PCF launch

CO2 AI knew how to compute an entire company. It needed to compute a product. In 2023, the carbon footprint stops being a communications exercise and becomes accounting, auditor included, and the demand moves to the product: a retailer wants the number for the SKU it puts on the shelf, not the group’s total. The exchange standard comes out in January 2023. CO2 AI knew how to do the corporate footprint, not the product footprint, and leadership gave three months to ship a first product. The opportunity sat in that gap: carbon accounting platforms stopped at the company, and the people who knew how to compute a product were life cycle assessment firms, who do it one SKU at a time. Nobody covered an entire catalog.

This is the hardest subject I have ever had to learn. If a passage puts up a fight, it is not you.

A company’s footprint, and where it stops

My starting ground was the CCFCorporate Carbon Footprint: the total emissions of a company over a year., the carbon footprint of a company: thousands of activity lines converted into emissions, spread over three scopes, for an annual number that can be defended in front of an auditor. Barely charted ground, which Charlotte Degot (CEO of CO2 AI) summed up that same year: one organization in ten actually measures its emissions, one in a hundred manages to reduce them.

It becomes mandatory on both sides of the Atlantic the same year: the European sustainability reporting directive takes effect on January 5, 2023, and on October 7 California signs a law that imposes all three scopes on companies above one billion dollars in revenue, under penalty of $500,000 a year. A verified number is not published, it is justified: where each data point comes from, and how solid it is. But it remains a report, which goes to the regulator and stops there. The manufacturer who receives it learns nothing from it; the split between materials, transport and energy, they already have in their head.

A product’s footprint does the opposite: it leaves. It enters the calculation of whoever buys, and that is where responsibility moves, since a manufacturer inherits the emissions of its suppliers. To act, they need to see where that number comes from, challenge it, correct it, and build a decarbonization plan with them.

The scope says it literally. The sector’s exchange rule uses cradle-to-gateCradle to gate: everything before the product leaves the factory, extraction included, without distribution or use., from the cradle to the factory gate: everything upstream of your plant belongs to your supplier’s number, you add your own, and you pass it to the next link.

footprint passed on =footprint received+my own emissions

footprint received

the supplier's number

my own emissions

what happens in my own plant

Every company computes its share and inherits the rest. When a supplier cannot give its number, its client replaces it with an average, and that approximation travels all the way down the chain. It matters because upstream dominates: according to CDP in 2023, what a manufacturer buys weighs 26× what it emits itself. Most of its footprint therefore depends on numbers it does not compute.

1. Extraction
2. Upstream transport
3. Processing
4. Distribution
5. Use
6. End of life

Gate-to-gatewhat you do yourself

Cradle-to-gatewhat the PCF computed

Cradle-to-gravethe full life cycle assessment

Three months to open a second channel

The PCFProduct Carbon Footprint: the emissions of a single SKU, from the cradle to the factory gate., the footprint of a product, was not one more feature: it was another sales channel, other buyers, other cycles, other competitors. Leadership set the window at three months.

CO2 AI does not come from a garage: the Boston Consulting Group built and incubated it from 2020, and it became a standalone company on September 13, 2023, during my engagement. Its client is not a user who signs up, it is an industrial group that buys after months of discussion, and that already has its spreadsheets, its consultants and its obligations. The software does not replace a tool, it replaces a job.

Three months for software sold to industrial groups means you do not ship a finished product. You ship enough to decide. Three things had to fit in the window:

What we needed	Why
A test product	Show an end-to-end calculation, not a mockup
A data partner	Without a reference database, no product number can be defended
A way to share a PCF	The number is only worth something once it leaves the company that computes it

On paper we were four, in practice a pair with Sophie: the two engineers stayed on CCF tickets, which was running at clients and would not wait. The work started in May 2023 and carried on until December: the three months were the window for the proof, not for the delivery.

The calendar explains the window. On January 26, 2023, the WBCSD publishes version 2.0 of the Pathfinder Framework, the common rule for computing and exchanging a product footprint. The standard arrives before the goods: the corporate footprint, everyone knew how to produce; the number that would circulate inside it, almost nobody.

Reading a trade in three months

A product footprint cannot be designed before understanding what it computes. Four bodies of knowledge had to be absorbed, and none of them was my trade.

The bill of materials. The BOMBill of materials: the list of a product’s components, with their quantities. is the list of what a product is made of. For a yogurt: the milk, the sugar, the fruit preparation, the cup, the lid, the overwrap film, each with its quantity, its process and its origin. Without it, no number. A manufacturer holds thousands of them, in systems that were never built for this.

