Safety · 3 min read
Safe AI approvals: deciding what waits for a human
Approval rules are not red tape. They are how you let an AI team do real work without lying awake wondering what it did.
By the Makardo Team

The question owners ask most about AI agents is some version of: what if it does something I would never have done? It is the right question. The answer is not to avoid delegating; it is to decide, in advance and in writing, which actions an agent may take on its own and which must wait for a person. That set of decisions is your approval policy, and it is worth an hour of careful thought.
Sort actions by who they touch
A simple way to begin is to sort every action an agent might take by how far its effects reach.
- Internal and reversible: summarising notes, tagging enquiries, drafting documents nobody else sees yet. These are usually safe to allow freely.
- Internal but lasting: changing records, updating your knowledge base, reassigning tasks. Allow with a clear log, and review regularly.
- External: anything a customer, supplier, candidate or the public will see. Emails, posts, quotes, replies. Default these to approval.
- Money and access: payments, refunds, discounts, granting or removing access to systems. Keep these with a human, always.
Write rules an agent can actually follow
Vague rules produce vague behaviour. Instead of saying be careful with refunds, say: any reply that mentions a refund, credit or compensation waits for approval. Instead of saying do not overpromise, say: any message that gives a delivery date, a price or a guarantee waits for approval. Specific triggers are easier for an agent to detect and easier for you to audit later.
Least access, from the start
Approvals catch risky actions; permissions prevent them from being possible at all. Each agent should only see and touch what its job requires. The agent drafting social posts has no reason to read your bank statements. The bookkeeping helper has no reason to publish anything. When something does go wrong, narrow permissions keep the problem small and easy to understand.
One rule is worth stating plainly: an agent should never be able to grant itself more access. Changes to permissions are a human decision. In Makardo, that oversight is built in: anything that sends, posts, publishes or deletes waits in Approvals until you say yes, and only you can change what an agent is allowed to do.
Keep a log you would actually read
A record of what happened, which agent did it and who approved it turns a vague worry into a question you can answer. Make sure the log is written in plain language rather than technical events. Skim it weekly at first. You are looking for surprises: actions you did not expect, approvals you do not remember, or drafts that keep needing the same fix.
Loosen deliberately, not by drift
Over time, some approval rules will feel unnecessary. Perhaps you have approved hundreds of appointment reminders without changing a word. It may be reasonable to let that category send without review. Make that change on purpose, write it down, and note the date. What you want to avoid is the slow drift where approvals are technically required but nobody really reads them.
A starting policy you can adapt
- Nothing external is sent without approval for the first month.
- Money, refunds, discounts and access changes always need a human.
- Any reply that is not answered by the knowledge base pauses rather than guesses.
- Every approval and action is logged in plain language.
- Rules are reviewed monthly, and changes are written down.
Approval rules are what make delegation sustainable. They let you hand over more work with confidence, because you know exactly where the lines are and who holds the pen.
Want to apply this to your business?
Send a request and our team will review it and get back to you about next steps.

