It is late March 2026, and a trade-compliance manager at an EU steel importer is staring at a deadline she cannot move. By 31 March her company must have filed its application to become an authorised CBAM declarant, or it will lose the legal ability to bring its steel across the EU border later in the year. On her screen is an AI-drafted embedded-emissions figure for a shipment of imported steel. It looks clean. It is also, she suspects, a default value the model has quietly labeled as an actual measured value. If she files it that way, she is not saving time. She is misstating a customs declaration.

The Definitive Phase Is Live

The Carbon Border Adjustment Mechanism, CBAM, is the European Union's tool to put a carbon price on certain imported goods, so that products made in higher-emitting conditions abroad face a cost comparable to goods made under the EU's own carbon pricing. For its first years CBAM ran in a transitional phase: importers reported data but did not pay. That changed. The definitive phase went live on 1 January 2026. The reporting is no longer a dry run. It now carries financial and legal consequences, and the data behind every declaration is something a customs authority can test.

Two dates anchor the early definitive phase. Authorised-declarant applications were due 31 March 2026, the gate you must pass to legally import covered goods going forward. And the first certificate surrender falls in 2027, the point at which importers must hand over CBAM certificates corresponding to the embedded emissions of what they brought in. By early January 2026, more than 12,000 authorisation applications had already been filed, a measure of how many businesses this touches.

It is worth pausing on what the transition from transitional to definitive actually changed in the day-to-day life of a declarant. In the transitional years, the data you submitted was, in a sense, a rehearsal: it built familiarity, it surfaced data-availability problems, but a soft number did not cost you anything directly. The definitive phase removed that cushion. Now the embedded-emissions figure on a declaration is the basis for a real obligation to surrender certificates, and a real cost. The psychological shift this demands is the same one a financial controller makes between a draft management account and an audited statement: the same number, but now it has to be right and it has to be supportable, because someone with authority can examine it and there is money attached. A team that carries transitional-phase habits, rough estimates, undocumented sources, into the definitive phase is carrying a rehearsal mindset into a performance with consequences.

In the transitional phase a wrong CBAM number was a reporting error. In the definitive phase it is a misstated customs declaration with a price attached.

For the AI-aware declarant, this raises the bar on every number a tool produces. A model can help you organize shipment records, draft a methodology note, and structure a declaration at speed, and all of that is genuinely valuable under deadline pressure. But the moment a figure carries a financial obligation and is open to examination by a customs authority, the standard for that figure is not "did the tool produce something plausible." It is "can I prove where this came from." That shift, from plausible to provable, is the through-line of the entire definitive phase, and it is exactly the shift the rest of this lesson equips you to make.

What CBAM Covers, and the 50-Tonne Threshold

CBAM does not cover everything you might import. It targets a defined list of carbon-intensive goods where the risk of carbon leakage is highest. The covered goods are cement, iron and steel, aluminium, fertilisers, hydrogen, and electricity. If you import these into the EU, CBAM is your concern. If you import finished electronics or textiles, it is not, at least not directly.

The phrase doing the work here is carbon leakage: the risk that strict carbon pricing in one place simply pushes production to places with weaker rules, so the emissions move abroad rather than disappear. CBAM exists to neutralize that incentive by putting a comparable carbon cost on the imported version of these high-emitting goods. Understanding the why helps you reason about the what. The covered list is not arbitrary; it is the set of basic, carbon-intensive materials where leakage pressure is strongest. That also explains why embedded emissions, not the finished value of the product, are the unit of account: the regime is pricing the carbon built into the material, regardless of where it was made, so that an EU producer paying for its emissions and an importer of the same material face a comparable cost.

There is also a floor that exempts the smallest importers. A 50-tonne annual de minimis threshold means that an importer bringing in less than 50 tonnes of covered goods per year sits below the line and is exempted from the core obligation. This matters for two reasons. First, it tells you whether you are even in the regime: check your annual import volume of covered goods against the 50-tonne floor before you assume you must comply. Second, like every threshold in this chapter, it is a number to verify against your own customs and procurement records, not a number to guess at or to let a model assert for you.

