before escalation.

www.olx.ro/d/oferta/rothmans-international-3-cartoane-IDkM29J.… ↗
F-04 OLX.ro Classified ads Medium

Țigări Rothmans de colecție

100 lei · posted 29 June 2026 · Sulina

  • de colecție” wording — the classic workaround to slip tobacco into an allowed category
  • Location shown: Sulina (Danube delta, border area)
  • Still online 26 days after posting

What the machine does not settle Genuine collectible (empty, vintage pack) or tobacco sold under a collectible label? Settled from the photo. This is a Medium case: log it, watch for seller recurrence, don't escalate.

www.olx.ro/d/oferta/tigari-rothmans-de-colectie-IDkHWSY.html… ↗
F-05 craiovaforum.ro Regional forum High

Tigari ieftine - Rothmans

120 lei / cartouche · posted 2019 · still online and indexed · 120 lei / carton — 63 % below reference

  • The seller writes it himself: “tigarile sunt luate din aeroport, din duty free” — non-duty-paid origin stated outright
  • 120 lei per carton against the 325 lei reference: −63 %
  • Tags set by the seller: tigari duty free, tigari ieftine, rothmans
  • Photos attached to the post · account registered since 2006 · thread now closed

What the machine does not settle The thread is closed and the offer is seven years old: this is not an active sale, it is residual visibility. It still surfaces on “Rothmans ieftin” searches in 2026. Status “historical, still visible” — the call (takedown request?) belongs to the client.

www.craiovaforum.ro/showthread.php?t=1768373… ↗

And above all: what you didn't have to read

On OLX, “rothmans” returns 7 results. 2 are about tobacco. Here are the other 5 — discarded by the machine before a human spent a second on them. That is 71 % of the reading work gone, on one channel, on one keyword.

  • Vintage Seiko F1 Rothmans watch — 300 lei
  • 1/43 Porsche 956 Rothmans Le Mans 1983 model car — 100 lei
  • vintage seiko casio rothmans f1 watches — 700 lei
  • Cigarette tubes, Rothmans carbon filters — 80 lei
  • 1:43 Le Mans 24h model car collection — 47 lei

The station nobody delivers

Twelve isolated listings,
or one seller doing it again.

§4 explicitly asks you to detect “repeated seller activity across multiple platforms, phone numbers, usernames or groups”. It is the line of the brief that never gets honoured, because it means comparing everything with everything, continuously. A machine does it without thinking: it matches numbers, image fingerprints, turns of phrase and prices, across every platform at the same time.

The number

The same phone number behind two handles on two platforms: the simplest and most profitable match there is. A human only finds it by accident.

The image fingerprint

The same pack photo re-posted three weeks later under another account. Matching is done on the image fingerprint, not the file — a crop is not enough to break the link.

The phrasing

“din duty free”, “luate din aeroport”, a recurring misspelling: a seller's turns of phrase are as reliable a signature as a username, and far harder to change.

In practice this changes the nature of the deliverable: the weekly report stops being a list of listings and becomes a list of sellers, each with a history. That is the difference between reporting ten listings and shutting down a network.

The lever

Cost stops tracking
the breadth of surveillance.

The sliders start on the literal scope of your brief: 10 keywords, 8 channels, 2 sweeps a week. At that setting I am not cheaper than a team — the chart shows it without cheating, and I'd rather you saw it immediately. Now push the sliders towards what §5 actually asks for: the variants, the abbreviations, the slang, every day.

160 passes per week 1 pass = 1 keyword × 1 channel × 1 sweep

€0k €20k €40k €60k €80k brief scope €1,248 / mo €434 / mo 0 10,080 passes / week
  • Human team — 25 passes/h, €45/h loaded
  • Agent — €420/mo infrastructure + €0.02 per pass

The assumptions are bent against me, on purpose. The analyst handles 25 passes an hour — including evidence capture, which is optimistic. The hourly rate used is €45, the low end of agency pricing. And the machine cost already carries a safety margin: I'd rather quote slightly above reality than have that conversation in six months.

How one case runs

Seven stations,
only one is human.

