Everyone is using AI now. And almost everyone uses it the same way.
You log into Google Ads, send a report, paste the CSV into ChatGPT or Claude, get the analytics, and repeat the whole process for Meta, Google Analytics 4, or whatever else is on your plate that week. Same painful process, every field, every week.
That’s not AI-powered marketing. AI-assisted copy-pasting.
The AI in that workflow works on a static snapshot. Not live. It is not connected to your real account. You don’t know what happened yesterday or what your cost-per-acquisition (CPA) goal is. It’s a powerful engine that runs on old fuel, and it explains why its output feels inconsistent: good one day, the next, normal, which always requires more planning than it should.
The problem is not the model. The problem is the setup. There is a three-layer stack that fundamentally changes this: MCP to access live data, Capabilities to behave in the same way, and Claude Projects to package everything into a reusable team environment. Each layer addresses a different failure mode. Together, they are the difference between AI as innovation and AI as infrastructure.
Layer 1: MCP Gives AI Eyes on Your Real Business
The Model Context Protocol (MCP) is an open standard designed to connect AI models to external tools and data sources. Think of it as Zapier’s AI layer, except instead of moving data between apps, it gives AI the ability to read, query, and in some cases act on that data directly.
Without MCP, your AI works blindly. It knows a lot in general, but it doesn’t know anything specific about your business, your campaigns, your customers, or your operations. You copy-paste the numbers into the dialog window and ask it to analyze them. That’s not smart on the scale. That’s an expensive clipboard.
With MCP connected, AI can pull live data directly from your tools. Google Ads has an official MCP server, which means you can ask Claude to check which campaigns are underperforming compared to the CPA you’re currently targeting, pull search term reports, budget compliance issues, or compare the performance of campaigns across campaigns, and it’s asking a real account rather than waiting for you to attach a report. No exporting, no copy-pasting, no manual formatting steps.
The same principle applies to GA4, your CRM, or any other data source that has an MCP server available. But Google Ads is the clearest starting point for PPC teams because the data is live, the decisions are time-sensitive, and the performance gap between working on Monday data and Friday data is real and measurable.
For marketing teams, this is important because performance data is always moving. The analysis you do on Monday is out of date on Wednesday. An AI that can see live data is completely different from one that can’t.
Layer 2: Skills That Tell the AI How to Behave in Your Situation
MCP handles the data problem. Skills handle the problem of consistency.
A skill is a continuous set of instructions that tells Claude how to approach a particular type of task. Not what you should do once, but how to behave all the time. You define the rules once, and all conversations using that Skill automatically gain them.
For agencies, this is one of the biggest job openings available right now.
Think about how much confidential information resides within your agency that is never written down. Your senior analyst knows your reporting format, your preferred attribution model, how to frame recommendations for conservative versus growth stage clients, which shows what your most common client types really care about. A junior recruit takes six months to acquire that knowledge by osmosis.
Skill captures it in a few hundred words. You document your agency’s best practices and: how to plan campaign evaluation, how to draft budget recommendations, what tone to use in client-facing summaries, what key performance indicators (KPIs) to automatically mark. Each team member who uses Claude and that Skill works receives the judgment of the chief analyst from day one.
A concrete example: your agency has a standard way of auditing the Google ads account. He examines Quality Score distribution, search impression share by campaign type, lag windows before hitting return on ad spend (ROAS) targets, and always makes recommendations against the client’s stated growth goal instead of field benchmarks. All that checklist, installed as a skill, means that Claude runs those checks regularly through the Google Ads MCP connection, which pulls live account data and uses your framework automatically, not just when your most experienced person does it manually.
Layer 3: Claude Projects Pack Everything For Teams
Projects are Claude’s way of creating continuous, contextual environments. Each project has its own instructions, its own knowledge base, and its own memory that holds all the conversations. It is a functional container that enables the combination of MCP plus skills to be used at the team level.
For agencies, the setup is straightforward: One project per client.
Each client project gets the context of the client loaded: in their business model, their target audience, their historical performance rates, their seasonal patterns, any product guidelines that are relevant to the copy or message. And you connect Agency-level Skills, so they work automatically. Now, every conversation about that client starts from a place of complete detail.
The result is that anyone on your team who opens a client project, whether it’s an account lead, a strategist working while someone else is out, or a junior pulling a quick report, starts from the same base of knowledge.
For inside sales teams, Projects work differently but are just as powerful.
Instead of one Project per client, the in-house team typically creates one Project per task or workflow. The paid search project manages brand campaign design, naming conventions, bidding philosophy, and targeting metrics. When that project is linked to Google Ads via MCP, the question of “which campaigns are running over budget this week and underperforming against impression share goals” becomes a two-second question rather than a 20-minute reporting exercise. The content project manages the brand’s voice index, approved messaging frameworks, and current content calendar. The reporting project knows the stakeholders who receive the report, what they care about, and what format they expect.
Skills in internal settings manage the company’s own institutional knowledge rather than agency best practices. If your brand is always measuring success by new customer acquisition cost rather than combined ROAS, it’s always in Skill. If your growth team uses a specific exposure model for budget allocation, that resides in Skill. If a senior marketing executive prefers a one-page summary over a data dump, Ability handles that as well.
The practical effect is that the AI’s output stops feeling generic and starts to feel like something produced by a well-rounded team member.
Why A Stack Is More Important Than Any Single Tool
Each layer of this stack addresses a different failure mode that makes AI ineffective in real-world marketing environments.
MCP solves the data access problem. AI without access to data is amazing in demos and disappointing in production, because production is always about your specific numbers, not assumptions.
Skills solve the consistency problem. The quality of information varies among team members on a daily basis. A well-written skill does little quality and makes the output predictable enough to be trusted.
Projects that solve the content problem. Marketing is not a series of isolated questions. It is a continuous process in which context accumulates. Projects take that context forward, so every conversation builds on the last one rather than starting from scratch.
The teams that are reaping the real productivity benefits from AI right now aren’t the best informed. They are the ones who built a better place.
That difference is more important than it might seem. Much of the frustration with AI in marketing right now comes from teams that have embraced the tool but not the infrastructure around it. They gave their team access to a capable model and wondered why the results weren’t consistent, why the younger team members got worse results than the older ones, why nothing felt ready to be produced without hard planning. The answer is almost always the same: the model was powerful, but the environment was not designed to support it.
The switch is not technically complicated. Setting up a Google Ads MCP connection takes an afternoon. Writing a core Competencies document for your agency or team takes a few hours and one honest conversation about what your best people do differently. Creating a project structure takes less time than onboarding a new hire. The barrier is not professional. It’s a decision to treat AI as an infrastructure rather than a shortcut.
Once that decision is made, the merger begins. Every client project gets better as you add context to it. Each skill improves as you refine it based on the output you receive from customers. Nature gets smarter over time without the underlying model changing at all.
That’s what separates the teams that are building something solid from the teams that are still sending CSVs and hoping for the best.
Additional resources:
Featured image: Stokkete/Shutterstock
