Our product

/ Case study / Textile manufacturing AI

AI agents for textile factories that ask a person before they act

Mavick, formerly Safock Textile, is Safock's AI operating system for textile factories. Each department gets its own AI agent that answers questions from that department's data, cites its sources, and stops for a person to approve before it changes anything. Mavick is in live beta at mavick.si.

Client
Mavick
Industry
Textile manufacturing
Stack
AI agentsERP connectorsMCP serversAppend-only activity trail

What it includes

  • Department-scoped AI agents
  • Human approval before writes
  • Answers with citations
  • Live ERP connectors
  • Append-only activity trail
  • Monthly spend ceilings
  • MCP server tools
mavick.si
The Mavick website at mavick.si, presenting an AI operating system for textile factories

/ At a glance

  • 0132Department templatesFrom merchandising to knitting
  • 02One AI agent per department
  • 03Three build pieces per departmentAgent, users, connectors
  • 04Two surfacesDashboard and workspace
On this page
  1. 01The answers exist, but they are scattered and risky to act on
  2. 02One department, one agent, one set of permissions
  3. 03How a factory sets up its AI operating system
  4. 04Connected to the ERP and tools a factory already runs
  5. 05A morning on the dyeing floor
  6. 06Trust rules people can see on screen
  7. 07Where Mavick stands today

01The challenge

The answers exist, but they are scattered and risky to act on

Most questions on a textile floor already have an answer somewhere. It sits in the ERP, a buyer's email thread, a Google Sheet, an SOP, a purchase order or a QC report. Finding it usually means walking over to someone, or messaging them and waiting.

Acting on that answer is riskier. One wrong write into a live ERP or a shared production sheet can stall a shipment, so any AI tool allowed to change those systems carries that risk with it.

Departments also keep their information apart, for good reason. Merchandising has no business reading dyeing's recipes. A chatbot that pools the whole company's data into one index fixes the search problem and creates an access problem in its place.

We set Mavick one job. Give each department a trustworthy answer from its own data, quickly, and let people act on it without breaking anything.

02The solution

One department, one agent, one set of permissions

Mavick models the factory the way it is organised. The owner sets up each department as a separate room. A room holds one AI agent, the people who work in that department, and the systems that department is allowed to use.

Members ask questions in chat. The agent looks up the answer in the department's files and connected systems, and every answer comes with citations, a confidence badge and a trace of the tools it used. When the next step would change something, such as writing a row to a sheet, sending an email or pushing data into the ERP, the run stops and waits.

An admin sees the proposed action frozen on an approval card and decides once. The action then runs exactly once, or it does not run at all. If the admin declines, the member is told plainly that the action was not taken.

03How it works

How a factory sets up its AI operating system

The owner picks departments from a catalog of 32 textile templates, including merchandising, dyeing and processing, quality assurance, stores, compliance, finishing and knitting, or creates a custom department. Each department is built from exactly three pieces.

Approvals, the activity trail and budgets come with every department automatically. The owner never has to remember to switch them on.

  1. 01

    AgentPlain-language instructions, with tools that appear from whatever the department has connected. Publishing creates a new version, and a rollback republishes an earlier one.

  2. 02

    UsersDepartment admins and members, invited by email. A member sees only their own department.

  3. 03

    ConnectorsERP systems, read live and written to only after approval. Email, sheets and chat tools. A knowledgebase of POs, SOPs and certificates. MCP servers as a further source of tools.

04How it works

Connected to the ERP and tools a factory already runs

Mavick works with the systems textile businesses already depend on. On the ERP side it connects to SAP Business One, SAP S/4HANA, Odoo, Oracle NetSuite, Microsoft Dynamics 365, WFX Cloud ERP and Datatex NOW. For everyday communication and records it connects to Gmail, Outlook, Google Sheets, Slack and WhatsApp Business.

ERP access follows one rule. Reads happen live, so an answer reflects the system as it stands now, with a citation showing the source and when it was fetched. Writes happen only after a person approves them. Connectors can be attached for the whole organisation or for a single department.

05How it works

A morning on the dyeing floor

A dyeing operator opens the Dyeing department and asks what is holding up the navy export batch. The agent reads Dyeing's files and the ERP, then replies with the answer and its sources.

Clearing the hold needs a re-dye entry written into a sheet. The agent prepares that write and stops. The department admin opens the approvals inbox, reads the frozen action and approves it. The write runs once, and the operator sees the result in the same chat thread.

Procurement never saw the dyeing recipe at any point in the exchange. The question, the answer and the approval all sit on the activity trail.

06How it works

Trust rules people can see on screen

A factory owner has to see the controls before trusting an agent near production systems, so Mavick puts them in the interface.

Roles are scoped the same way. The factory owner works from the organisation dashboard, department admins see only the departments they run, and members land straight in their department's chat. Removing someone ends their access immediately, and the change goes on the append-only activity trail, which records whether a user, an agent or the system did each thing and can be exported.

  1. 01

    Approvals move through a labelled lifecyclewaiting, approved, executing, done, declined, failed, stalled or expired. Nothing changes silently.

  2. 02

    Answers carry citation pills, a confidence badge and tool traces that are still there after a page reload.

  3. 03

    Every turn is metered against monthly spend ceilings. Mavick warns first, then blocks chat, while waiting approvals can still be decided. A superadmin can put any agent on hold.

  4. 04

    A member who opens another department's URL gets page not found, which reveals nothing about that room. One access.denied entry is written to the trail.

  5. 05

    A scope card in each workspace shows what the agent can do, what it can see and what it cannot, generated from the same grants the server enforces.

07The outcome

Where Mavick stands today

Mavick is in live beta. This case study covers the product and its design, and it reports no customer results.

The design gives a factory specific things. Each department gets answers from its own data, with sources anyone can check. Every change to a live system has a named person behind it and runs once. Owners keep control through approvals, spend ceilings, a hold switch and a full record of who did what.

The same rules shape the AI agents Safock builds for clients in other industries. We scope each agent to the data it needs, cite every answer, and put a person in front of any write to a live system.

/ FAQ

About this project

Which ERP systems does Mavick connect to?

Mavick connects to SAP Business One, SAP S/4HANA, Odoo, Oracle NetSuite, Microsoft Dynamics 365, WFX Cloud ERP and Datatex NOW. It also connects to Gmail, Outlook, Google Sheets, Slack and WhatsApp Business, and accepts MCP servers as a further source of tools. ERP reads happen live, and connectors can be attached for the whole organisation or a single department.

Can a Mavick agent change our ERP data on its own?

No. Mavick reads ERP data live, but every write, whether a sheet row, an email or an ERP update, waits on an approval card for a department admin to decide. An approved action runs exactly once, and a declined action never runs. The member who asked is told either way, and the decision goes on the append-only activity trail.

Can one department see another department's data?

Members see only the departments they have been granted. Opening another department's page returns page not found, and the attempt is written to the activity trail. Each workspace also shows a scope card listing what its agent can and cannot see.

How can we see Mavick working?

Mavick is in live beta at mavick.si. Book a 30-minute call with Safock and we will walk you through the live platform: setting up a department from the 32 templates, asking a question in chat, approving a frozen action in the approvals inbox, and reading the activity trail that records it.

/ Next step

Want Mavick running in your operation?

Tell us what you are working on. You will speak with a founder, and you will leave the call with a clear next step, even if that step is not us.

Prefer writing? Send us a brief and a founder will reply.

WhatsApp usBook a call