Maaz Zindani

Loading the index…

↑↓ move↵ openesc close4072 pages

Help for owners

I want an AI agent to handle internal staff requests

Short answer

An internal operations agent takes requests such as access, leave, purchases or reports, finds the answer in your documents and, for actions, passes them to a person for approval. Give it the least access that works, log every action, and treat anything inside requests and attachments as untrusted.

  • Written for business owners
  • Updated
  • By Maaz Zindani, Senior Software Engineer

Does this sound like your situation?

  • Staff ask the same questions in chat all day.
  • Requests get lost between email and messages.
  • I want it to do things, not only answer.
  • I am worried about it seeing salary or personal data.
  • I cannot tell who approved what.

Why does this happen?

  1. The agent can only follow rules you define

    Which requests it handles, who approves each and what it may change must be written down before building.

  2. Broad access makes mistakes costly

    Anthropic recommends least privilege: do not give the model secrets it does not need and scope permissions as narrowly as possible.

  3. Requests and attachments carry text

    Content from emails, documents and tool results may contain instructions. Anthropic advises treating it as untrusted data.

  4. Actions are tool calls your code runs

    The model asks for a tool and your application executes it, so your code decides what is allowed, not the model.

  5. Answers can be wrong

    Anthropic says models can produce incorrect text, so policies need to come from your documents.

What can I do about it today?

  1. List the top requestsPick three that happen often and are low risk.
  2. Say who approves whatA named person for each type of request.
  3. Keep sensitive data out at firstPayroll, health and legal files wait until access control is proven.
  4. Choose where requests come fromOne channel, such as a form or chat, rather than all of them.
  5. Plan a logEvery request, answer and action should be kept and findable.

When should I bring someone in?

  • The agent will touch payroll, HR or finance systems.
  • You must show an audit trail.
  • Your tools have no safe API to call.

How do I handle it?

  1. Narrow the first scopeOne or two request types with clear rules.
  2. Give read access before writeIt answers first, then prepares actions for approval.
  3. Approve every changeA person confirms anything that modifies data or spends money.
  4. Keep a full logWho asked, what the agent did, who approved.
  5. Extend after reviewNew request types are added only after the earlier ones behave.

The price depends on what I find, so I quote in writing after a short discovery call. Tell me what is happening and I reply within 24 hours.

What do owners ask about this?

Can it approve its own requests?

It should not for anything that changes data or spends money.

What about confidential staff data?

Keep it out until permissions are tested, and check your privacy duties.

Sources

How was this page made?

I drafted this page with AI assistance from the official documents listed under Sources. Found a mistake? Tell me.

Next step

Want this off your desk?

Send me a few lines on where things stand. I reply within 24 hours.