Could AI Replace An Entire Marketing Team?

AI adoption is not the bottleneck any more. Results are. 88% of organisations now use AI in at least one business function, and most are still stuck in experimentation. The gap between adoption and results keeps widening. Marketing feels this most sharply. The supply of tools grows faster than any team's ability to absorb them, and budgets have not moved to match. What looks like an AI problem is really an accumulation problem: more software, more pilots, no architecture connecting them. The waste is not caused by one bad vendor or one slow team. It is structural, and it shows up in three places at once: what teams buy, what they actually use, and what they can afford to run.

The numbers from McKinsey’s State of AI 2025 report are stark. The 2025 Gartner Marketing Technology Survey paints a similar picture: 81% of martech leaders are piloting or fully implementing AI agents, yet utilisation of their overall martech stack sits at just 49%. That’s an improvement from 33% in 2023, but still means half of marketing technology capabilities go unused.

Meanwhile, the 2025 Marketing Technology Landscape now catalogues 15,384 solutions, up 9% from the previous year, and 100x growth since 2011. Of the new tools added this year, 77% were AI-native.

More tools. More AI. Same fundamental problem.

CMOs Are Paying the Price

The 2025 Gartner CMO Spend Survey reveals that 59% of CMOs report insufficient budget to execute their strategy. Marketing budgets remain flat at 7.7% of company revenue. CMOs are expected to do more with less, and 65% believe AI will dramatically transform their role within the next two years.

Yet while budgets stagnate and expectations rise, waste accelerates:

• SaaS Bloat: The average enterprise manages 275 SaaS applications (Zylo 2025 SaaS Management Index), yet uses only 47% of the licenses purchased.

• Rising Costs: SaaS spend per employee has risen to $4,830, up 21.9% year-over-year, driven by unexpected consumption-based AI pricing models.

• Projected Waste: Organisations without centralised visibility will overspend by at least 25% by 2027 due to redundancy (Gartner Magic Quadrant for SMPs).

Marketing is both a contributor to and a victim of this waste. Bleeding budget into tools that don’t connect. Paying for potential rather than performance.

More AI won’t fix this. This isn’t a technology problem. It’s an architecture problem.

Tool selection is the wrong starting point. Most AI marketing conversations open with "Which AI tools should I use?" or "How do I automate this task?", and both lead somewhere unhelpful. I have watched teams collect AI tools the way you collect puzzle pieces, assuming the picture will assemble itself. It rarely does. The reason is structural. Those questions optimise for the smallest unit of work, so the answer is always another purchase. They never ask what connects the purchases, or what the connected whole should produce. Ask a question scoped to a single task and you get a tool. Ask one scoped to the whole operation and you get an architecture. The starting question decides the outcome, which is why so many AI programmes are well equipped and going nowhere.

A better question: How do I turn disconnected tools into a governed marketing system that executes bounded workflows with human approval where it matters?

This reframe changes everything. It shifts focus from tools to architecture. From features to outcomes. From prompts to workflows.

But to build a system, you need to understand how the components think. And for a long time, that was impossible.

Early large language models hid their reasoning. In the GPT-3 era, around 2020, a prompt like "write SEO-optimised content for our product launch" produced finished text with no visible path from request to result. You saw the input. You saw the output. Everything in between stayed sealed inside the model. That opacity mattered more in marketing than in most fields, because the value of an experienced marketer sits in the judgement before the writing, not in the writing itself. A model that jumps straight to a draft has skipped the part that makes the draft correct, or at least has not shown that it did any of it. Without visibility into those steps, there is nothing to inspect, nothing to correct, and nothing durable to build on.

Contrast this with how an experienced content marketer works:

• Audience: Who is this for? What stage of the journey are they in?

• Landscape: What are competitors saying? Where’s the white space?

• Strategy: What’s our unique angle? What proof points support it?

• Success: What does success look like? Traffic? Conversions? Brand lift?

They run through this mental checklist, consciously or not, before executing. LLMs didn’t do that then. I called it the Black Box Gap. And I needed to close it.

Language models can reason in steps, and that stopped being a hunch in 2022. Chain-of-Thought prompting showed that a model asked to work through a problem performs very differently from the same model asked for an answer. The capability was already inside the weights. It stayed dormant until something in the prompt switched it on, and the switch turned out to be almost absurdly small. That finding reset what AI could be trusted with in marketing. If a model can be made to show a route to its answer, it can be handed the analytical work that sits before execution, not only the execution. Reasoning became something you could request and then inspect. That is the precondition for trusting a model with anything more consequential than a first draft.