I am not selling “less human”. I am selling the human moved to the one place where judgement has value: deciding whether a case goes to Legal. Everything upstream is copying, capturing and comparing — machine work currently done by people.

  1. 01 Sweep machine

    The §5 keyword list plus variants, misspellings and slang, across every §3 channel.

  2. 02 Noise filter machine

    F1 watches, Porsche models, cigarette tubes: discarded before anyone reads them.

  3. 03 Evidence machine

    Timestamped full-page capture, URL, archive. All of §6, on every single case.

  4. 04 Pack reading machine

    Language, health warning, fiscal stamp, design consistency.

  5. 05 Seller graph machine

    Same number, same photo, same handle — across 12 platforms at once.

  6. 06 Qualification human

    Counterfeit? Non-duty-paid? Youth access? The machine surfaces indicators, you decide.

  7. 07 Deliverables machine

    Excel register, weekly report, urgent alert, monthly deck. Generated, not written.

↺ every case a human settles flows back into the machine: rejection reasons sharpen the noise filter, escalation reasons sharpen the risk matrix.

The limits, before you ask

§5 sets four prohibitions.
Three of them are capabilities the agent doesn't have.

A human team promises to respect these rules. An agent cannot break them: the corresponding functions do not exist in the system. It is verifiable in the code, and it is the only compliance argument I find honest to make.

No purchasing

No payment method is connected to the agent. There is no purchase function.

No seller interaction

No account, no messaging, no contact form. Read-only.

No impersonating minors

No profile is created anywhere. The agent reads what is public and unauthenticated.

No encouraging illegal activity

The only one of the four that remains a design question rather than a capability one — it rests on nothing ever being published or re-published. Reviewed at acceptance.

Personal data

Numbers, handles, profiles: this is processing. It needs its legal basis, retention period, processing record and a data processing agreement. All three documents ship with the machine.

Social platforms

Facebook groups, private messaging: no defeating of protections. Access through a licensed data provider, priced separately, decided with you.

How it runs

Four weeks,
one hour of your time per week.

Every week ships something you can look at — not a status update. The weekly meeting is not there for you to track me, it is there to extract your judgement: what deserves escalation, what is only noise, what we refuse to put in a report.

  1. Week 1 Collection

    The channels that don't fight back.

    Classifieds, marketplaces, dedicated shops, forums, search results. These are the channels where the evidence is cleanest and the blocking weakest. By the end of the week the sweep runs on its own twice a day and you get the first register.

    • OLX · Publi24 · Okazii · eMAG · tobacco shops · forums · SERP
    • The §6 Excel register, filled in
    • The keyword list expanded to variants and slang
  2. ↓ I build — you validate
  3. Week 2 Judgement

    The machine learns to read a pack.

    Vision over listing photos: pack language, presence and rotation of the Romanian health warning, fiscal stamp visible or not, suspicious design. Then the §8 risk matrix, with a written rationale for every case.

    • Low / Medium / High / Urgent classification
    • One rationale sentence per case
    • The urgent-alert threshold calibrated with you
  4. ↓ I build — you validate
  5. Week 3 Recurrence

    The seller who comes back under another name.

    This is the station nobody runs by hand: cross-matching numbers, handles, image fingerprints and phrasing across every tracked platform, turning twelve isolated listings into one seller. §4 asks for it explicitly. It is never delivered.

    • Cross-platform seller graph
    • Re-posting detection (same listing, new account)
    • Seller typologies for the monthly deck
  6. ↓ I build — you validate
  7. Week 4 Delivery

    The four §7 deliverables, generated.

    Weekly report, evidence log, urgent alert, monthly deck. Plus the review screen: a queue of High and Urgent cases, each settled in thirty seconds by a human. That is all the manual work that remains.

    • The 4 deliverables, in the expected format
    • The human review screen
    • The escalation procedure to Legal & AIT

The price

It's written here,
because a page without a price can't be decided on.

Two lines, no tiers, no quote to wait for. The setup is real work: collectors, pack reading, seller graph, deliverable generators, review screen. The monthly covers infrastructure, operation and evolution.

