Insights
What is an AI agent mandate?
Mark Tack, Founder of RealMandates and author of Marketing Rebuilt · October 31, 2026 · 6 minutes readWhy agents need authority, not just instructions
For most of the history of software, people decided and software executed. A person clicked send, approved the budget, or published the page. AI agents change that. They can notice something, decide what to do about it, and act, sometimes in seconds and sometimes thousands of times a day.
Most of the work we do to prepare agents is about instructions: what the agent should do and how it should do it. Instructions matter, but they answer the wrong question for a leader. When an agent sends an offer to a customer, the question your CEO asks is not "what was it told to do?" It is "who allowed this?"
That is what a mandate answers. In everyday business, a mandate is authority someone grants to act on their behalf. An AI agent mandate applies the same idea to software that acts.
The three questions every mandate answers
- What may the agent handle within its mandate? Some work is low risk and can run without a person: monitoring signals, summarizing research, preparing drafts.
- What needs a human decision? Actions that reach customers, spend money, change records, or speak for the company should go to a person before they happen.
- Who is accountable? Every mandate names an owner: the person who granted the authority and answers for the outcome.
If a mandate cannot answer all three, it is not finished.
Mandates, instructions, and guardrails
These ideas are often blended together. They do different jobs.
Instructions tell an agent what to do. Guardrails limit what an agent can do in general, usually set once in a tool. A mandate grants authority for a specific kind of work, sets how far the agent may go, and names the person accountable. Guardrails are about prevention. Mandates are about permission and ownership.
A mandate in action
Illustrative example
A demand generation agent notices a surge of buying intent from 2,000 target accounts and drafts a time-limited discount offer. Its mandate lets it draft customer-facing offers but not send them. Because the offer reaches customers and affects margin, it goes to the VP of Demand Generation first. She narrows the list to 600 accounts and approves. The record shows the signal that started it, the rule it followed, and her name as the owner.
How to write your first mandate
Start with one agent and one kind of work. Then write down:
- The work: what the agent does, in one sentence.
- The level: observe, recommend, draft, or act.
- The limits: what is in bounds, such as spend ranges, audiences, or channels.
- The decisions: which actions must go to a person first.
- The owner: one named person, not a team or a committee.
- The review date: when you will look at how the agent is performing and decide whether to give it more or less responsibility.
Common mistakes
- Naming a team instead of a person. Teams cannot be accountable; people can.
- Giving every agent the same responsibility. A research agent and a pricing agent should not have the same freedom.
- Writing mandates once and forgetting them. Responsibility should move as trust is earned or lost.
- Treating logs as accountability. A log tells you what happened. A mandate tells you who allowed it.
The bottom line
AI agents will do more of the work every year. The companies that benefit most will be the ones that can always answer two questions: who allowed this, and who answers for it. A mandate is how you answer both before the action, not after.
Frequently asked questions
Written by Mark Tack. Examples labeled illustrative are composites created to explain an idea; they do not describe a specific company.