Every extraction is logged. Every pipeline run is traced. The fee-code mapper is continuously regression-tested. A verification step self-checks each extraction. And no money-affecting action ships without a human approving it.
Most AI tools ask you to trust a number on a slide. Oracron does the opposite. Every change to the AI that classifies freight charges is automatically regression-tested against a fixed, labelled benchmark before it can ship. If a change would make classification worse, the gate stops it. Accuracy isn't a marketing claim here — it's a release condition.
Benchmark covers fee-code classification — how each charge line is identified and mapped to our standardised code taxonomy. It is one measured stage of the audit, not an end-to-end accuracy figure.
Four built-in checks. Each one already shipped, already running in production.
Every AI decision is recorded with a tamper-resistant audit trail — the model used, the prompt context, the response, and a SHA-256 identity hash. Your auditor can see exactly why any charge was flagged.
Every pipeline run — ingest, mapping, enrichment, rate validation, dispute drafting — is traced as a connected sequence of spans. We can replay any audit decision step by step, with the inputs, the model, and the output for each stage.
The fee-code mapper — the AI step that decides what each line on a carrier invoice represents — is tested continuously against a labelled set of real freight charges. New model versions are gated on this evaluation before they touch your data.
A separate verification step re-reads each extracted invoice and cross-checks it against the source PDF before it reaches your review queue. If the verification disagrees, the extraction is held for review instead of acted on.
The AI proposes. Your team disposes. Money-affecting actions never auto-fire.
Every overcharge finding the engine produces sits in a review queue until a human signs off. Dispute letters are drafted by the system but only leave once your approval policy says they can. The same gate applies to credit-note matching, contract overrides, and any other action that moves money or sends a message to a carrier.
This is also the shape the EU AI Act mandates from 2 August 2026 for high-risk systems (Article 14, human oversight). We built it that way from day one because freight finance teams demand it — the regulatory alignment is a side effect, not a retrofit.
The short version — full detail on the security and data-governance pages.
EU or US — chosen at onboarding. Your data is stored and processed in your selected region.
Your invoices and contracts are never used to train AI for other customers. Anthropic and OpenAI retention ≤ 30 days (abuse monitoring only), then auto-deleted.
TLS 1.3 in transit, AES-256 at rest, tenant-isolated via row-level security in PostgreSQL.
For the full posture see Security & Compliance and AI data governance.
The questions enterprise procurement teams ask before they sign.
We treat accuracy as something to measure, not claim. The AI step that classifies each charge line into our standardised fee-code taxonomy is tested against a fixed, labelled benchmark of real freight-charge cases every time we change it — inside our CI pipeline, before any release reaches customers. A change that would lower accuracy is blocked automatically. Recent runs score 100% on that benchmark (verified June 2026), and because the benchmark runs continuously, that number is checked again on every release rather than frozen on a slide. Note this measures fee-code classification specifically — one stage of the audit — not a single end-to-end "the whole audit is X% correct" figure, which no honest freight-audit tool can claim across every carrier and document type.
No. Your invoices, contracts, and dispute data stay in your tenant and are never used to train models for other customers. We use Anthropic and OpenAI under their no-training enterprise terms; each retains prompt/response logs for at most 30 days for abuse review, then auto-deletes.
Every extracted charge is cross-checked against the source PDF by a separate verification step before it reaches the review queue. If the verification disagrees with the extraction, the line is held for review instead of acted on. On top of that, every AI decision is recorded in an audit log you can replay line by line.
Yes. Every AI call, pipeline run, and dispute action is logged with a SHA-256 identity hash and the exact policy version that produced it. For any flagged charge you can pull the trace, the model used, the inputs, and the human sign-off — the kind of evidence chain SOX, GoBD, and ISO 27001 reviewers expect.
There's always a human in the loop before anything that affects money. Oracron's AI reads, classifies, and audits each shipment and explains every discrepancy it finds — but disputing a charge with a carrier requires a person on your team to review the evidence and approve it. Every AI decision is logged in a tamper-resistant audit trail and traced end-to-end, so findings can be reviewed, explained, and reversed. This design maps directly to the EU AI Act's human-oversight requirement (Article 14, applicable 2 August 2026). To be precise: we describe Oracron as designed for human oversight and Art. 14-ready, not as formally "AI Act certified" — certification is a separate, evolving process. Download the procurement brief.
EU or US — you choose at onboarding. Your tenant data sits in your selected region (AWS Frankfurt for EU customers; AWS us-east for US customers). AI calls go to Anthropic and OpenAI endpoints in the same region. Cross-region transfer never happens by accident; if you need stricter residency contractually, ask us for the DPA on the data-governance page.
The conservative path is the default. Mismatches the audit can't resolve sit as no-benchmark verdicts instead of fabricated findings — disclosed gaps, not invented amounts. The fee-code mapper is regression-tested continuously; the verification self-check catches extraction errors before they reach your queue; and the human-approval gate stops anything wrong from leaving the platform.
We answer enterprise security questionnaires directly. If you're evaluating Oracron for your organisation and need a DPA, an AI-governance write-up, or a walk-through of the audit-log evidence chain, we respond within one business day.
Last updated: June 2026 · Version 1.0 · Privacy policy · Terms of service