The Roles and Skills an Agentic Marketing Operations Team Needs
The five human roles an agentic marketing operation needs, the skills behind each, a team structure by company size, and when to reskill current staff or hire.
On this page
The team roles and skills needed for agentic marketing operations fit on one page: five human jobs, the skills behind each, and a rule about which jobs one person can hold at once. The five are agent owner, workflow designer, evaluator, data steward and approver. Those names are ours. We found no standard or vendor document that prescribes a marketing org chart for agents, so this piece separates what published studies and documentation say, as we read them on October 7, 2026, from the structure we propose on top.
Which human roles does an agentic marketing operations team need?
An agentic AI marketing operations team needs five roles held by people, however much of the execution the agents take on. Agents plan and act across tools (our primer on what agentic AI is covers the mechanics), yet a person still answers for the outcome, the design, the quality bar, the data and each risky action.
| Role | Owns | Decides alone |
|---|---|---|
| Agent owner | The business result of one workflow, and its budget | Whether the agent keeps running |
| Workflow designer | Instructions, steps, tool access and handoffs | How the work gets done |
| Evaluator | The test set, the grading rubric and the weekly sample | Whether output meets the bar |
| Data steward | The fields the agent reads and the fields it may write | Which data is fit to use |
| Approver | The review queue for customer-facing or hard-to-reverse actions | Whether one action ships |
The five-way split is our proposal; the call for a written split is documented. The AI Risk Management Framework playbook from NIST states it under GOVERN 3.2: "Policies and procedures are in place to define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems." The 2026 Work Trend Index from Microsoft poses three questions it says every Frontier Firm will need to answer: "Who reviews agent performance? Who has the authority to update the workflows that agents run? How does a local win get captured and scaled across the organization?" Our evaluator answers the first, the workflow designer the second and the agent owner the third.
Some of these roles already exist in software. HubSpot documents that when an agent runs from an automation, "the agent owner's permissions are used," and that "By default, custom agents only read information, and you must turn on write access." The prospecting agent documentation from HubSpot describes two Autonomous mode options: "Review before sending," where "you'll review and approve each message before it goes out," and "Send automatically." Both pages list HubSpot Credits as required. An owner, a write-access decision and an approval queue exist in the product whether or not anyone has been named to hold them. Our piece on what marketing agents can write back to a CRM goes deeper on the write side.
The skills behind each role
The agentic marketing operations roles and skills pair up one to one, and most of the skills are judgment marketers already practice. The same Microsoft survey (20,000 knowledge workers who use AI at work, 10 markets, February to April 2026) asked which human skills are more important as AI takes on more work, and two led the list: quality control of AI output at 50% and critical thinking at 46%. The January 2025 Future of Jobs Report from the World Economic Forum, drawing on over 1,000 employers that represent more than 14 million workers, likewise found analytical thinking the most sought-after core skill, with seven out of 10 companies considering it essential in 2025.
| Role | Core skills | Usually grows from | First deliverable |
|---|---|---|---|
| Agent owner | Unit economics, prioritization, vendor and credit budgeting | Marketing ops lead or channel lead | Charter with target metric and stop conditions |
| Workflow designer | Process mapping, instruction writing, tool permissions, exception handling | Marketing automation specialist | Step map with escalation rules |
| Evaluator | Test design, rubric writing, sampling, basic scripting for automated checks | Marketing analyst | Test set with edge cases and a pass mark |
| Data steward | Field definitions, consent and suppression logic, deduplication, access scoping | CRM admin or data analyst | Read and write field list |
| Approver | Brand and claims judgment, product knowledge, fast written feedback | Content lead or campaign manager | Review checklist with reject reasons |
The matrix is our proposal, and the evaluator's row has the most published guidance behind it. Evaluation documentation from Anthropic advises building tests that mirror the real task mix, edge cases included, and describes human grading as "Most flexible and high quality, but slow and expensive. Avoid if possible." An evaluator who can script simple checks saves human review for the cases that need it.
How the structure changes with company size
An agentic AI marketing team structure keeps the same five roles at every company size and changes how many people share them. The size bands below are our working assumption; we found no study that sets headcount per role.
| Marketing team size | Who holds the roles | Rule that matters most |
|---|---|---|
| Up to 10 people | Two people. One is owner, designer and approver; the other is evaluator and data steward | The person who builds a workflow never grades it |
| 11 to 50 people | An owner per workflow, a shared designer in marketing ops, a part-time evaluator from analytics, the CRM admin as steward, approvers in each channel | One named owner per workflow |
| More than 50 people | A small agent operations group of designers, evaluators and a steward serving channel teams, which keep their own owners and approvers | Evaluators report outside the build team |
Design and evaluation are separated first because the NIST playbook describes that arrangement under GOVERN 2.1: in some risk management structures, "professionals carrying out test and evaluation tasks are independent from AI system developers and report through risk management functions or directly to executives." At larger sizes IT and security join in, two of the four roles the Microsoft report says have to change together. Per-workflow accountability and permission tiers are in our governance operating model for marketing agents.