The regulations. French, European, sector-specific. They set what a number must contain to be enforceable, and the obligation goes down to the product: the European battery regulation requires, since August 17, 2023, a footprint declaration per model and per factory, verified by a notified body. The manufacturer was no longer buying a communications argument, it was buying a way not to be in breach.

The exchanges. Who passes what to whom, in which format, and what each party agrees to reveal: a supplier who hands over its footprint hands over a clue about its recipes and its margins, and that is the real reason these numbers circulate badly. The modeling. How a product breaks down into steps, and where the calculation is allowed to approximate without ceasing to be defensible.

That volume cannot be read by two people inside a three-month window. So Sophie and I had ChatGPT read in our place. In 2023 there was no agent, no tooling, no published method: a conversation, and that was it. The flaw showed up fast. A model of that era makes things up, and on a regulatory text it makes them up credibly: the right tone, the right vocabulary, and it is wrong. So we verified, a lot. Every answer had to cite its source, and we went to read it in the original document before it entered the product. The machine read, it did not decide. Those were the early days, and it became a way of working.

The partner

A product number is computed against a reference database; without one, every calculation is an opinion. The target partner was Quantis, which holds the World Food LCA Database: more than 2,600 datasets, 150 countries. Food and agriculture weigh about 28% of global emissions, which makes it the sector where a product number changes the most decisions.

The work happened during the engagement, the announcement came on July 23, 2024. A data partnership does not get signed inside a three-month window, and that is not what is at stake there: what you do is make the integration obvious enough that the other party wants to build it.

Choosing the right emission factor

A product footprint is a sum: for each line of the bill of materials, a quantity multiplied by an emission factor. The quantity, the client has. The factor has to be fetched, and from one factor to another the same product does not give the same result.

product footprint = Σ (quantity×emission factor)

quantity

known to the client, it sits in their systems

emission factor

to choose: a library, then a degree of precision

Plain yogurt, 125 g0.180 kgCO₂e

kgCO₂edu total

• Materials0.11765 %

  • Raw milk0.09050 %

  • Polypropylene cup0.0158 %

  • Cardboard sleeve0.0095 %

  • Cultures0.0032 %

• Production0.03620 %

• Upstream transport0.02715 %

The factor cannot be invented: you fetch it from a library, public and free like Agribalyse, or by subscription like ecoinvent. CO2 AI carried several, and that was its edge: the client has every factor at hand at the same time, and a model trained to pick the best one for each line.

The emission factor for a kilo of raw milk

Database averagethe same number for every supplier

Regional averagemoves with the country, not with the farm

Supplier's own datamoves when they decarbonize

less precise

more precise

The degree you pick decides the reliability. Poore and Nemecek (the sector’s reference meta-analysis, 38,700 farms across 119 countries) measures up to 50× spread between two producers of the same product. Taking an average when the real spread is of that order is not measuring approximately, it is not measuring.

So the point is not to avoid error, it is to choose. A catalog holds thousands of lines, and nobody matches them by hand: the model fetches the right factor in the right database for each one, and when it has no exact match, it declares it instead of deciding. It intervenes nowhere else in the number itself, which rests on nothing but rules and additions.

✦Matching factors0.0 s

1. ✓Fresh milk→Agrifootprint (6.3 (2022))raw milk at farm including flag
2. ✓Plastic cup→Ecoinvent 3.9polypropylene, granulate
3. ✓Cardboard sleeve→Ecoinvent 3.9corrugated board box
4. ✓Milk cultures→Agrifootprint (6.3 (2022))lactic cultures
5. ≈Printing ink→no exact match, sector average

That choice does not stay implicit: the product measures the precision obtained and displays it, and that indicator is what will later say which line to revisit.

From the bill of materials to the number

A complete product footprint does not rest on the bill of materials alone. The calculation calls for five families of data, and all of them come from the client: their own files, pulled from their own systems, dropped into the product.

What it takes	What it gives
The bill of materials	what the product is made of, and in what quantity
The production volumes	a way to bring a footprint down to the unit sold
The supplier list	who each component comes from
The manufacturing sites	where it is processed, with what electricity
The manufacturing activities	what happens there, and what it consumes

A client rarely brings all five. Two full folders and three empty ones is a common state, and the calculation has to land anyway.

That data does not become a number in one go. It goes down a chain.

