Click anywhere to close
A worker that writes to your ERP is trusted the way a new hire is trusted: with a defined identity, scoped permissions, supervision that lifts in stages, and a record of everything it did. The controls below are enforced by the platform on every action, not by the model's good judgment.
Where the system supports it, every write carries an idempotency key or a transaction identifier. A timeout after a write triggers a status check before any retry. The resulting document state is re-read and recorded; a case is complete only when that re-read matches the intent.
Numerical validation and rule enforcement use explicit logic. Model conclusions are checked against source documents and system state before they become actions. Agreement between two agents is not proof of correctness.
Reverting a procedure version does not reverse a transaction already committed in your system. Correction follows the proper reversal, compensating entry or authorised recovery, with its own audit history.
Workers act through an integration identity your administrators create, with read access to what the remit needs and write access limited to the actions in the service description. Tested outside production and accepted before live execution. Never unrestricted production access.
A worker holds exactly the authority you delegate, the way you delegate it to a person in the role today. Approval requirements, thresholds and reserved decisions are rules the platform checks before any write; a learned preference cannot grant a transaction right. Credentials are held outside any prompt and never appear in a worker's context.
Invoices, remittances, emails and portal messages are treated as evidence. Text inside them that reads as an instruction cannot change a worker's permissions, its rules or its next action.
Your records, procedures and lessons are isolated in your own environment. Nothing is pooled. A lesson learned on your engagement stays on your engagement; general method improvements carry no client data.
Every read, write, approval, intervention and incident is on an audit log you can export. Each case carries a signed evidence pack. Your work records are yours if the engagement ends.
Model selection, processing location, retention and any third-party service are agreed at deployment and written into the service description. For clients in the European Union and the United Kingdom, in-region processing and self-hosted models are available. Humetry trains no models on client data.
A security review precedes any write access. Humetry provides its security policy, the control description for each engagement, the audit-log export, and the insurance and attestation documents a review asks for. Independent attestation of the controls is on the company's launch path and will be stated here on the day it is obtained and nowhere before.
Tell us where the work sits today, the systems it runs in and the volumes. We come back with a scope, a pilot proposal and a date. We reply from a named person.