In 2022, researchers at Google published a landmark study on Chain-of-Thought (CoT) prompting. When LLMs are given step-by-step reasoning examples, their performance on complex tasks improves dramatically: from 17.9% to 58.1% on a mathematical reasoning benchmark.

A follow-up study by the University of Tokyo and Google found that simply adding "Let’s think step by step" before a problem, with no examples at all, triggered similar reasoning capabilities. On one benchmark, accuracy jumped from 17.7% to 78.7%. A 4x improvement from five words.

This Zero-shot Chain-of-Thought research revealed something profound: LLMs contain latent reasoning capabilities that emerge when explicitly activated. The models weren’t just pattern-matching. They could decompose problems the way experienced practitioners do when properly prompted.

The implication for marketing: AI doesn’t just generate content. It can reason through problems (audience analysis, competitive positioning, content strategy) when given the right structure.

Chain-of-Thought stopped being a technique and became architecture. Modern models ship with step-by-step reasoning built in, so the prompt trick that mattered in 2022 is now largely redundant. That changes the job. When reasoning is native, the interesting question is no longer how to phrase a request but what the system is allowed to do once it has reasoned. This is the line between AI as a tool you operate and AI that operates on your behalf. Such systems choose a path, take the action, and adjust on the next run instead of returning a completion and stopping. Marketing sits right on that line today. The technique era rewarded whoever wrote the better prompt. The agentic era rewards whoever decides what the system may do without asking first.

OpenAI’s reasoning models (o1, o3, o4-mini) perform step-by-step thinking automatically, without explicit prompting. As Microsoft documentation notes, models like the o1-series have built-in chain-of-thought reasoning, meaning they "internally reason through steps without needing explicit coaxing."

The Wharton Prompting Science Report (June 2025) confirmed this evolution: "For models with built-in reasoning capabilities, CoT prompting produced minimal benefits… Many models perform CoT-like reasoning by default, even without explicit instructions."

The evolution from CoT prompting to built-in reasoning has culminated in what the industry now calls "Agentic AI." These are autonomous systems that don’t just respond to prompts. They make decisions, trigger actions, and learn across cycles.

BCG’s recent research describes the shift: "Past innovations, from CRM to marketing automation, helped streamline discrete steps. Agentic AI goes further. These systems introduce autonomy: they make decisions, trigger actions, and learn across cycles."

McKinsey’s analysis is direct: "Success calls for designing processes around agents, not bolting agents onto legacy processes."

Sound familiar? This is the "pile of parts" problem at a new scale.

Marketing work is not one thing, so it cannot be automated as one thing. Split it first. I borrowed Jobs To Be Done from product thinking, where it normally explains why a customer hires a product to solve a problem, and turned the lens around onto the work itself. What job is this piece of marketing hired to do? Asked that way, marketing separates into two kinds of job. Atomic jobs are self-contained: one input, one output, done. Composite jobs are sequences of atomic ones with decisions between the steps. The distinction decides everything downstream. It tells you what a model can be handed directly, what needs orchestration built around it, and why a tool that handles the first will never handle the second.

In my synthesis, every marketing activity can be decomposed into discrete jobs with clear inputs, outputs, and success criteria. This led me to identify two types:

Atomic JTBDs

Single, well-defined tasks with clear parameters:

• Analyse competitor pricing pages

• Generate 10 headline variants for A/B testing

• Score leads based on engagement signals

• Extract key themes from customer reviews

Composite JTBDs

Complex workflows requiring multiple atomic jobs in sequence:

• Develop Q1 content strategy (requires: audience analysis → competitive review → theme identification → content calendar → brief creation)

• Launch product campaign (requires: positioning → messaging → creative development → channel planning → execution → optimisation)

The insight: AI excels at atomic jobs today. Composite jobs require orchestration: the systematic integration of atomic jobs into coherent workflows.

Example: Content Creation as Composite JTBD

A "create blog post" request seems simple. But decomposed:

Example: Competitive Response as Composite JTBD

A competitor launches a new feature. Your response seems reactive. But decomposed:

What appears to be instinct is, in fact, a workflow. The experienced marketer runs this loop unconsciously. AI makes it explicit and repeatable.

