Our point of view
AI agents need real mandates, not just guardrails
The question is no longer whether AI can do the work. It is who gave it permission, and who answers for the result.
AI has joined the team
For years, AI answered questions and people decided what to do. Now AI agents work alongside people: they send emails, adjust budgets, update records, and talk to customers. That changes the job of leadership. When software acts, someone has to own the action.
The accountability gap
Most companies are adding agents faster than they are adding rules for them. Agents get built in different tools by different teams, and each one decides for itself what it is allowed to do. When an agent makes a costly mistake, the usual answers are "the system did it" or "we will review the logs." Neither is accountability.
You cannot automate accountability.
Instructions, guardrails, and mandates
They are often confused. They do different jobs.
| Instructions | Guardrails | Mandates | |
|---|---|---|---|
| What it does | Tells an agent what to do | Limits what an agent can do | Grants authority to act, within limits |
| Who decides | Whoever wrote the prompt | Whoever configured the tool | A named, accountable person |
| Before a consequential action | Nothing | A general rule | A human decision when it matters |
| After the action | Little or nothing | A log entry | An owner on record |
| Question it answers | What should the agent do? | What must the agent never do? | Who allowed this, and who answers for it? |
What a mandate is
A mandate is the authority a person grants an AI agent to act, within limits that person sets, with their name attached. A good mandate answers three questions before the agent acts: what it may handle within its mandate, what needs a human decision, and who is accountable for the outcome. Read more in What is an AI agent mandate?
The autonomy threshold
Not every action deserves the same freedom. The autonomy threshold sets how far an agent may go for a given kind of work: observe, recommend, draft, or act. Agents earn more responsibility as they prove reliable. High-stakes actions always get a human decision, no matter how confident the agent is.
Measuring what matters: the Governed Action Rate
The Governed Action Rate (GAR) is the percentage of observed consequential AI agent actions within a defined measurement scope that meet the organization's authorization, approval, and named-owner requirements. RealMandates proposes 100% as the operating goal for in-scope actions. The rate must be read alongside governance coverage, because actions the organization cannot observe cannot be assumed to be governed.
Governed Action Rate = In-scope observed consequential actions that meet defined governance requirements ÷ All observed consequential actions in that same scope × 100.
Governance coverage describes how much of the organization's identified consequential agent activity is included in that governed measurement scope. A 100% GAR with limited coverage is not equivalent to 100% enterprise-wide governance. These are proposed management metrics and should be validated against actual system instrumentation before being represented as measured product outputs.
Five principles we build on
- Authority comes before action.
- Every consequential action has a named owner.
- Responsibility is earned, one level at a time.
- High-stakes actions always get a human decision.
- Accountability should make companies faster, not slower.
The belief behind RealMandates
Signals initiate. Agents execute. Humans govern.
Machines are very good at noticing and doing. People are responsible for deciding what should be done and answering for it. RealMandates exists to keep that division of labor intact as AI takes on more of the work.
See it in practice on The Mandate Engine page.