Varco
From a photo of the goods to a customs code you can defend.
Firmamento Technologies, Genoa, Italy
The system already runs. It sits on a European server, it has been tested on real goods, and it checks every answer against 122,874 binding rulings. When it is not sure, it says so.
Startup track · Ports, logistics and applied AI · varcotrade.com
Customs clears most declarations in minutes. A shipment stops when a field is wrong: it goes to inspection, and the bill can arrive years later. The first field on every entry is the classification code of the goods. Duty, tax and inspection all follow from it.
Hours of a licensed professional for every hard classification. Goods picked for inspection. Products that do not ship until the code is safe. And the general chatbot does not help: it always answers and never gives a reason, and in a dispute that code cannot be defended.
From a photo, or from a description, it works out what the item is, what it is made of and what it is for. Then it asks you to confirm.
It walks down the tariff tree applying the General Rules of Interpretation and the official section and chapter notes, one step at a time, citing the source at each step.
It returns the code with the justification written in the words of the law and the comparison against European precedent: an opinion a professional can put their name on.
In practice: a new catalogue gets classified today, not in 120 days. The goods ship with a code that has an argument behind it. Every case stays on file with its trail. Nothing to install: it opens in the browser, in 35 languages.
It does not invent. It stops, says which piece of information is missing and asks for it. If two headings stay tied, it declares the tie. Confidence is computed on every single case.
On a European server, at TRL 7. In 105 seconds it walks the entire tariff tree, level by level, and writes the legal reasoning for the heading it lands on. The number is about completeness, not speed: that is the whole tree, argued.
The Alpha, in preview to invited customers from the end of 2026, goes all the way down to the declarable code, CN8 and TARIC10 in Europe. US HTSUS depth is the next build, and the CBP corpus is already in the system.
Wired to the engine means the system reads them while it decides, not that it was trained on them once and now recalls them from memory. Every case comes back with the specific rulings it was compared against, by number, so the user can open them.
The 70,930 CBP rulings are the reason a US product is a build, not a rewrite. The first six digits of the Harmonized System are the same on both sides of the Atlantic. What differs is the national depth, and CBP rulings are the corpus that teaches it.
230 cases taken from EU classification regulations, each labelled with the correct code and the legal reasoning next to it. It runs on every release and it drives what we build next.
Every case carries a calibrated confidence. Below the threshold the system abstains, says what is missing and asks. That behaviour is the product, not a safety net bolted on.
No accuracy percentage until development closes. A number measured today would describe a version superseded next month. The full measurement is published with the Alpha, before the first customer, not after.
A system that answers every time demos better. This one occasionally says it cannot decide, which is a worse demo and a better product. We would rather lose the demo.
This is the applied side of research on detecting and preventing hallucination in generative models, which is what our engineer does for a living.
Two years of tariff changes turned the code on the entry into a line the finance team reads. The same item can now carry very different duty depending on a decision that used to go to whoever had time.
Search demand for tariff classification in Italy is up 37% in five years, and across eight European markets it is 10.4 times the Italian volume, 72% of it in English and German.
Transparency obligations for generative systems become applicable. Varco is compliant by design: it already declares what it is, cites what it used and states its confidence.
What we do was not economically possible two years ago. The cost per case now sits inside a subscription a single professional can afford.
Clearing through Savannah, Atlanta's inland ports and Charleston. The person who signs is the owner or the licensed broker of record. They buy because classification is billable expert time they cannot hire fast enough, and because their licence is on the line for every entry.
Home improvement, flooring and textiles in northwest Georgia, apparel, auto parts, consumer electronics. The buyer is the Director or Manager of Trade Compliance, sitting under Supply Chain or Legal.
We are not selling to enterprise procurement in year one. We are selling to the person who personally signs the entry.
108 licensed customs assistance centres in Italy, 26 of them in Liguria. Between 1,500 and 2,000 licensed Italian customs agents. 639 registered representatives in Belgium. 21,112 agents in Poland.
| Who | What they are | Where the gap is |
|---|---|---|
| Avalara, ONESOURCE, Descartes | Trade compliance suites | Bought by enterprises with a compliance department, priced for them |
| Digicust, MIC | Enterprise customs automation | Built around the forwarder's flow, not around the classification decision |
| 3CE, Zonos | Standalone classification | Return a code, not the legal reasoning behind it |
| General chatbots | The real default, used off the books | Always answer, never cite, never abstain |
| Varco | Reasoned classification for the professional who signs | Legal reasoning cited, precedent compared, abstention declared, priced for one professional |
Not us against the suites. Us against the way the work is done today: a senior professional classifying from memory, a junior copying the code from the last similar shipment, or a chatbot nobody will admit to using. That is the incumbent.
No absolute claim of being first. What we claim is a verifiable set that no one covers in full today: reasoning in the words of the law, comparison against binding precedent, declared abstention, multilingual reading using the Court's own method, and a price a single professional can approve without a committee.
| Plan | For whom | Per month |
|---|---|---|
| Professional | One licensed professional | € 99 |
| Company | An import or export desk | € 199 |
| Firm | A brokerage or customs centre | € 349 |
| Enterprise | Volume and integration | By project |
European list prices, cases included with metered overage. The model cost sits inside the subscription and is measured: no unlimited use at a flat price. US pricing is one of the things we want to test in Atlanta.
At these prices, the person who needs the tool is also the person who can approve it. No procurement cycle, no security questionnaire, no annual committee. That is a deliberate design choice, not a discount.
Per-case pricing punishes exactly the behaviour we want: checking a code you are unsure about. The subscription makes checking free at the margin, which is how the product earns its place in the daily routine.
| When | What ships |
|---|---|
| Q4 2026 | Company incorporated. Alpha in preview to selected customers, down to the declarable code, CN8 and TARIC10 |
| Q1 2027 | Pilot with 10 firms in Liguria, run in person |
| Q2 2027 | Production release, with the full measurement completed and published |
| Q3 to Q4 2027 | 25 customers. Belgium, and the goods directory in German and English |
| 2028 | European scale and the enterprise tier |
The dates above are the European plan, and it stands on its own: the pilot is in Liguria because that is where a quarter of Italy's licensed customs centres are, and it is run in person. Nothing below depends on Atlanta going well.
Three US pilots inside the Q4 2026 Alpha instead of an all-European cohort. HTSUS depth wired to the CBP corpus we already hold, brought forward to Q1 2027. A US entity when the third pilot pays, not before.
US customs brokers or trade compliance teams willing to run a 30 day pilot on their own SKUs before the end of 2026. Three, because one is an accident. Paid, because a free pilot tells you nothing about whether the problem is worth money.
Through the Georgia Ports Authority and the Metro Atlanta Chamber, to the people who clear goods through Savannah every day. We would rather be told no by ten of them this week than guess for a year.
Business development with an interest in supply chain. In month one: build the broker list for Georgia, South Carolina and Tennessee, book fifteen calls, and write down the objections verbatim.
The business plan reaches operating break-even in the second year on members' contributions alone. That is deliberate: it means we can say no to the wrong customer, and it means a pilot is a pilot, not a disguised round.
The team is small, and the plan says so: a commercial profile funded from year two, a technical one from year three, critical phases run in sequence rather than in parallel. No US entity, no US reference customer, no revenue yet.
Firmamento Technologies Società Cooperativa, Genoa, founded September 2025, 9 founding members, incubated at the FI.L.S.E. business innovation centre. A cooperative that builds ventures and turns university students into working members. Varco is one of them.