AI marketing automation with approval: the gate is the design.

The unit of delegation is prepared work, not a sent message.

Four verbs, in a fixed order: you direct, it prepares, you review, it releases. Every difficult question in this area turns out to be a question about where the third verb goes — and almost every failure is somebody quietly removing it.

That order matters more than it looks. Asking a general assistant for a launch produces a good plan and a bad afternoon: you then copy the page into one system, the emails into another, the event into a third and the automation into a fourth. The plan was never the expensive part. The copying was, and so was reconstructing the context at every boundary — which customers bought before, who is showing intent now, what is already live.

An operator is different in one specific way: it prepares the work where the work lives, using the business knowledge and customer context it has been permitted to use, and it stops in front of you with something you can inspect. Not text you have to move. Not an action already taken. A prepared thing, waiting.

Agentic does not mean unsupervised

The interesting design question is not how much an agent is capable of doing. It is which actions require a human decision, and how that human is shown enough to make it quickly. Get that pair right and you can delegate far more than you would dare to otherwise, because the cost of being wrong is bounded. Get it wrong in either direction — gates everywhere, or gates nowhere — and you end up doing the work yourself either way.

Repeatable, reversible, context-rich — and checkable in ten minutes.

The first three are the same tests that decide whether a task should be automated at all. The fourth is specific to delegation, and it is the one people discover the expensive way.

1. 01Repeatable. You have done it by hand at least twice and the steps did not change between them. A job you have never done once yourself cannot be delegated, to a person or to an agent, because you have no way of recognising a good version of it.
2. 02Reversible. If the output is wrong, you can undo it. A draft is entirely reversible. A configuration change you can switch back is nearly so. A send to your whole list is not reversible at any price, which is why that is the line the gate sits on rather than a step you speed up.
3. 03Context-rich. Everything a competent new colleague would need is written down somewhere the agent is permitted to read — the offer, the terms, the voice, what you already sell, what you tried last time. If you would have to explain it in a paragraph every time you asked, write that paragraph into your knowledge sources first. That is not preparation for the work; it is most of the work.
4. 04Checkable in about ten minutes. You can tell whether the output is right quickly, without rebuilding it to find out. If reviewing costs more than doing, you have not delegated anything — you have created a second job and given yourself the duller half.

Notice that the third test is the only one that generates work for you. Writing down the offer, the terms, the voice and the history is the part that makes everything afterwards possible, and it is the part people skip because it does not feel like automation. An agent working from thin context produces confident, structurally perfect, entirely generic work — and you will spend longer fixing that than you would have spent writing the source material once.

The right-hand column is not permanent. Most of it moves left once the source material exists, the job has been done by hand a few times, and you have watched a gated version run. What does not move is anything you would put in a contract.

A gate is a place work stops, not a message saying it happened.

Treat the gate as a design pattern rather than a safety feature bolted on afterwards. You are deciding, once, which classes of action a person must authorise — and then building everything else to run freely inside that boundary.

What stays gated

1. 01Customer-facing. Anything a person outside your business will read, receive or land on. The gate is not about the quality of the writing; it is about the fact that you cannot recall it.
2. 02Consequential. Spends money, changes a price, grants or removes access, deletes something, or is materially hard to reverse. The test is not how complicated the action is — it is what it costs to be wrong.
3. 03At scale. The same action against twelve records and against twelve thousand are different risks wearing the same name. Blast radius belongs in the gating decision, and it is the one people leave out.
4. 04Irreplaceable. Anything touching a record you could not reconstruct — historical customer data, the audit trail, the one document nobody has a copy of.

And what does not: work that only produces a draft, analysis and summaries, and changes scoped to a single record you can undo in a click. Gating those feels responsible and is actively harmful, because every unnecessary gate spends the attention you need for the ones that matter.

Four properties of a gate that is real

• →The reviewer sees the change, not a description of it. A summary saying “prepared five emails and a segment” is a notification. The gate is only real when you can read what will actually go out, to whom, and when.
• →It has a named owner and lands on their list. A gate in a system nobody opens is a delay, not a control. Somebody specific has to be responsible for clearing it, and it has to appear where they already look.
• →The default is not-approved. A gate that expires into “send” after 48 hours is a countdown with a polite name. When nobody responds, the correct outcome is that nothing happens.
• →It leaves a record of who approved what. Not for compliance theatre — for the conversation you will have with yourself in three months when you are trying to work out how something went out. Confirm what your workspace actually retains before you depend on it.

Check the inputs before you read the output.

If the brief was wrong or the source material was stale, no amount of careful reading will rescue the result — and checking takes about ninety seconds. What did it use? Is that current? Is that the audience I meant? Then read.

