Skip to content

Instar in Your Organization

This page is for the person who looks after AI use in an organization: a director, an office manager, whoever owns the policy. You don’t need to be technical.

An Instar agent is an AI assistant that lives on a computer your colleague controls and works through their own Claude or ChatGPT account. It remembers what it’s told, keeps track of tasks, and can do work on a schedule. It asks before doing anything risky. Each colleague’s agent belongs to that colleague.

What an organization policy can and can’t do

Section titled “What an organization policy can and can’t do”
  • It can give every agent the same rules. It can make agents ask before certain actions, keep them from sharing certain information, and limit how independently they act.
  • It can’t give you access to your colleagues’ conversations, files or questions. It doesn’t control anyone’s computer. Your real control stays where it is today: access to your own systems (shared drives, email, databases), which you grant and remove as usual.
TodayBeing built
The policyEach agent keeps your policy in its organization rules and reads them at the start of every conversation.A published, versioned policy that agents check for updates.
Giving it to colleaguesEach colleague’s agent is given the same policy text.An invitation link or code that sets it up in one step.
How strict it isThe rules guide the agent. An agent can also check its outgoing messages against the “never” rules, but that check is off by default, and actions, like deleting a shared file, aren’t checked.Rules marked “never” that block reliably, judged by meaning rather than keywords.
Seeing your agentsAsk each colleague.A Work agents page: each agent, its person, its policy version and when it last checked in.

We’ll update this page as each piece arrives. Until then, treat the policy as clear instructions the agent follows, not as a guarantee.

The decision sheet: twelve questions to settle first

Section titled “The decision sheet: twelve questions to settle first”

Settle these with whoever owns policy before setting anything up. This is the real work. The setup itself takes about half an hour.

  1. Which information must never leave the organization?
  2. May an agent write to people outside the organization: always, never, or only with a yes each time?
  3. May an agent speak in the organization’s name?
  4. May an agent spend money? Up to what amount, and with whose yes?
  5. Which AI accounts are acceptable: personal subscriptions, or ones the organization pays for?
  6. May work happen on a personally owned computer?
  7. Should work agents be separate from personal ones? (We recommend yes.)
  8. Which shared tools should work agents be able to use?
  9. Must an agent say it’s an AI when writing to people outside?
  10. What happens to work notes when someone leaves?
  11. Who are the two people who can change the policy, and who answers staff questions?
  12. When will the policy next be reviewed?

Copy this, change what doesn’t fit, and delete what you don’t need.

Purpose. Our agents help staff do their work well: preparing drafts, organizing tasks, researching approved material, and keeping track of commitments. Each agent works for one staff member and within our policy.

Never (the agent will not do these, whatever it is asked):

  1. Share personal information about clients, participants, donors or staff outside the organization.
  2. Speak publicly or sign anything in the organization’s name.
  3. Spend organization money.
  4. Use an AI service or outside tool that is not on our approved list for work material.
  5. Pretend to be a person when writing to someone outside.

Ask first (the agent asks its person each time):

  1. Sending any message to someone outside the organization.
  2. Changing or deleting shared documents.
  3. Accepting meetings or commitments on someone’s behalf.
  4. Using sensitive information for a purpose other than the one it was shared for.
  5. Anything it isn’t sure the policy allows.

Normally (defaults the person can change for their own agent):

  1. Write plainly and briefly.
  2. Keep work notes in the approved shared location.
  3. Leave the person alone outside working hours unless something is urgent.
  4. Say clearly when it is unsure or doesn’t know.
  5. Prefer a useful draft over a question.

When a rule gets in the way, the agent explains which rule applies and who to ask, and offers an alternative that is allowed. A rule that keeps getting in the way is a signal to improve the policy, not to work around it.

Most independence allowed: Balanced (the agent does small, safe things and shows its work; it asks before changing or deleting anything).

Policy owners: [name], [backup name]. Review by: [date].

Fewer rules work better. Every “never” costs some usefulness, so prefer “ask first” where you can.

  1. Set up your own Instar agent first.
  2. Tell it: “Set up our organization’s AI policy,” and paste in your policy (or the starter above, adapted).
  3. Ask it to read the policy back in its own words. Fix anything it got wrong.
  4. Ask it to try twenty everyday requests against the policy and tell you which ones it would refuse or ask about. If it refuses too much, move some “never” rules to “ask first”.
  5. Give each colleague the final policy text. When they set up their agent, they tell it the same thing: “Set up our organization’s AI policy,” and paste it in.

Tell them plainly: what their agent will never do, what it will ask about first, and what the organization can and can’t see. The organization sees nothing of their conversations, files or questions.

Their access to your systems ends the usual way. Ask them to remove the work policy and work tools from their agent, or to retire the work agent. A personal agent, if they have one, stays theirs.

This is a template, not legal advice. Your organization owns its policy.