Custom AI workers

What a custom AI worker is

A custom AI worker is a reusable role configured around a business function. It can have its own identity, purpose, knowledge sources, model preferences, skills, communication style, autonomy level, and escalation rules.

Lead-to-role projects

A custom AI worker project still starts with a Lead. The Lead converts into the business role the project needs: CMO for marketing, compliance lead for regulated review, support lead for customer operations, or another accountable owner. Marlow is the Lead-to-CMO example used for an enterprise marketing worker team.

Start with the role

Describe the role in plain language before configuring advanced settings. The best custom AI workers have a clear job, clear boundaries, known inputs, expected outputs, and a human owner.

• What should this worker help with?
• Who will use or review its work?
• What information should it know?
• What tools or actions should it be allowed to use?
• When should it ask for help or approval?

Create flow

A customer should start with a plain-language description, review the generated draft, adjust the role and boundaries, attach approved knowledge, choose model capabilities, bind only necessary tools, set escalation rules, then test the worker before assigning real project work.

Identity and role

Identity explains what the worker is for. It should state the business function, the audience it serves, the kind of work it should accept, and the kind of work it should reject or escalate.

Knowledge sources

Knowledge sources tell the worker what it can rely on. Useful sources can include project files, approved documents, URLs, FileVault paths, templates, policies, or business process notes.

• Attach only approved material.
• Prefer current source documents over stale notes.
• Remove sources that are no longer valid.

Model capability

Model choice should match the job. A research-heavy worker may need stronger reasoning. A visual-review worker may need vision capability. A routine support worker may need lower-cost reliable models and clear escalation rules.

Autonomy and guardrails

Autonomy controls how much the worker can do before asking for approval. Start lower for new workers, regulated workflows, customer-visible actions, access changes, billing, or anything that could be hard to undo.

Tools and skills

Give the worker the smallest useful set of tools. If a tool lets the worker affect a customer, system, file, or external service, make sure the tool is necessary and the approval path is clear.

Settings users should understand

Most users should focus on role, knowledge, model capability, autonomy, tools, communication style, and escalation. The goal is to make the worker useful without giving it unnecessary freedom.

Test before project use

Use the worker test flow before attaching the worker to important project work. A useful test checks whether the worker understands its role, uses the right sources, refuses out-of-scope requests, and escalates at the right time.

Expected result

A good custom worker behaves predictably. It knows its function, stays inside its boundaries, uses the right sources, explains what it did, and escalates when the task needs a human decision.

What can go wrong

Custom AI workers become risky when their role is too broad, their knowledge is stale, their tools are too powerful, or their approval rules are vague. Fix those settings before increasing autonomy.

Read next

Read Models and providers when choosing model capabilities. Read Steering and guardrails when configuring behaviour, approvals, and safe autonomy.