After that, the review is not a general act of scrutiny. It is a short routine per kind of artifact, and the routines are short on purpose: a review you can complete in ten minutes is a review that will actually happen every time, which is worth far more than a thorough one that happens twice.

One rule sits above all three. If a review is taking substantially longer than ten minutes, the problem is the size of the job rather than your thoroughness. Ask for it in pieces — the segment first, then the sequence, then the schedule — because a prepared artifact you cannot hold in your head is one you will eventually approve on vibes.

Know what the agent can spend before you approve the work.

Delegation has a budget as well as a boundary. In Agorix that budget is AI Credits, which measure eligible Operator work — preparing funnels, pages, emails, segments, workflows, course material and audits — and each paid plan includes a monthly allowance.

Those are the current published allowances; the pricing page is the authority on them, along with credit packs for months that need more than the plan carries.

Four habits that make the budget predictable

1. 01Treat the monthly allowance as this month’s work budget. Not as a background utility you never think about. Included credits reset each month, so an allowance you did not use is an allowance you gave back — and one you overspend in week two leaves the rest of the month without capacity for the small, useful jobs.
2. 02Sequence the large jobs deliberately. The published example puts a full course build at roughly 1,200 to 1,800 credits, while an email rewrite or a small page revision is a small task and a coordinated funnel is a medium one. Two large builds and a launch in the same month is a scheduling decision, not an accident to discover on the 20th.
3. 03Read the estimate as a scoping signal. Before a larger task begins, Agorix should show an estimated credit range so you can review the scope and decide whether to proceed. An estimate much larger than you expected almost always means the brief is vaguer than you thought — narrow it and ask again rather than approving and hoping.
4. 04Know which capacity expires. Included plan credits reset monthly. Purchased credit packs are prepaid and currently roll over for twelve months. So a predictable heavy month — a launch, a course build, a migration — is an argument for a pack rather than for permanently moving up a tier.

The estimate shown before a larger task is the most useful part of this whole mechanism, and it is worth treating as a conversation with yourself rather than a toll. It arrives at exactly the moment when changing the scope is still free — which is the same reason an approval gate is placed before execution rather than after it.

Trust is granted to a job, never to an agent.

That sentence is the whole safety model. An agent that has prepared forty good follow-up sequences has earned nothing at all about pricing pages, and treating competence as transferable is how organisations end up surprised.

1. 01Drafts only. Output goes nowhere near a customer and nowhere near your live configuration. Stay on this rung until your corrections stop being structural — the moment you are changing wording rather than approach, the job is understood well enough to move up.
2. 02Gated execution. The agent prepares and configures the work where it belongs, and nothing customer-facing or consequential is released without approval. Most jobs should live here permanently, and that is not a halfway house. It is the design.
3. 03Standing approval, for one proven job. A narrow, named job runs without a per-instance gate, inside a cap, reviewed on a schedule. This rung is earned by a specific job rather than granted to an agent, and the difference between those two sentences is the entire safety model.

Five conditions before a standing approval

Write the demotion rule at the same moment you grant the approval, in the same sentence: one materially wrong output returns this job to gated review. Deciding that in advance is the entire trick, because the day after something goes out wrong is the worst possible day to have a conversation about policy — everybody is defending a position instead of applying a rule.

What each rung looks like in practice depends on the roles, permissions and controls configured in your workspace, and on what the platform currently supports. Confirm the activity history and the reversibility you actually have before granting anything on the third rung — that rung assumes you can see what happened and undo it.

Four ways prepared work is wrong, and what each one is telling you.

Wrong output is information about your inputs far more often than it is information about the agent. Reading it that way is the difference between improving the system and rewriting the same brief forever.

1. 01Subtly wrong facts. The dangerous one, because everything around it is right. A date that shifted, a price that changed last quarter, a bonus that no longer exists. Almost always caused by two source documents disagreeing and nobody having noticed.
2. 02Right-shaped and hollow. The structure is perfect and there is nothing specific in it. Usually means the source material was thin, so the work had nothing to be specific about.
3. 03Correct but off-voice. Accurate, competent, and not how you talk. Usually means the brand guidance available to it is a list of adjectives rather than examples of your actual writing.
4. 04Correct but aimed at the wrong people. Good work for an audience you did not mean. This is a brief failure rather than an output failure, and rewriting the output fixes it exactly once.

In every case the response is the same shape: shrink the job, not the ambition. Put that job back behind a gate, fix the input the failure pointed at — the disagreeing documents, the thin source material, the brand guidance that is a list of adjectives — and run it again on a smaller scope. What does not work is asking more insistently.

Start with one job. Run it as drafts until the corrections are cosmetic, move it behind a gate, and leave it there while you learn what it does on an unusual week. Add the second job only when the first has stopped being interesting — which is the same restraint that governs your first marketing automations, for the same reason.