AI at Bext360: it reads the rules, checks the land and the lots, and a person decides

AI reads the rule. Your analyst approves it.

Ethica uses Anthropic’s Claude models to read regulation text and draft each obligation with the passage it came from. Nothing is used until a person approves it.

1 · The text

“…the geolocation of all plots of land where the relevant commodities … were produced…”

EU Deforestation Regulation, Article 9(1)(d)

2 · The obligation, drafted

Give the location of every plot the product came from

Points for small plots, boundaries for plots over 4 ha. Source linked.

3 · Mapped to your data

Plot locationPlot boundary

Collected once, at the farm, in the Origin app. The plot location is asked for by 14 of the 50 regulations in the dictionary.

4 · A person approves

Approved by your analyst

Only now are lots checked against it.

When a rule changes

Add the new text and the same loop runs: read, draft, map, approve. The forms in the field change only where the new rule needs data you do not collect yet.

1 Read

Reads the regulation, label or buyer standard you add.

2 Draft

Drafts each obligation it contains, with the passage it came from.

3 Map

Maps each one to a data point and the step in your chain where it is collected.

4 Approve

Your analyst approves it before any lot is checked against it.

Simulate a rule change in the monitoring showcase

Satellites check the land, in under a second

Every parcel mapped in the Origin app is checked against the 2020 forest map and every loss since, from three independent sources and this year’s Sentinel-2 images. Bext360 runs the check itself.

1 · Forest in 2020

How much of the parcel was forest on 31 December 2020, from the EU’s own 2020 forest map.

2 · Loss since

Any of that forest lost after the cut-off, as three independent sources see it: yearly loss maps, the tropical forest change map, and Sentinel-2 images of the last twelve months compared with 2020, which catch this year’s clearing before the yearly maps do.

3 · Low, medium or high

A fixed rule: loss confirmed by two sources is high and the lot is held; an alert from one source is medium and becomes a field check; nothing is low.

4 · Field check

Medium parcels go to the review queue: field staff take geotagged photos in the Origin app and the reviewer’s decision and reason are recorded.

5 · Evidence card

Per parcel: forest in 2020, loss by source, the 2020 and today images, the data versions and the rule version, sealed on a public ledger.

The sources

EU Global Forest Cover 2020 · 10 m

Forest on the cut-off date. European Commission, Joint Research Centre.
Global Forest Change (tree cover loss) · 30 m

Loss by year since 2021. University of Maryland (Hansen et al.).
Tropical Moist Forests · 30 m

Deforestation and degradation by year, tropics. European Commission, Joint Research Centre.
Sentinel-2 imagery · 10 m

Change since 2020: cloud-free composites of 2020 and of the last twelve months, and the 2020 and today pictures. Copernicus, European Union.

How we test it

54,171

forest clearings that people confirmed on high-resolution imagery, in Peru, Colombia, Bolivia and Indonesia
40,319

cocoa and coffee farms mapped in Ghana and Colombia, which the check should leave alone
1,280

EU reference points for forest in 2020, used only to measure the check, never to tune it
Held out

settings are chosen on one group of regions and scored on others they have never seen
290 of 290

real parcels give the same verdict in our two independent builds of the check, on every release

Run the check on example parcels

Every lot, checked at every handoff

At each handoff the platform checks the lot against the rules you have approved and the weights against what came in. What does not fit is flagged with the reason, for a person to decide.

SID-0412

Sidama → Hamburg

Proof complete

At the roaster · 1,196 kg

Every rule checked. Proof sealed and ready for the roaster’s buyers.

MAR-0108

Marcala → Oslo

On track

At the importer · 1,380 kg

All checks pass so far. Proof builds at every handoff.

SID-0421

Sidama → Hamburg

Data missing

At the exporter · 1,100 kg

Rainforest Alliance certificate expired 30 Sep 2026 Flagged until the data arrives

MAR-0119

Marcala → Oslo

Data missing

At the exporter · 1,035 kg

Missing: export declaration Flagged until the data arrives

MAR-0113

Marcala → Oslo

Flagged for review

At the exporter · 6,100 kg

Volumes don’t add up: 4,800 kg received, 6,100 kg shipped (+27 %, tolerance 3 %) Needs a person: nobody guesses where 1,300 extra kg came from

Example coffee lots on demo data, from the regulation monitoring showcase.

What the AI does not do

It does not file your statements

Bext360 builds the evidence and the due diligence statement; your team files it.

It does not approve its own rules

Every obligation it drafts waits for a person before it is used.

It does not guess

When the weights do not add up, the lot is flagged for a person, never explained away.

It does not grade by camera

Grades are entered on your own forms and checked against the limits you set.

Where the AI works, and where a person decides

AI reads the rule. Your analyst approves it.

One dictionary, so each fact is collected once

When a rule changes

Satellites check the land, in under a second

The sources

How we test it

Every lot, checked at every handoff

What the AI does not do

Bring the rules you answer to.

Tell us about your supply chain

What best describes your role?

Which supply chain are you working with?

Where do you sit in that chain?

What's driving the need right now?

What's the best way to reach you?

Anything you'd like us to cover?

Request sent — thank you!