The threshold also rewards a habit worth naming early: aggregate before you conclude. A company might import 30 tonnes of steel and 25 tonnes of aluminium and reason, wrongly, that each sits below 50 and so it is exempt. The de minimis line is about your covered-goods volume, and a casual single-product glance can produce a false exemption. This is exactly the kind of reasoning a model will get wrong if you feed it one product at a time, and exactly the kind of thing that has to come from a clean view of your actual customs records. The threshold is simple arithmetic, but the arithmetic only protects you if you give it the right, complete inputs.

The Four Concepts You Must Get Right

CBAM has its own vocabulary, and four terms carry almost all the risk. Get these right and most of the regime falls into place, and get any single one of them wrong and an otherwise clean-looking declaration can still be a serious misstatement.

Embedded Emissions

Embedded emissions are the greenhouse gas emissions released during the production of the goods you import, the carbon built into the product before it ever reached the border. They are measured per unit of the good, so the embedded-emissions intensity of a tonne of steel, multiplied by the tonnes you import, drives the obligation. For a tonne of imported steel, the embedded emissions are what the foreign mill emitted to make that steel. This is the core quantity CBAM is built around. Everything else, the certificates, the surrender, the cost, flows from how much embedded emissions your imports carry. Why you care: this is the number, and like every disclosed number in this program, it must trace to evidence, not to a plausible estimate.

Actual Values vs Default Values

You can establish embedded emissions two ways. Actual values are the real, measured emissions data from the specific installation that produced your goods, obtained from your supplier. Default values are fallback figures provided for use when actual data is not available, typically more conservative. The distinction is not cosmetic. Using a default value is legitimate when you genuinely lack actual data, but the declaration has to be honest about which one it is. The cardinal failure mode, the one our trade-compliance manager spotted, is an AI output that takes a default value and presents it as an actual measured value, or vice versa. A customs authority reviewing your declaration can ask which it is and demand the basis. Mislabeling the two is exactly the kind of fabrication-by-relabeling this program exists to prevent.

The Authorised Declarant

Only an authorised CBAM declarant may import covered goods into the EU in the definitive phase. Becoming one is the purpose of that 31 March 2026 application deadline. The authorised declarant is the accountable party: the entity that files the CBAM declaration and answers for its accuracy. The word that matters is accountable. The obligation sits with the declarant, not with any software, model, or consultant that helped produce the numbers. If an AI tool generated a wrong embedded-emissions figure, the authorised declarant still owns the misstatement.

This is the CBAM-specific form of the program's cardinal rule, and it has a sharp edge for anyone tempted to outsource the hard part to a vendor. A reporting platform or an AI assistant can speed your data handling, but it cannot become the accountable party. When the customs authority asks who stands behind a figure, the answer is the authorised declarant, every time. So a declarant who waves through a tool's output without verifying it has not transferred the risk to the tool; they have simply failed to discharge their own accountability. The practical posture this demands is the same one a controller takes toward any number they sign: you may use every tool available to build it, but you must be able to stand behind it as if you had built it by hand, because in the eyes of the regime, you did.

Certificate Surrender

Certificate surrender is the act of handing over CBAM certificates to match the embedded emissions of the goods you imported, effectively paying the carbon cost. The first surrender falls in 2027. Surrender is where the embedded-emissions number turns into money: more embedded emissions means more certificates to surrender means more cost. That direct line from the number to the cost is exactly why an inflated or fabricated emissions figure is not a harmless approximation. It is a financial misstatement in one direction and an under-declaration risk in the other.

The surrender step also closes the loop on why labeling cannot be sloppy in 2026. The figures you declare now are the figures that flow into the 2027 surrender calculation. A mislabeled default that looked harmless on a 2026 declaration does not stay harmless; it propagates into the number of certificates you buy and surrender, and into the cost you report. So the discipline you apply to a declaration today is not bureaucratic box-ticking for its own sake. It is the thing that determines whether your 2027 obligation is correct. An error caught at declaration is a quick fix. The same error discovered at surrender, or by an authority reviewing your surrender, is a correction with a price tag and a credibility cost.

