Skip to content

AI Copilot

An assistant that proposes, and people who decide

Ask about stock, vendors and orders in plain language. The copilot prepares drafts and answers from your real data, and waits for a person before anything consequential happens.

Typed tools

It works through tools, not through guesses

The copilot does not read your database directly. It calls typed tools such as stock lookup, vendor price comparison, replenishment suggestions and report generation, and each call runs with the permissions of the person who asked.

  • Every call is checked against the signed-in user's permissions, not just at the start of the conversation
  • The organization always comes from the session. No tool accepts an organization id, so the model cannot pick one
  • Report narratives may only quote figures that are present in the data
  • Supplier documents, order notes and imported files are labelled as untrusted data, and the model is told never to follow instructions inside them
Read the safety model

Three kinds of action

The more consequential, the more human

Every tool declares what it does, and the platform decides how much oversight it needs.

  • Read

    Search products, check stock, inspect movements, compare vendor prices and look up order status. These run immediately.

  • Reversible

    Prepare a draft purchase order, a draft transfer or a draft sales order. A draft changes nothing until someone submits it.

  • Consequential

    Submit a purchase order, adjust stock, buy a label or dispatch a shipment. These wait for an explicit approval.

Approvals

You see exactly what would happen

A consequential action becomes an approval request showing the arguments and the reason. A person with the required permission can edit it, approve it or reject it. Approval is checked again against live data and current permissions when it is given.

  • Edit the arguments before approving
  • Arguments, permissions, policy limits and preconditions are validated again at approval time
  • An approved action runs once, as the approving user, with an idempotency key
  • Requests nobody decides expire
How approvals work

Guardrails

Limits you set, and a record of everything

Controls for the people who sign off on AI in the business.

  • Full audit

    Decisions and executions are written to the audit log alongside everything else that happens in the organization.

  • Budgets and AI credits

    Each plan includes a monthly AI budget and credits. Administrators can lower the organization's budget, and requests stop when it is used.

  • Policy limits

    Set a maximum purchase order value and maximum adjustment quantity or daily value, and larger actions are refused or need approval.

  • Anomalies for review

    Anomalies and recommendations are surfaced in a list for a person to review. They are not acted on automatically.

  • Rate limits

    A cap on copilot messages per minute per user, and a bounded number of tool steps per run.

  • AI imports with confidence bands

    Catalogue extraction and SKU matching show confidence bands, and matches are never confirmed without a person.

Providers

Bring the model you prefer

Each task is routed through a provider abstraction with fallbacks, so you are not tied to one vendor.

  • Anthropic Claude

    The default primary for the copilot and for document extraction.

  • OpenAI

    A fallback for chat, and the provider of the embeddings used to match products.

  • Google Gemini

    A fallback with PDF input.

  • OpenAI-compatible

    Any OpenAI-compatible endpoint, with presets for DeepSeek, OpenRouter and Ollama.

Copilot questions

Let AI do the preparing and people do the deciding

Start a 14-day trial of the Growth plan. No card needed.