The Marketer as Orchestrator

Orchestration means the marketer stops being the person who executes each task and becomes the person who specifies, supervises and verifies it. The work product changes. Instead of finished emails, lists and workflows, the orchestrator produces three things: the context an agent works from, the test set that proves its output is correct, and the permissions that decide what it may do without asking. Judgment stays human. What changes is where that judgment is written down.

What changes when the executor is software?

The quickest way to see the shift is to take tasks a marketing ops function already runs and ask what the human produces in each era.

Task	Executor-era work product	Orchestrator-era work product
Nurture programme	The emails, the workflow or engagement programme, the smart lists	The brief, voice and claims rules, the exclusion logic, a test set of personas with expected outputs, the approval gate before send
Lead routing	The round-robin workflow or assignment rules	The routing specification with edge cases (existing account owner, partner-sourced, unknown country), the escalation path, a replay test on last month’s leads
Data enrichment	A CSV import and a lookup	Which fields may be written and from which source, the overwrite rule, a 50-record test set with known true values
Campaign reporting	The dashboard	The metric definitions, the source-of-truth field for each, the anomaly thresholds that trigger a human review
List building	The list	The ICP expressed as field values, the suppression rules, the check that the list matches the definition

In every row, the artefact the human owns becomes more upstream and more durable. A good routing specification outlives the tool that executes it. A good test set outlives the model.

What is context engineering for a marketer?

Context engineering is writing down what an agent needs to know to do a task correctly, in a form it can read. For most marketing teams, the hard part is not the prompt. The definitions have never been written down anywhere.

The context pack we build for any agent that touches a CRM or MAP contains:

• ICP as field values, not prose. “Mid-market fintech” is not usable. industry in a named set, numberofemployees in a range, country in a list, and the enrichment source that populates each.
• Lifecycle definitions in internal values. In HubSpot, lifecyclestage uses internal values such as subscriber, lead, marketingqualifiedlead, salesqualifiedlead, opportunity and customer. The agent needs the entry criteria for each stage and the rule that stages only move forward. By default HubSpot will not move a record backward through a workflow unless the property is cleared first, and an agent writing through the API needs to know the same rule applies to it. In Salesforce, the same applies to Lead.Status picklist values and the conversion criteria.
• Source-of-truth map. For each fact the agent might read (owner, industry, lead source, ARR), which system and which field is authoritative. Chapter 1’s field dictionary is exactly this.
• Naming conventions. Campaign, programme and UTM naming rules, so agent-created assets can be found and reported on.
• Claims and voice rules. What you may say about the product, the proof points you are allowed to use, and phrases that are off-limits.
• Hard stops. Actions the agent must never take: sending to all contacts, deleting records, changing owner on a record with an open opportunity.

Two rules make the pack work. It lives in versioned files, not in someone’s chat history, so a change to a definition is a reviewed change. And it is the same pack for every agent working in that domain, so two agents never work from different ICP definitions.

How do you evaluate an agent’s output before trusting it?

Most agent pilots we are asked to review were judged on a demo and a few spot checks. That is how an agent gets promoted on the strength of its best output, not its typical one.

The alternative is an evaluation set: a fixed collection of inputs where you already know the right answer, built before the agent runs.

• Enrichment agent. Pick 50 accounts where you have verified industry, employee band and HQ country. Run the agent. Score each field as correct, wrong or blank. Wrong is worse than blank, because blank triggers a fallback and wrong does not.
• Routing agent. Take last month’s inbound leads and the owner each one should have received under the current rules, including the awkward cases. Replay them. Every mismatch is either an agent error or a gap in the routing specification. Both are useful.
• Content agent. Write ten briefs with known constraints (persona, stage, claim limits). Score output against a checklist: correct persona, no forbidden claims, correct CTA, correct naming.

Then keep the set fixed and rerun it after every change to the prompt, the model, the context pack or the underlying data. A score that drops after a model update is the signal you would otherwise find in a quarterly report.

In production, sampling replaces the test set. Pull a fixed share of the agent’s writes each week, check them, and log each defect by type. The defect log is what earns an agent promotion up the autonomy ladder; Chapter 4 covers that ladder in detail.

There is a reason to be strict here. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls (Gartner, June 2025). An evaluation set addresses the last two directly: it shows what the agent is worth, and it catches drift before a customer does.

What does governance look like at the task level?

Governance sounds like a committee. At the task level it is four settings and a register.

1. A dedicated user per agent.

• Salesforce: an integration user licence with a minimum-access, API-only profile, plus a permission set granting only the objects and fields the task needs.
• HubSpot: a private app or app user with the narrowest scopes that work. An agent that updates contacts does not need deal or settings scopes.
• Marketo: an API-only user with a custom role that includes only the API permissions the task needs, attached to its own LaunchPoint custom service so its calls are identifiable.

