Platform · 06
AI analysis
How the platform finds inconsistencies, weak links and missing evidence in a company's own records, and how Humboldt, the platform's AI analyst, explains them, groups them and suggests what to do next.
What it is for
The records a company keeps can contradict each other in ways nobody notices by hand: a processing step that puts out more coffee than went in, a lot delivered before the harvest date recorded for it, a certificate issued after the goods left, a plot recorded as a point where a polygon is needed. The AI analysis looks for those, every day and after new activity, and shows them on the Findings page before a buyer, an auditor or an authority finds them.
Two layers
- Checks. The platform checks the records against control packs: sets of controls for a regulation, a certification scheme or the company's own policy (EUDR, FSMA 204, certification schemes, company policy). Each control is repeatable, and each pack is versioned, so the same records checked against the same pack always give the same findings. A finding is opened for each fact that is inconsistent with the records or each piece of evidence that is missing, with the records behind it, its severity and what to check.
- Humboldt, the analyst. On request, Humboldt, an AI analyst that runs on Claude by Anthropic, reads the open findings and the records behind them, explains each one in business words, groups findings that share one cause into cases, ranks them and suggests next steps. Every sentence it writes about the records cites them.
The controls decide what is a finding. Humboldt only explains, groups and suggests.
Humboldt groups findings into cases and suggests next steps; people decide. Humboldt never creates anything: a person turns a suggestion into a draft. A person asks for an analysis with Ask Humboldt on the Findings page. Humboldt runs on Claude by Anthropic.
Examples of what the controls look for
| Area | Examples |
|---|---|
| Mass balance | a processing step puts out more than went in, beyond the usual ratio for the product; deliveries waiting add up to more than an account holds; lots made from a supplier's lot claim more than was received of it |
| Timelines | a lot delivered or processed before its harvest date; a lot registered before the supplier lot it comes from was received |
| Certificates | a certificate issued after the lot was delivered, expired at delivery, withdrawn or from an issuer no longer trusted; the same document attached to lots from different origins |
| EUDR and FSMA 204 | a lot of an EUDR product from a site with no geolocation; a plot over 4 hectares recorded as a single point; a lot delivered abroad without a due diligence statement reference; overlapping plots; a lot shipped to the United States without a traceability lot code |
| Network | lots received from a company the company is not connected to; a new supplier delivering far more than the others; a lot named in an open recall that is still held |
| Approvals | a large delivery approved by the person who prepared it; a burst of proposals from one integration; proposals waiting far longer than usual |
Working with findings
Each finding opens with the records behind it, what to check and its history. A person can acknowledge it, resolve it, reopen it or dismiss it with a reason. A dismissed finding stays dismissed while its evidence is unchanged, and comes back if the records change. A finding that is no longer detected is resolved by itself. High-severity findings notify the owners and administrators, and an automation can start from a new finding.
Humboldt's next steps are suggestions. With one click a person turns one into a draft: an evidence request to a supplier, an evidence check, a recall review or a correction. Humboldt never creates or sends anything itself, and it proposes at most three next steps per run.
What it never does
- It never says that a lot, product or company complies or does not comply. Findings say "evidence recorded", "evidence missing" or "inconsistent with the records"; Humboldt's text is checked for verdicts before it is stored.
- It never reads another company's records. It works on the company's own records, under the same access rules as the person who asked.
- An AI check is never evidence. A finding points at records; it does not replace them.
- Plot coordinates are read only inside the platform to find overlaps, and are never stored with a finding.
Further reading
- In the help: Prepare your EUDR evidence, Work with the assistant.
- On this site: Certificates and evidence checks.