Setup · 4 weeks €8,900 fixed price, once
  • Open channels in production: classifieds, marketplaces, shops, forums, SERP, public Telegram
  • Pack reading, §8 risk matrix, seller graph
  • The four §7 deliverables + the human review screen
  • Legal basis, retention period, data processing agreement
  • Everything stays yours if we stop
Then · per month €1,490 infrastructure included, no lock-in
  • The machine sweeps, captures, classifies and reports
  • Keywords and channels uncapped — add as many as you want
  • One new collector per month included (new platform, new format)
  • A monthly session on the borderline cases
  • One instruction pauses everything

What that gives over twelve months. €26,780 in year one, setup included. Against that, the same coverage held by humans — 24 keywords, 10 channels, 7 sweeps a week — costs €157,248 a year on the assumptions most favourable to the human side, and still doesn't produce the seller graph. The ratio isn't a few percent: it's 5.9× cheaper, for surveillance 11× broader than the brief's own minimum.

And the reverse, stated just as plainly: at the strict scope of the brief, the agent costs roughly what a team costs. If the goal is to tick §5 at the minimum, keep the agency. If the goal is to cover what §3 actually describes, no agency can keep up, at any budget.

The objections

The questions you'll ask me
in two hours.

Might as well answer them now — it leaves time for the real ones.

So an AI decides on its own that a pack is counterfeit?

No, and this page claims that nowhere. The machine records visible indicators — foreign-language pack, missing warning, no visible stamp, a price 63 % below the maximum retail price — and presents them together. Qualification stays human: that is station 06, the only one of the seven that is not automated. What you gain is not automatic judgement, it is only ever having to judge cases that are already fully documented.

Facebook, Instagram, TikTok — are you going to scrape private groups?

No. And this is the genuinely hard part of the brief, so let's put it on the table now: §3 includes social media and messaging channels, yet there is no API for groups, and defeating anti-bot protection is both against the platforms' terms and impossible to defend for a client your size. My answer: open channels (classifieds, marketplaces, shops, forums, SERP, public Telegram channels via the official API) at a fixed price, and social through a licensed data provider — a separate cost line, decided with you, not a box quietly ticked.

We'd be collecting phone numbers and profiles. What about GDPR?

This is unambiguously personal data processing. It needs a legal basis (legitimate interest, documented), a retention period, a record of processing, and a data processing agreement between us. I deliver those three documents with the machine — and this very page is the demonstration: no seller identifier appears on it, while the delivered register contains all of them, exactly as §6 requires.

§5 forbids buying, entrapment, impersonating minors and engaging sellers. How do you guarantee that?

Better than a human team can, and this is the argument I like most: the agent is read-only by construction. It has no seller account, no payment method, no message-sending function at all. This is not a rule in a handbook someone might forget on a Friday evening — it is a capability that does not exist in the system. An audit can verify it in the code.

What happens when a platform changes its HTML or blocks us?

It happens, and it is planned for: three of the eight platforms in the sweep below refused the first request. Collection runs in a real browser, every collector has a health check, and a silent platform raises an alarm instead of quietly vanishing from the report. The worst failure mode for a monitoring system is not breaking down — it is returning an empty report that looks like a reassuring one.

Why not simply the agency, which already does this?

On the literal scope of the brief — 10 keywords, 2 passes a week — I am not cheaper, and the simulator above shows it without cheating. The gap opens on what §5 actually asks for: “equivalent misspellings, abbreviations and local slang”, across every §3 channel. Nobody does that by hand twice a week. The machine does it twice a day for roughly the same money, and the urgent alert goes out in eleven minutes instead of twenty-four hours.

And if we stop after six months?

Everything stays with you: the infrastructure runs on your accounts, the historical register and the captures are yours, the code is delivered and readable. No closed platform subscription, no data held hostage. The monthly covers operation and evolution, not a right of access.

Next

Give me one week.
I'll hand you the §7 weekly report, for real.

Not a mockup: the expected deliverable, on your keywords, with the cases found that week, the Excel register filled in and the timestamped captures attached. You compare it to the one you receive today. If it adds nothing, we stop there and it cost you nothing.

Open question for the call: over the last twelve months, how many reported sellers came back under another account — and how did you find out?