1. Model Inputsthe client drops off
2. Allocation Ruleswe fill in
3. Product Activities (LCA)we sort
4. EF Matchingwe connect

bom-2025.csvFresh milk0.115 kg340 kmmaterials× Agribalyse 3.10.090 kgCO₂e

Allocation rules decide what is computable: when a data point is missing, a written rule says what to replace it with, and the calculation lands.

An approximation hidden inside the calculation would be invisible to an audit. Written as a rule, it carries a name, fits in one sentence, and declares the products it touches and their weight. At any moment you know what share of the result comes from a measurement and what share from a convention.

Covering the whole catalog

What remains is deciding for which products, and in what order. That is the decision of the engagement.

The reflex is tempting: keep the products whose data is complete, compute them carefully, and cover 3% of the catalog. A retailer asking for a SKU then has 97 chances out of 100 of hitting a blank. Allocation rules allow the opposite, and the whole catalog at average quality beats 3% at perfection: the first one sells, gets passed on, and improves afterwards.

A more radical path had been proposed, and it deserved to be: starting from the corporate footprint and spreading it across the catalog gives total coverage in a day, with no bill of materials and no supplier.

It was adopted by half, and it is the right half. Spreading works for what is common to the factory, the energy. Not for what tells two products apart, their materials: the local organic yogurt and the one with imported fruit would come out at the same number, and a supplier who decarbonizes would move nothing. The method published with the CDP therefore keeps the spread for energy and requires declaring the rule used.

Hence an unusual screen constraint: every value must carry where it comes from and its precision. An indicator scores it, the Quality Data Level, and under each gauge a recommendation says the next thing to do and what it would gain: the screen does not grade, it points.

Refining without breaking what shipped

The catalog is covered, at average quality. Raising that quality is the next piece of work, and the screen says where: line by line, what rests on a measurement and what rests on a convention. Refining means moving a line from the second to the first, starting with the ones that weigh.

It is not a small thing. An emission factor changes source, vintage, database, and the day it changes it does not change one number, it changes every product that used it. Updating it therefore opens a task, not a form: you walk through what needs review, you validate at the end, and every product touched ships in a new version, with the recap of what moved, the log and the rollback. A single role can launch it.

Sharing closes the loop. Catalog passed on and clients connected, a footprint corrected at the supplier updates theirs without anyone asking again: the data becomes a flow. The channel works both ways; whoever receives a number can challenge it, propose another factor, request the missing data. Without traceability, challenging would amount to saying I do not believe you.

Once the catalog is computed, the question changes: no longer which product emits the most, but which one weighs the most. A low-carbon SKU sold by the millions outweighs a heavy SKU produced in short runs, and that ranking cannot be guessed.

Not every client was there. Volvo Group, whose case CO2 AI publishes, works at the part level: 6 million data points attached to more than 2,000 emission factors, and 1,500 buyers who use them to discuss decarbonization with their suppliers. At that level, a company no longer asks its suppliers for the number, it builds it with them. It is the other end of the same market: a house that far along wants precision, a house just starting wants coverage, and the same product has to serve both on the same day.

The pace of a launch

The window held, at the cost of a permanent trade-off that I settled the same way every time: move rather than consolidate. It is the right call when you are opening a sales channel or not opening it. What I keep from it fits in one rule: that pace is a launch regime, not a cruising regime, and the date you exit the regime gets decided at the same time as the date you enter it.

Three trade-offs stayed written in the roadmap, and I still stand by them. Report personalization, color and logo, is cut and noted as cut. The factor search is flagged as bad rather than worked around: the taxonomy allows neither finding nor choosing, and writing that down beats designing on top of it. And simulation, the most sellable subject of the four, carries a caveat: no design until we have clients to establish the use case.

What remains

The test product became a full product line, Product Footprinting. It reports 25,000 SKUs computed for Reckitt and, at Symrise, 10,000 raw materials across 90 production sites. It declares itself compliant with PACT, TfS, PEF, ISO 14067, ISO 14044 and the GHG Protocol. PACT is the standard whose version 2.0 had just come out when we started. That was the bet of the window, and it held.

Two convictions from that era are still recognizable in it: bill-of-materials lines still attach to factors automatically, and a supplier space brings primary data up into the calculation instead of asking for it again.

The rest happened without me. What I keep from it fits in one sentence from the co-founder and CTO, written when I left and quoted further down: the PCF opened a new trajectory for the product and for the company.

Credits

The launch was done with Sophie-Madeleine Meunier, Lead Product Manager, and Staff Software Engineers Pascal Corpet and Jonathan Wadin.

Kudos