2. Field-level limits. Salesforce field-level security and HubSpot property permissions (on tiers that support them) restrict what the agent can edit. In Marketo, Block Field Updates protects specific fields from specific sources.

3. An audit trail. Every agent write should be attributable to the agent’s user in HubSpot property history or Salesforce Field History Tracking. If the agent runs under a shared admin, you have no trail.

4. A kill switch someone has tested. Deactivate the user or revoke the token, and confirm the agent stops. Do it once in the first week so the procedure is known before anyone needs it.

The register. One row per agent: its purpose, user, scopes, writable fields, autonomy rung, owner, eval score and last review date. It takes an hour to set up and it is the document an auditor, a new ops lead or a security review will ask for first.

What does an orchestrator’s week look like?

The cadence matters more than the tooling. The rhythm we run and recommend:

• Daily, 15 minutes: the exception queue. Anything the agent escalated, anything it refused, anything that tripped an anomaly threshold.
• Weekly: the sample audit. Pull the week’s sample, log defects by type, and decide whether any pattern needs a context-pack change.
• Monthly: rerun every evaluation set. Compare scores to last month. Review the agent register for scope creep.
• Quarterly: permission review for every integration and agent user, retire agents that are no longer used, and refresh the context pack against current definitions.

Notice what is missing: nobody is building lists or cloning emails. That time moves to the specs, the test sets and the exceptions.

Which marketing work should stay human?

The line is not “creative versus routine.” It is reversibility and blast radius.

Keep these with a named human approver, even when an agent does the preparation:

• Positioning, messaging and claims. The agent can draft; a person decides what the company says.
• Pricing and discount decisions. Recommend, never act.
• Whole-database sends. Any send to all contacts, or to a list above an agreed size, needs a human approval step.
• Destructive data operations. Bulk deletes and merges at scale. A HubSpot merge cannot be undone, and several of its effects on associations and property values are not obvious in advance. Our notes on what HubSpot does on merge cover the details.
• Anything that changes a customer’s experience in a way that is hard to take back. Ownership changes on open deals, renewal communications, sunset notices.

Everything else can move toward agent execution as it earns trust, one workflow at a time.

How do you start this quarter?

A 30-day plan we use with teams starting from zero:

1. Week 1: pick one workflow. High volume, reversible, already well understood. Enrichment and lead routing are the usual first choices.
2. Week 1: write its context pack. Definitions, source-of-truth fields, hard stops. If you cannot write it, the workflow is not ready for an agent. It is not ready for a new hire either.
3. Week 2: build the evaluation set. 50 records or the last month’s real cases, with known answers.
4. Weeks 2–3: run in recommend-only mode. The agent proposes; a human applies. Score against the eval set and a weekly sample.
5. Week 4: decide. Promote to act-with-review if defects are low and understood, or fix the context pack and rerun. Create the agent’s own user and add it to the register before it writes anything.

What have we seen in the field?

• The ICP that lived in a slide. An enrichment agent was scoring accounts against an ICP that existed only as a paragraph in a board deck. Two regional teams had interpreted it differently for years. Writing it as field values took a single working session and surfaced the disagreement. The agent was the least interesting part of the fix.
• The routing replay. Replaying a month of inbound leads against a new routing specification found more gaps in the specification than errors in the agent. Partner-sourced leads and existing-customer contacts had no rule at all; the old workflow had been handling them by accident.
• The admin-credential pilot. A content agent was connected to a MAP under a senior admin’s login. It worked, but none of its changes were distinguishable from the admin’s own. Moving it to an API-only user with a custom role took an afternoon and should have happened on day one.

What did we get wrong, and what is still open?

We started by treating context engineering as prompt writing. The prompts were fine. The definitions underneath them were not, and no prompt fixes an ICP that two teams define differently. We now start every agent engagement with the context pack and the field dictionary.

Still open: how far orchestration changes the skills a marketing ops hire needs. Evaluation and specification writing look more like QA and product management than like classic campaign ops. We think the best orchestrators are ops people who can write a clear spec, but the market’s job descriptions have not caught up, and neither have we fully.

Sources

• Gartner, press release, 25 June 2025: “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.”

Want to know which of your workflows are ready for an agent?

A GTM Systems Audit tells you where your definitions, field dictionary and permissions stand across HubSpot, Marketo and Salesforce, which is the groundwork every agent needs before it writes. Audits start from USD 1,500. To check your own instance first, work through the GTM Systems Audit Checklist. If you would rather have the orchestration run for you, that is what Embedded Ops is for.

Book a GTM Systems Audit