Scope 3 Classification at Machine Speed
Claude for the conversation, Jev for the decisions, people for the judgement
Scope 3 is usually the largest part of a company’s footprint and the hardest to measure. The first step is unglamorous: someone has to sort thousands of procurement lines into the right category and match each one to an emission factor. Analysts do it by hand, or companies skip it and accept a rough guess.
We built a tool that does that sorting at machine speed. Claude, Anthropic’s frontier model, is the front end: you talk to it, it explains results and drafts the narrative. Jev, a decision-only model from TypeSafe AI, does the volume: it reads each line and picks a category and a factor from a fixed list, with a calibrated confidence score. Plain code does the arithmetic. People confirm anything the model is unsure about.
How it works
Claude runs the conversation
You work in Claude: hand it a ledger, ask for progress, review exceptions, draft the disclosure text. Claude never reads the 50,000 lines itself. It calls the backend and reads the summary and the exceptions.
Jev makes the decisions
For every line, Jev answers typed questions: which category, is the description specific enough, which emission factor. Each answer comes with a probability, and large factor lists are narrowed in two steps, industry group first, then the factor.
Code does the arithmetic
Currency conversion, price-year adjustment and spend times factor all run in ordinary code. Amounts never go to a model, so there is no risk of a model miscounting a number.
People make the calls that matter
Lines below the confidence threshold go to a review queue, sorted by spend so the most material lines come first. Every model answer and every human override is logged with its probabilities, which gives auditors a trail to follow.
Why it matters, technically
The right-sized model for each job
Most AI deployments send every small decision to a large generative model and parse its prose. This splits the work: a frontier model where reasoning and language matter, a decision model where the answer is one of a known set. It is a pattern that applies well beyond Scope 3.
Answers that can’t be malformed
Jev can only return one of the options it was given, so there is no free text to parse and no invented category. It can still choose the wrong option, which is what the confidence score is for.
Confidence you can route on
Jev is trained to give calibrated probabilities. That turns review into a rule rather than a judgement call: confident answers pass, uncertain ones go to a person. Human time goes where it changes the result.
An audit trail by design
Every decision is stored with the model version, the probabilities and any human override. That is the evidence an assurance provider asks for under CSRD and similar regimes, and it comes as a by-product, not an afterthought.
Low cost, low impact
Jev charges $0.042 per million input tokens and nothing for output, at TypeSafe’s published price. With the full US EPA factor library, we estimate a 50,000-line ledger costs roughly $10 to $15 in model fees. A manual first pass over the same ledger takes an analyst weeks.
The footprint is small too. Jev scores a fixed set of answers in a single pass instead of writing text word by word, so each decision needs a fraction of the computing of a generative call. TypeSafe reports responses 190 times faster and 440 times cheaper than generative models on its own decision workflows; independent figures aren’t published yet. The whole tool runs on serverless infrastructure that costs nothing when idle.
That is the broader point: AI for sustainability doesn’t have to mean large models running all day. Matched to the task, it can be cheap enough to run every month and light enough that its own footprint is a rounding error next to the emissions it measures.
First test
On a 10-line test ledger with known answers, Jev placed all 10 lines in the right category. The one line we made deliberately vague, an invoice number with no description, came back at 76% confidence and went to review instead of being guessed.
What it doesn’t do
- Spend-based factors give screening-level estimates. For the largest suppliers, supplier-specific data should replace them over time.
- Energy purchases are flagged for activity-based accounting (kWh, litres) rather than estimated from spend.
- Jev is new and in early access; we validate it against hand-classified ledgers before relying on it for a client.
- A person signs off on anything that goes into a disclosure.
Run it on your own ledger
Start with an export from your accounting system. We run it, review the exceptions with you, and hand back emissions by category with a full audit trail.
Discuss a pilot