State machines for campaign orchestration: making approvals impossible to skip
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Agents bring flexibility to marketing work. They decide what to do next based on what they observe, and that is exactly what makes them useful for individual tasks. Across a whole campaign, the same flexibility can make work unpredictable when nothing constrains how it moves from one stage to the next.
An agent might publish something that was never approved, or keep optimizing a campaign that a person paused an hour ago. Neither failure comes from a bad agent. Both come from a system that never wrote down which moves are allowed. A simple tool from software design closes that gap: the state machine.
What a state machine is
A state machine is a written model of how something changes over time. It has four parts, and each one answers a question a reviewer would otherwise have to ask.
- States: the conditions something can be in, such as draft or approved.
- Transitions: the moves allowed between states.
- Conditions: what must be true before each transition can happen.
- Actors: who or what may trigger each transition, whether an agent, a person, or a schedule.
For a campaign asset, the states might be draft, in review, changes requested, approved, scheduled, live, paused, and retired. The model says that an asset can move from approved to scheduled, for example, but never from draft to live.
Why it helps with agents
The main benefit is that agents cannot skip steps. If the only path from draft to live passes through approved, and only a person can approve, an agent cannot publish unapproved work, whatever it decides. The rule lives in the system, not in the agent's instructions, so a confused or manipulated agent still cannot break it.
Everyone also knows where things stand. The current state of every asset and campaign is explicit and visible, so the answer to "is this live yet" is a lookup rather than a search through several tools.
Pausing works the way people expect. When a person moves a campaign to paused, every agent sees that state and stops acting on it. Without a shared state, a pause is only as good as the list of agents someone remembered to tell.
Audit trails come almost free. Each transition is an event that can be logged with who triggered it, when, and why. Our post on audit trails covers what legal, brand, and finance reviewers expect that record to contain.
An example for email campaigns
A state model for a single email might have seven states. Here is one version, in the order an email usually moves through them.
- Draft: created by an agent or a person.
- In review: submitted for approval.
- Changes requested: the reviewer sends it back with reasons.
- Approved: the reviewer accepts it, and only people can make this transition.
- Scheduled: a send time is set within approved windows.
- Sent: delivery completed.
- Paused or cancelled: a person stops it at any point before sending.
Agents can move assets between draft, in review, and scheduled. Only people can approve or cancel. That split is the whole point: agents do the routine moves, and people own the moves that commit the brand.
Where the rules live
A state model only helps if it is enforced somewhere agents cannot reach around it. We recommend two layers. The hub tracks state and refuses transitions that break the model.
The tool connectors then check state before acting. A send call for an email that is not in the scheduled state fails at the connector, even if a bug lets it past the hub.
Permissions follow the same lines. The actor field in the model should match the roles in your permission system, which we cover in roles and permissions for a mixed team. If the model says only a person can approve, no agent credential should carry approval rights.
A worked example
Consider a hypothetical retailer running a promotion email to a loyalty segment. A content agent drafts the email and moves it to in review. A reviewer requests a change to the price claim, which moves it to changes requested.
The agent revises it and resubmits, and the reviewer approves it. The scheduling agent then sets a send time for Thursday morning, moving the email to scheduled.
On Wednesday evening, the merchandising team learns the promoted product has a supply problem. A person moves the email to paused, and the scheduling agent sees the paused state on its next check and does nothing. On Thursday morning, the connector receives no send call. If one did arrive, it would fail because the email is not in the scheduled state.
The run log shows every transition with its actor and reason. When the team reviews the incident, the record answers who paused the email, when, and why, without anyone pulling logs from three tools.
Keeping it simple
State machines can grow complicated quickly. Every new state adds transitions, and every transition needs a rule and an owner. We suggest a few habits to keep the model useful.
- Start minimal: model only the states where a decision or a stop matters, usually approval and pause.
- Add states on evidence: add one only when a real case shows the current model cannot express it.
- Name states plainly: use words a reviewer would use, not internal shorthand.
- Review the model quarterly: remove states nobody uses and transitions that never fire.
An overgrown model has its own cost. If reviewers cannot hold the states in their heads, they start working around the system, and the model stops describing what actually happens. Approval states in particular can become queues that slow everything, which our post on approval queues addresses.
What to measure
A state model makes several useful measures straightforward. We recommend tracking these from the start.
- Time in state: how long work sits in each state, especially in review.
- Rejected transitions: attempts to make a move the model does not allow, and which actor made them.
- Rework loops: how often assets cycle between in review and changes requested.
- Pause response time: time from a person pausing work to every agent stopping.
Rejected transitions deserve a close look. A steady trickle from one agent usually means its instructions describe a process the model does not allow, and that mismatch is worth fixing before it turns into a workaround.
Where it fits
In our orchestration run diagram, state transitions mark the boundaries between planning, execution, review, approval, and action. The hub enforces them so individual agents do not have to. We are now studying how state models should work across a whole campaign, where dozens of assets share one parent state. A related question is how a campaign-level pause should cascade to assets already mid-transition.
How to apply this: draw the states your campaign assets move through today, and mark which transitions require a person. If any agent can currently skip an approval state, close that path in the hub and at the tool connectors.
Related research
Orchestrating across channels: one plan, many executors
Customers meet a brand across email, ads, social, sales, and the website. Orchestration keeps each channel's agents telling one story from one shared plan.
Holiday campaign orchestration: a week-by-week agent plan
Holiday campaigns pack channels, offers, and deadlines into a short window. A week-by-week plan shows where agents help and where people stay in charge.
The Agentic Hub pattern: one place to plan, route, and review agent work
As agent programs grow, scattered coordination becomes the main failure point. The Agentic Hub puts planning, routing, context, and approvals in one layer.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
