Skip to content

Security principles

This page sets out the five security principles our forward-deployed engineers work to at your site, and what each one means in practice.

Every engineer and every agent gets the access the work needs and no more.

  • The Discovery Sprint runs on read access to the ERP data.
  • Write access is added one workflow at a time, when that workflow is ready to write back. It covers only the transactions that workflow creates.
  • An agent that drafts sales orders can read the item master, the price list and the customer’s credit limit. It cannot see payroll or change a supplier’s bank details, because its account does not reach them.
  • Access ends when the need ends: when an engineer leaves your engagement, when an agent is retired, when a stage ends.

Accounts in your own systems are created and removed by your IT team, under your own password and sign-in rules.

  • Each engineer has a named account. Each agent has its own service account.
  • There are no shared logins, and our engineers never use an employee’s login.
  • You can disable any of these accounts at any time, without asking us.
  • Who holds the AI model provider account and the hosting account during the engagement is agreed in your contract. If either is ours, it moves to you at handover.

Access and credentials covers requests and offboarding.

  • Every agent action is recorded in the audit trail with its inputs, its decision and its result. Audit trail and monitoring covers how to review one.
  • Each engineer works under their own name, so your systems’ own logs show which person did what.
  • Every release is recorded with what changed and who signed it off.
  • A workflow cannot move up a level unless every action is logged and reversible. That is one of the four checks in a promotion review.
  • Your ERP stays the system of record. We read from it and write to it only through interfaces it supports.
  • You own the custom code and the data built for you.
  • Where the software and its data are hosted is agreed in your contract.
  • An AI model receives only the data a workflow step needs. AI providers and your data covers what is sent and to whom.
  • After exit we keep none of your data. Data ownership and portability covers exports.

An agent works inside limits that people set. It cannot change them.

  • Every agent has limits: amounts, customer and supplier lists, time windows. A case outside them is an exception and goes to a person.
  • Limits are set with the workflow owner and change only with the workflow owner’s sign-off. Our engineers build them in.
  • The stop switch halts an agent at once, the way an emergency stop halts a machine.
  • A workflow moves up a level only after a promotion review, with the workflow owner’s sign-off. Moving it down needs no review: the workflow owner decides alone.

Guardrails and limits covers limits, permissions and the stop switch in detail.

This page describes how we work. It is not a contract term, and it leaves three things out on purpose:

  • Security terms. They are agreed in each contract. If your security team has requirements or a questionnaire, raise them before the Discovery Sprint, because that is when our engineers first get read access to your ERP data.
  • Hosting location. It is agreed in each contract.
  • Support response times. They are agreed in each contract.

Stages and terms has the full list of what your contract records.

We claim compliance with no standard. What you can check for yourself is on this page: the accounts you created, the logs in your own systems and the audit trail.