The Twelve Technology Layers You Didn’t Know You Needed to Make Your AI Strategy Work

Ilana Shabtay, VP of Marketing at Fullpath, sets out her twelve-layer framework for dealership AI and the questions to ask about what a vendor provides, what your team must build and who owns the ongoing work.

There is a highly attractive pitch circulating right now where you just point a general AI model at your CRM, flip a magic switch, and suddenly you have a brilliant AI agent running effortlessly on your dealership data.

It sounds incredible and too good to be true. Unfortunately, it is exactly that.

If you haven’t already, you will soon start hearing the acronym MCP thrown around. MCPs stands for Model Context Protocol. An AI model out of the box is completely isolated and knows absolutely nothing about your dealership. To make the AI useful, you have to connect it to your proprietary data, a process that historically was an absolute nightmare as every software vendor has a different API that requires extensive work to hook up and manage.

MCPs are a standardized, open-source protocol that solves this problem by allowing the AI to securely talk to your local data sources and your dealership tools using one universal language. The MCP provides a secure, direct connection between the AI and your dealership data, dictating exactly what the AI can read and what tools the AI can use.

But the MCP is just the connection. Making the AI actually execute tasks demands 12 distinct layers of technology that fall into four categories: collection, identity and rules, the agent, and action.

The Twelve Tech Layers of Effective AI

Group 1: Collection

Collection is the floor everything else stands on. Skip any piece of it and your AI is doomed to confidently hallucinate absolute nonsense. It starts with sourcing (layer 1), where you bring together every fragmented system in your store. Think CRM, DMS, digital advertising tools, F&I, fixed ops data and more pulled into one place. Next comes behavioral capture (layer 2) because a CRM record only tells you who someone is, it doesn't tell you what they want right now. Pages browsed on your website, opened emails, and clicked ads all make the difference between an AI that knows a customer and one that's just reading a file.

Now, all of that information requires storage (layer 3), be it a data lake, a software platform, or a secure server. Storage is not free, so know what it costs to house every click and open before you flip the system on. Storage means nothing until the data is structured (layer 4), either. Dealership software is notoriously bad at agreeing on a format which means your CRM writes a phone number with parentheses and your DMS uses periods while your database quietly decides these are different people. Structure is what forces it into a single version of the truth.

Group 2: Identity and Rules

Identity and rules is where structured data turns into something you can actually act on, and where getting it wrong stops being an accuracy problem and becomes a legal one. Identity resolution (layer 5) is the quiet, unglamorous work of deciding that the person who clicked an ad on Tuesday is the same person sitting in your CRM from three years ago, every time, continuously. Get it right and your agent has a complete customer. Get it wrong and it has three strangers who happen to share a phone number.

From there, consent (layer 6) and permissions (layer 7) are what keep an agent from becoming a liability. The most common thing dealers want AI to do is talk to shoppers, which means you need a hard rule the agent can't ignore about who it's legally allowed to text or email. Permissions decide who the agent answers to at the dealership so it knows who's asking, and what they're actually authorized to see. None of these things ship with an MCP. All of it is a decision somebody has to build on, and a mistake somebody has to live with if it is built incorrectly.

Group 3: The Agent

No AI agent just arrives at your store knowing what "aged inventory" means to you. If nobody defines it, the agent will guess. While it may be an educated guess, it will guess wrong about as often as it guesses right, unless you build a semantic layer (layer 8) that picks one answer and sticks to it.

Tools (layer 9) are what let the agent actually do something with that semantic understanding like, "Build an audience of lease returners with high equity." Skills and evaluations (layer 10) are what keep all of it honest over time. Skills are the written procedure for recurring tasks so the agent isn't reinventing the task from scratch every single time. Evals are the background checks that catch it when, say, you switch providers and the agent quietly starts getting things wrong.

Group 4: Action

Action is the grouping everyone actually sees, and it's exactly where the "it's simple" pitch runs out of road. The chat interface (layer 11) for your AI agent is just the screen and it's the least impressive piece of the whole stack, which is exactly why every demo leads with it. Activation (layer 12) is what separates the toys from the tools. Real activation builds an audience, pushes it to your ad platforms, sends the text, or launches the campaign, all entirely on its own with no human standing in the middle doing the copy-paste job the AI was supposed to replace.

What an MCP Actually Gives You

An MCP gives you a sliver of the sourcing in layer one. Maybe a taste of the tools in layer nine, if someone's handy with a prompt. Everything else is still entirely yours to build which leaves you paying for the tool itself, the training, the 24/7 monitoring across every team touching it, and the security posture nobody inherited where every system wired up independently is now also independently breakable, bringing us all the way back to square one.

Evaluating Your Options

The vendor market is flooded with empty promises which means you need to know exactly what you are buying before you get started.

Option one is the Data Lake approach. A vendor will collect your data and store it, handling the basic storage layer. You are entirely responsible for identity resolution, rules, the semantic layer, and activation. If you choose this path, you need to have a massive internal engineering team in place to build the rest of the engine or your AI agent will never make it past the first query.

Option two is the CDPs and APIs approach. A CDP will collect your data, normalize it, and handle identity resolution. This covers groups one and two. However, you are still entirely responsible for the agent layer and the action layer. You have to teach the AI your semantics and build the execution tools yourself. Ask these vendors if their system respects your internal permissions or if it just blindly guesses from column names.

Option three is the complete off-the-shelf solution. All twelve layers are built and managed for you - CDP included - and your AI agent is an automotive-trained model that lives and breathes your dealership where data collection, identity resolution, semantic training, and activation are fully integrated. You just log in and use it.

No matter which option you choose, security guardrails are non-negotiable. Identity resolution, consent, permissions, and more don’t become optional based on the starting point you’ve chosen. The guardrails either become your responsibility to build, or someone else’s responsibility to prove they’ve already built them.

Pick the Layers, Not the Pitch

When it all comes down to it, you have to do what is best for your dealership and your team. A dealership with a real engineering team might genuinely want the data lake and build the rest its own way. A marketing team already living inside a CDP might be perfectly happy teaching the agent layer itself. What's wrong is walking into any of these conversations believing an MCP can do most of the work.

So before you sign anything, make whoever's selling it answer the following: which of these twelve layers are you actually handing me, which ones do I still need to build myself, and what is the entire package going to cost me? If they can't answer that in a single sentence, you already have your answer. A connection got you in the room, but it still has to earn your business.