The GTM Agent Operating Model

Most teams adopting AI agents for go-to-market skip the only question that matters and go straight to tooling. The question is not "which agent writes the best LinkedIn post". It is "which decisions am I keeping, and what does the agent need in order to make the rest of them well".

This is an operating model, not a tool comparison. It works whether you run agents through Orbitable, through a general assistant, or through something you built yourself.

Start with the three kinds of GTM work

Go-to-market work is usually discussed by channel: content, outbound, paid, events. That split is useless for deciding what to delegate, because every channel contains all three of the following.

Judgement work decides what is true and what matters. Who the customer is. What the product is actually for. Which of three positions to take. This work is compressed, high-stakes, and almost impossible to check afterwards because a confident wrong answer looks exactly like a confident right one.

Craft work turns a decision into an artefact. The sequence, the page, the brief, the battle card. It is checkable: you can read the output and tell whether it is good. It is also where most of the hours go.

Maintenance work keeps the system honest. Refreshing a target list, checking whether a competitor changed their pricing, noticing that four calls in a row raised the same objection. It is unglamorous, it compounds, and it is the first thing dropped when a team gets busy.

Craft and maintenance delegate cleanly. Judgement does not, and every disappointing agent programme we have seen delegated judgement by accident, usually by giving an agent a vague brief and letting it decide the strategy inside the draft.

The four handoffs

Any single piece of work sits in one of four modes. Naming them stops the argument about whether "AI can do marketing", which is not a real question.

Two rules follow from the table, and they are the whole model.

The first: never move a piece of work up a mode to save time. Teams that jump from "delegate and review" to "automate" because review is slow do not get faster. They get a backlog of published work nobody has read, and they find out about it from a customer.

The second: the mode is a property of the work, not of the agent. A better model does not move competitive positioning from Direct into Automate. It makes the artefact better inside Direct.

What the agent needs in order to be good at craft

An agent producing craft work is only as good as the judgement it inherits. That inheritance has a name in most systems: shared context, a knowledge base, a world. Whatever it is called, it needs four things, and the order matters.

1. Who the customer is, specifically enough to exclude people. An ICP that excludes nobody is not an ICP.
2. What the product is for, in the buyer's words rather than the roadmap's.
3. What has already been decided, including the positions you rejected and why. Without the rejections, an agent will rediscover them and argue for them.
4. What is off limits, which is the cheapest and most-skipped of the four. Claims you cannot make. Numbers you cannot cite. A competitor you do not name.

Teams supply the first two and skip the last two, then are surprised when the output is generically correct and specifically useless.

The review contract

Review is where an agent programme succeeds or quietly rots. Three rules, stated as rules because soft versions of them do not survive a busy week.

A human approves anything that leaves the building. Not "reads". Approves, with a recorded decision attached to a specific version. If nobody can say who approved the thing that went out, nobody did.

Feedback attaches to the artefact, not to a conversation. A note that says "make it punchier" in a chat thread is lost the moment the thread scrolls. A note pinned to the second paragraph survives, and can be checked against the next round.

An unresolved note carries forward. The most common failure in round two is that round one's real objection quietly disappears because the draft changed enough to hide it. Carrying unresolved notes into the next round is a mechanical fix for a human failure of memory.

The four failure modes, named

You will hit at least two of these. Naming them in advance is most of the cure.

Plausible sludge. Output that is grammatical, on-topic, structurally correct and completely unusable, because it was briefed with a topic rather than a decision. The fix is upstream, in the brief, not in the model.

Confident fabrication. A number, a study or a quotation that does not exist, produced because the brief demanded authority and gave the agent no source to be authoritative from. The fix is to say explicitly what may be asserted and what must be sourced, and to make "I do not have a source for this" an acceptable answer.

Review theatre. Everything technically gets approved, but approval means someone clicked a button within four seconds. This shows up as a sudden drop in the number of change requests. If nothing is ever sent back, review has stopped happening.

Volume drift. The system produces more each month and the team reads less of it. Output goes up, outcomes do not. The counter is a hard cap on published work per week, set below capacity on purpose.

How to start, in order

• Pick one craft workflow you already do badly because you never have time. Not your best one.
• Write down the four context items above for that workflow. If you cannot write down what is off limits, you are not ready to delegate it.
• Run it in "delegate and review" for a month with every single item reviewed, and count how many you send back.
• When the send-back rate is stable and low, and only then, consider sampling instead of reviewing everything.
• Keep judgement work in Direct. Revisit that in a year, not a quarter.

The teams that get value from agents are not the ones with the best prompts. They are the ones who wrote down what they were not willing to delegate before they delegated anything.