Agency Operations Guide: Systems, Roles, and Delivery Workflows

This is the reference piece for everything else I have written about running the delivery layer of a marketing agency. If you arrived from another article in this cluster, this is the map they hang off. A note on the phrase itself: “agency operations” means one thing to a government department and something quite different to a forty-person marketing agency. I am writing about the second one.

I have spent twelve years running delivery operations for digital marketing agencies, across 79 completed engagements with a 4.9 out of 5 client rating. What follows is the whole territory in one place: the delivery workflow, who owns what at five, fifteen and fifty people, the systems layer underneath, the metrics worth watching, and where AI now fits. Each section points to the deeper piece where one exists.

What agency operations actually covers

Operations is the layer between the strategy you sold and the work that ships. It is not the work itself. Your strategists, designers, writers and media buyers do the work. Operations decides how a request becomes a project, who picks it up, when it moves, who checks it, and how the client hears about it.

The clearest test I know is this: if a process exists only because one specific person remembers it, you do not have operations, you have that person. That is an acceptable answer at four or five people, where the founder is the system. It stops being acceptable when the founder is the last line of defence on every deliverable, and it becomes a business risk the day a good account manager leaves with the only working copy of the process.

Everything below is one of four things: the workflow, the roles, the systems, or the measurement. Agencies usually try to fix them in the wrong order, and the wrong order starts with buying a tool.

The delivery workflow is the spine

Every agency runs the same seven stages whether it has named them or not: intake, scoping, scheduling, production, QA, delivery, reporting. The stage-by-stage version, where each one breaks in real agencies and the single discipline that fixes it, is written up in full in project management for agencies. I am not going to repeat it here.

What belongs at this level is the property that makes the workflow hold together. Every stage needs a named owner and a written condition for leaving it. Without exit conditions, work moves between stages on assumption. Someone thinks the brief was clear, someone assumes design has seen it, someone believes the report went out. That is the mechanism underneath most missed deadlines, and no tool prevents it.

The second property is capacity visibility. An agency is many concurrent projects drawing on one shared pool of people, so a per-project plan cannot see the overload that breaks the week. You need one view showing what each person is carrying across every project. If you change nothing else this quarter, build that view.

Who owns what at 5, 15 and 50 people

The responsibilities do not change as an agency grows. It is always intake, scheduling, quality, client communication and measurement. What changes is how many people they are spread across, and every painful growth stage I have watched is the same moment: one person holding more of that list than a person can hold. The sizes below are patterns I have seen repeatedly, not rules.

Around five people

The founder is the system, and at this size that is the right answer. Operations is one person holding the whole picture. The useful discipline is not hiring, it is writing down the two or three processes that repeat most often so the picture survives a holiday. The common mistake at five is buying a project management tool before knowing which process it is meant to hold.

Around fifteen people

This is where operations becomes a job rather than a habit. Fifteen people means enough concurrent projects that nobody holds the schedule in their head, and a founder’s day fragmented past the point of usefulness. Something has to be delegated, and the split that works is coordination first: one person owns intake, scheduling and status, while the founder keeps client relationships and quality. Delegating quality first is the standard error, because quality is the thing a founder is least able to release and will quietly take back.

Fifteen is also where agencies discover the difference between wanting the function handled and wanting a person in the seat. Both are legitimate. Project management run as a service moves coordination out of your building. A fractional project manager sits inside your team and your tool, making priority calls as they arise.

Around fifty people

At fifty, operations is a layer rather than a person: teams or pods with their own coordinators, someone senior owning the system across them, and a real separation between running delivery this week and improving how delivery runs. Those are two different jobs, and when one person holds both the second one loses, because live client work always wins that argument. This is the size at which the leadership version of the role appears, whether you call it a head of operations, an operations manager or a fractional COO.

The failure I see most at this size is a fifty-person agency still running fifteen-person operations, with exceptions still routing through the founder and handoffs between pods still undocumented.

The systems layer underneath the workflow

Underneath the workflow sits the layer nobody volunteers to own: how work is named, where the single source of truth lives, what an SOP contains, and who updates it when the process changes. It is unglamorous, and it is the difference between a process that survives a resignation and one that leaves with the person.

The minimum I would build for any agency:

• One source of truth for project status, and a rule that anything not in it does not exist.
• A shared vocabulary: what in review means, what done means, and what separates a project that is pending from one that has started.
• Written SOPs for the processes that repeat and cost the most when done differently each time: onboarding, reporting and QA.
• A documented handoff format between every pair of roles that hand work to each other.
• A named owner for the system itself, with protected time in their week to maintain it.

On tools, briefly. I work inside ClickUp, Asana, Monday.com, Airtable and Notion, and I automate across them with n8n. I have no reseller relationship with any of them and no interest in migrating you as the price of entry. When a process is defined, most tools handle it adequately. When it is undefined, no tool saves you, and that undefined state is business systems work rather than a tooling decision.

The metrics that actually matter

I am not going to give you benchmark numbers. Any figure I quoted for a healthy utilisation rate or an acceptable rework percentage would be invented, and invented benchmarks get repeated in leadership meetings as though they were data. Here is what I actually watch.

• On-time delivery measured against the date the client was given, not the date the task was later moved to. The gap between those two is the honest number.
• Rework hours, separated from production hours. Work done twice is the cleanest signal that intake or scoping failed upstream, and most agencies never split it out, which is why it stays invisible.
• Time spent waiting between stages, not time spent working. Agency weeks disappear in the gaps, not in the production.
• Forward capacity load per person for the next two weeks. A report on last week’s overload arrives after the damage is done.
• Client-visible slips, counted separately from internal ones. A missed internal date costs hours. A missed client date costs trust. They should never sit inside the same number.

Trend beats precision. A roughly measured number you look at every week is worth more than a precise one calculated once a quarter.

Where AI fits, and where it does not

The sequence I hold to is process first, automation second, AI only where it beats a rule. A surprising share of what gets sold as AI is work an ordinary rule handles better, and rules are predictable, auditable, and do not drift when a model updates. Before paying for anything with AI in the name, it is worth knowing how much of your problem is plain rule-based automation.

Where AI has earned its place in agency delivery, in my experience: drafting recurring written artefacts that a person then edits, condensing long threads and call notes into something a project record can hold, and the assembly half of reporting where numbers are pulled and arranged. The pattern is consistent. AI does the assembly, a person keeps the judgement, and there is a named human checkpoint before anything reaches a client.

Where it does not belong: priority calls, client communication that depends on reading the room, and anything where being wrong stays invisible until a client finds it. It also does not belong on top of a process that changes shape every week, because automating an undefined process produces plausible output that is wrong in ways nobody catches until a client does.

I have written the longer version, including what I will not automate, on my AI automation consulting page.

If you take one thing from this guide

Every stage gets a named owner and a written condition for leaving it. That single rule is most of what twelve years of this taught me, and it costs nothing to apply. After that the order is capacity visibility, then the systems layer, then measurement, then automation, and AI last of all. Agencies that run that order backwards buy a tool first and conclude that operations is not the problem.

The full set of ways agencies engage me on this work is on my work page, and the diagnostic version of the role, where something is clearly broken but nobody can name it, is operations consulting.

Where to start

Diagnosis before purchase. The Agency Ops Audit is $1,500 fixed, takes two weeks, and gives you a written 90-day roadmap plus a 60-minute readout call: which part of your operation is failing first, what to fix in what order, and what to leave alone. If it calls for ongoing help, the fractional retainer runs $4,000 to $6,000 a month and implementation projects are $8,000 to $14,000 fixed. If your process is already sound, I would rather tell you that in the readout.