A Worked Example: Default or Actual, Labeled or Laundered

Return to the trade-compliance manager and her imported steel. The AI tool has produced a tidy line: "Embedded emissions: 2.1 tonnes CO2e per tonne of steel." Clean number, confident format. Here is what goes wrong if she files it as is.

She asks the obvious question the model did not answer: is 2.1 an actual value from the producing installation, or a default value? She digs. It turns out the supplier never sent installation-level data. The model, given an incomplete prompt, reached for a plausible figure that happens to align with a default for that product, and presented it without a label. On the declaration, that 2.1 would read as the embedded emissions of her specific steel. It is not. It is an unlabeled default dressed as a measured fact, and in 2027 the certificates she surrenders, and the cost she pays, would be built on it.

Now the disciplined version. She treats the number as unverified until its origin is established. She goes back to the supplier and requests actual installation data. If it arrives, she uses the actual value and records its source. If it does not, she uses the official default value, and she labels it clearly as a default on the declaration, with a note that actual data was unavailable. Either path is defensible, because either path is honest about what the number is. She then uses AI for the genuinely safe and useful work: structuring the declaration, organizing the supporting documents, drafting the methodology note that explains the data source. The AI never decides whether the number is actual or default, because that determination depends on whether real supplier data exists, and only she can verify that.

The lesson generalizes beyond steel. In a CBAM declaration, the most dangerous AI failure is not an obviously wrong number. It is a correct-looking number with the wrong label, a default wearing the clothes of an actual, an estimate wearing the clothes of measured data. The fix is the same discipline that runs through every regulated disclosure: establish the provenance of the number before it ships, and never let the tool collapse the distinction between what was measured and what was assumed.

One reason this failure is so easy to commit is that the two kinds of value can look identical on the page. A default value of 2.1 and an actual value of 2.1 are the same digits; the difference lives entirely in their provenance, in whether real installation data stands behind the figure. A model has no way to know which one it produced unless you have given it the supplier data, and even then it should not be the party that decides the label. The label is a claim about where the number came from, and only a human who can see whether the supplier evidence exists is in a position to make that claim honestly. This is why the right CBAM prompt discipline forces the model to state a source for every figure and to flag, rather than guess, any figure whose origin is not established. A flagged unknown is recoverable. A confidently mislabeled figure is a trap that springs later.

There is a useful parallel to the rest of this program. In the GHG inventory you learn never to let an estimate masquerade as measured activity data; in materiality you learn never to let a model's clustering stand in for a documented judgment; in CBAM you learn never to let a default masquerade as an actual. They are the same rule wearing different clothes. Across every framework, the integrity of a disclosure rests on the honest labeling of where each number came from, and AI's particular danger is its fluency at producing numbers that look authoritative regardless of whether anything supports them. The authorised declarant who internalizes that, and who keeps provenance as a human-verified gate, captures the genuine speed AI offers on the structuring and documentation work without ever letting the tool decide the one thing it must not: whether a number is measured or assumed.

Key Takeaways

• CBAM's definitive phase went live on 1 January 2026. Reporting now carries financial and legal consequences, and a customs authority can test the data behind every declaration.
• Two dates anchor the early phase: authorised-declarant applications were due 31 March 2026, and the first certificate surrender falls in 2027. More than 12,000 applications were filed by early January 2026.
• Covered goods are cement, iron and steel, aluminium, fertilisers, hydrogen, and electricity. A 50-tonne annual de minimis threshold exempts the smallest importers, so check your import volume against it.
• Embedded emissions are the greenhouse gases released producing the imported goods. This is the core quantity, and it must trace to evidence, not to a plausible estimate.
• Actual values are measured installation data from the supplier; default values are fallback figures for when actual data is unavailable. Both are legitimate, but the declaration must be honest about which one it is.
• The authorised declarant is the accountable party. The obligation does not transfer to any software, model, or consultant, so a wrong AI-generated figure is still the declarant's misstatement.
• Certificate surrender turns embedded emissions into cost, so an inflated or fabricated figure is a financial misstatement, not a harmless approximation.
• The most dangerous CBAM AI failure is a correct-looking number with the wrong label, a default dressed as an actual. Establish provenance before filing, and never let the tool collapse measured versus assumed.