---
title: "Expert judgement from Peab, scaled to thousands of accounts."
nav: Peab
summary: "Customer case: Peab wanted to understand Scope 3 for its purchases, with thousands of accounts and large volumes of transactions. We built the method and trained a language model on judgements from Peab's experts and the LCA consultancy Miljögiraff, and the model handled the rest. Peab now has an AI tool of its own that can be audited and becomes more reliable over time."
kind: case
status: written
cta: Talk to an expert
ctaTo: booking
heroCta: none
eyebrow: Customer case · Peab
crumb:
  - { label: How we work, href: "/how-we-work#cases" }
facts:
  - { label: Customer, value: "Peab, the Nordic Community Builder" }
  - { label: Together with, value: "Miljögiraff, LCA consultants" }
  - { label: Scope, value: "Scope 3 for purchases, thousands of accounts" }
  - { label: Result, value: "Its own AI tool, which becomes more reliable over time" }
image:
  src: /images/cases/peab-e45-goteborg.webp
  alt: "The E45 at Gullbergsvass in Gothenburg one evening, with the light trails of the traffic between the concrete walls and a row of street lamps down the middle."
  credit: "The E45 in Gothenburg. Photo: Markus Esselmark, Peab press images"
related:
  - /how-we-work
  - /intelligence/in-depth/emission-factors
doc: docs/44-formagorna-i-stallet-for-omradena.md
---

Peab wanted to understand Scope 3 for its purchases, with a method that holds up under external audit.

## The challenge

Thousands of accounts, large volumes of transactions and figures that had to be both reliable and useful in the business. From the start, Peab required the method to hold up under external audit.

## What we did.

We built the method and the model, in close collaboration with Peab's own experts and the LCA consultants at Miljögiraff. They classified a first sample of accounts by hand, and we trained a language model on their judgements. The model learned Peab's purchase descriptions, supplier names and account structure, and then proposed classifications for the rest. Each classification maps to a spend-based emission factor, with a justification grounded in LCA practice.

The collaboration also catches misunderstandings. Early on, the model read the purchases from the workwear company Björnkläder (literally "bear clothes") as fur goods. Expert feedback catches things like this quickly.

Each round of training has five steps:

1. Weighted sample. A small sample of accounts nobody has reviewed, say twenty, is drawn with a probability proportional to each account's estimated climate impact.
2. The experts judge. Miljögiraff and Peab classify the sample by hand.
3. Comparison. Each judgement is compared with the model's proposal.
4. New version. The new judgements are added to the training data, and the model is retrained.
5. Measurement. The difference from the previous version is measured, and the round is repeated until it falls below an agreed threshold.

AI is only as good as the people who train it.

## The result: Peab's own AI tool

:::rows
### Peab's own
The model is trained on Peab's own judgements and account structure. The same method runs on new data every year, and every new judgement makes the next proposal better.

### Auditable
Every step, from the first manual classifications to the result, can be followed in an external audit. Changes in purchasing show up as trends, without the method having to be rebuilt.

### All the small items included
The smaller purchases are too many to classify by hand, but together they add up. The model links them to an emission factor, so that Scope 3 is complete without anyone having to classify every row.

### Detail where it matters
The model points out the largest sources of emissions, so that Peab sees where the biggest reductions can be made and where a proxy factor is enough.
:::

::figure{name="peab-cumulative-emissions"}
Cumulative share of emissions against share of accounts. Each point is an account. The five largest accounts, three percent of all, make up 24 percent of the emissions.

## Next step: real figures where they matter most.

For the largest sources of emissions, no proxy factor is good enough. Setting meaningful targets and tracking them takes the suppliers' own data, year by year. That is what [MASSIV+](https://massivplus.org) is working on: sharing climate data in the value chain, so that proxy factors can be replaced with real figures where it matters most.

## The capabilities behind it

The case rests on three of [our capabilities](/how-we-work#capabilities): the right classification on every row, traceability from total to source document, and AI that proposes while people decide.