Where the agents sit and who approves what
The roles AI agents take on a marketing team are execution roles, such as researching, scoring, drafting and routing, and in our model each agent reports to a named person. A hypothetical example: a 12-person B2B team puts inbound lead follow-up on an agent. Every figure is an illustrative assumption.
- The data steward lists the fields the agent reads (lifecycle stage, consent status, last engagement) and limits its writes to two properties.
- The workflow designer writes the instructions and one escalation rule: a lead from a named target account goes to a person.
- The evaluator assembles 60 past leads with known outcomes, 10 of them awkward (missing company, competitor domain, opt-out pending), and sets a pass mark of 54 sendable drafts out of 60.
- The approver reviews every draft for the first three weeks and logs a reason for each rejection.
- The agent owner reads the weekly sample and decides whether one segment moves to automatic sending.
Three failure modes are worth planning for in structures like this. The designer grades their own build and the pass rate flatters it. The approval queue grows until review becomes a reflex click, and the Microsoft report notes that "the more agents execute, the higher the stakes around human evaluation." Or the owner changes jobs while the agent keeps running on that person's permissions, the HubSpot behavior quoted earlier. The build sequence for one workflow is in our guide to putting one marketing workflow on agents.
Should you reskill current staff or hire?
For agentic AI marketing team hiring and reskilling in 2026, our recommendation is to fill all five roles from current staff where you can and to hire only for the evaluator's scripting when no analyst on the team can write automated checks. Employers in the published studies plan to use both routes. In the World Economic Forum report, employers estimated that of every 100 workers, 59 would need training by 2030: 29 could be upskilled in their current roles, 19 upskilled and redeployed, and 11 would be unlikely to receive it. Among surveyed employers, 85% planned to prioritize upskilling and 70% expected to hire staff with new skills.
The 2025 Work Trend Index from Microsoft, which surveyed 31,000 knowledge workers across 31 markets from February 6 to March 24, 2025, shows which roles draw hiring: 28% of managers were considering hiring AI workforce managers to lead hybrid teams of people and agents, and 32% planned to hire AI agent specialists, while 47% of leaders listed upskilling existing employees as a top workforce strategy for the following 12 to 18 months. We found no comparable hiring figures in the 2026 edition.
The reskilling paths are short. Marketing ops staff already map processes and manage tool permissions, analysts already sample and test, content leads already judge claims, and CRM admins already own field definitions. What is new is narrow: instructions an agent follows, a graded test set and scoped write access. When nobody can take those on, a hire or outside help such as our agentic AI automation service is the sensible route.
Training is also a legal matter for teams the EU AI Act covers. The AI literacy Q&A from the European Commission, last updated July 27, 2026, says Article 4(1) requires providers and deployers of AI systems to take measures to support the AI literacy of their staff, with no specific level mandated since amendments entered into force in mid-July 2026. The page adds that "The employees and the human-in-the-loop need the appropriate skills, targeted as to the system they are using," and that "no specific governance structure is mandated to comply with article 4 of the AI Act." The five roles stay a design choice, while skills for whoever sits in the loop are the part the Commission addresses.
How to tell whether the structure is working
Start with the number the workflow was built to move, such as qualified pipeline, retention or cost per lead, against a baseline from before the agent. Under it sit four operating measures:
- Evaluator: pass rate on the test set and on the weekly live sample.
- Approver: hours from draft to decision, and the share of drafts rejected with a logged reason.
- Data steward: records skipped or escalated because a field was missing or in conflict.
- Agent owner: cost per completed run against the value of the outcome.
The cost side is broken down in what a marketing AI agent costs per contact, and our AI ROI calculator is a further resource for framing the return.
Written structure is still the exception in the 2026 Microsoft survey: 26% of Frontier Professionals, the report's term for its most advanced AI users, said agent workflows, human handoffs and quality standards are documented and repeatable at team level, against 19% of other respondents. Putting five names, the handoffs and the quality bar on paper for one workflow is the cheapest way to join that minority.
Sources
- World Economic Forum, Future of Jobs Report 2025 (January 2025)
- Microsoft, 2026 Work Trend Index Annual Report (May 5, 2026)
- Microsoft, 2025 Work Trend Index Annual Report (April 23, 2025)
- NIST, AI RMF Playbook, Govern (written for AI RMF 1.0, which NIST says is being updated, with the Playbook to follow)
- European Commission, AI Literacy Questions and Answers (last updated July 27, 2026)
- HubSpot, Run agents and review output
- HubSpot, Set up and use the prospecting agent
- Anthropic, Define success criteria and build evaluations