Example: Lead Nurturing as Composite JTBD

A new lead enters your funnel. The nurture sequence seems automated. But decomposed:

Each step is an atomic JTBD that AI can execute today. Most marketing automation handles fragments: steps 3 to 5 of lead nurturing, for example. The gap is orchestration.

A catalogue of jobs is not a system. Decomposing marketing work gives you parts, and parts sitting beside each other do not make an engine. Something has to specify how they fit, in what order, and with what passing between them. Someone has to be accountable for running the result. That someone is a role rather than a product, which is why no vendor sells it. Supply was never the constraint here: the martech catalogue lists 15,384 products. What is scarce is the judgement to choose a few of them, assign each a job, and hold the assembled thing to an outcome. Engines are designed and operated. Parts are only purchased. The difference between a stack and an engine sits entirely in that judgement.

I call this the Operator function: the strategic orchestration that connects atomic jobs into coherent workflows.

Think of the martech landscape’s 15,384 solutions as components. Each does something useful in isolation. But without a unifying architecture:

• Data doesn’t flow between systems

• Insights don’t inform decisions

• Optimisations don’t compound

• Strategy remains disconnected from execution

This is why "pile of parts" is the defining problem of AI marketing today.

The Evolution of Connection

For years, we solved this with "digital duct tape": linear automation tools like Zapier to glue APIs together. Today, we have far more powerful options:

• Standard Protocols: The Model Context Protocol (MCP) acts as a universal "socket," allowing AI to plug into data sources without custom code.

• Advanced Orchestrators: Platforms like n8n enable complex, agent-based workflows with loops and memory.

But having powerful tools creates a new trap: The Orchestration Illusion.

Just because you can build a complex autonomous workflow in n8n doesn’t mean you should.

• Privacy Risk: If you pipe customer data into an agent without privacy guardrails, you are automating compliance risk.

• Brand Risk: If you connect a content generator to a social publisher without a "Brand Voice" filter, you are automating brand damage.

• Cost Risk: "Connected" means "consumption." Continuous agent loops across multiple SaaS subscriptions drive massive API and compute costs. Inefficient orchestration creates "token bloat": paying for AI to read the same data thousands of times.

Connectivity is not strategy.

The technology solves the plumbing: how agents talk to each other. The Operator solves the design: what agents are allowed to say or do, and why.

Autonomy is a property of the job, not of the team running it. An Operator has to grade each piece of work before automating it, because a one-shot task and a multi-step workflow with branches need completely different handling. That grading needs a scale. I use five levels, L1 to L5, running from single prompts where nothing ships unreviewed, up to a system handed a goal and left to work out the route. Each level names who holds the decision, not how advanced the model is. The practical value is triage. Once a job carries a level, you know what to build around it, how much review it needs, and whether the technology to run it that way exists yet.

SAE International’s L0-L5 framework for autonomous vehicles has become the standard for discussing machine autonomy. Recently, researchers at the University of Washington adapted this thinking for AI agents, proposing five levels based on the user’s role: operator, collaborator, consultant, approver, and observer.

Drawing from both frameworks, I propose an autonomy model specifically for AI marketing systems:

According to McKinsey’s State of AI 2025 report, 23% of organisations are scaling agentic AI systems in at least one business function, with an additional 39% experimenting. But the use of agents is not yet widespread: most scaling efforts are occurring in only one or two functions.

Most marketing work operates at L2 to L3, often at different settings by engine. The target here is a durable, human-gated L3; interoperability and governance are separate axes.

AI can absorb bounded execution. It cannot absorb accountability. The honest answer to "could AI replace my entire marketing team" is that a skilled Operator still sets the intent, connects the engines, and signs off the decisions that carry consequences. What changes is how much of the work between those points a human touches. The target is a durable, human-gated L3, not a marketing function with nobody in it. Read the current state against that target and the shortfall is plain: 88% adoption, 49% stack utilisation, 15,384 available products. That is what accumulating looks like when nobody is architecting. Leaders who treat this as a shopping problem get more tools. Leaders who treat it as a design problem get something that actually runs.

The 2025 Gartner CMO Spend Survey found that GenAI investments are delivering ROI through:

• 49% improved time efficiency

• 40% improved cost efficiency

• 27% increased capacity to produce content

But these are efficiency gains, not transformation gains. The transformation comes from the architectural and orchestration expertise of an experienced marketing operator.

Next: The AI Marketing Framework →