In short
- Agentic analytics is not one tool. It is five layers: Data → Meaning → Access → Thinking → Action.
- The AI model sits on one layer only. Everything else is plumbing — and the plumbing is where projects fail.
- Most projects fail in the middle, at Meaning. That is your company dictionary, and nobody can sell it to you.
- A person stays in the loop on every layer. The agent does the typing, not the deciding.
Why I drew this
For years I built marketing analytics the same way: GA4 (Google Analytics 4), Google Tag Manager, BigQuery, Looker Studio. Collect the data. Clean it. Build a report. Wait for somebody to open it.
In 2026 the last part changed.
Earlier this year I built a conversational AI agent for an SEO (Search Engine Optimisation) and content team, so people could ask questions in normal words instead of SQL. It worked. But the hard part was not the AI. It was everything under the AI.
The agent kept giving confident wrong answers. Not because the model was weak. Because it did not know what our team meant by "organic traffic", or which table was the good one. Nobody had written that down. A human analyst keeps it in their head. An agent cannot.
So I drew a pyramid. It helps me explain to clients why an agentic analytics project is 80% boring data work and 20% AI.
A note on the names
Most versions of this diagram say "tooling layer" and "orchestration layer". I used those words too, until a reader told me they were confusing.
He was right, and it is worth explaining why. Those names are different kinds of thing. "Reasoning" is something a model does. "Orchestration" is how you manage several agents. "Tooling" can mean almost anything. You cannot compare them, because they are not comparable.
So I renamed the layers. Each one now answers one plain question about the same journey:
| Layer | Name | The question |
|---|---|---|
| 1 | Data | Where is your data kept? |
| 2 | Meaning | What do your numbers mean? |
| 3 | Access | How does the agent read your data? |
| 4 | Thinking | Who finds the answer? |
| 5 | Action | How do you get the answer? |
If you have seen the industry words: Data = data foundation, Meaning = semantic and context layer, Access = tooling, Thinking = reasoning and orchestration, Action = the top of the pyramid.
Each layer needs the one below it. That is the only rule the pyramid is trying to show.
LLM, AI agent, AI platform — what is the difference?
People use these three words as if they mean the same thing. They do not. And they are not three sizes of one product.
The easiest way to see the difference is to think about hiring somebody.
The LLM — the brain
The model that reasons and writes text: Claude, GPT, Gemini, Llama.
Think of a clever new person on their first day. Fast. Has read a lot. Has never seen your company.
Ask how your ROAS (Return on Ad Spend) was last month and you will get an answer. It will sound good. It will be invented, because they had nothing to look at.
A bigger model does not fix this. When a new hire does not know but wants to impress you, they guess something that sounds right. That is exactly what a wrong AI answer is: confidence without anything behind it. A bigger brain only makes the guess sound better.
This matters when you are choosing tools. Moving from a mid-size model to the biggest frontier model does not give the model your data. It buys you a slightly better guesser.
The AI agent — the employee
The same brain, after you give it three things:
- A key — access to your systems, so it can look instead of guess. (Layer 3)
- A dictionary — what your company's words and numbers mean. (Layer 2)
- A job — something to finish, and a way to report back. (Layers 4 and 5)
Same brain. Now useful.
One more thing has changed recently. In 2026 an agent is rarely one person. It is usually a small team of specialists:
- Employee A, the analyst — pulls the ROAS numbers.
- Employee B, the checker — tests that work against the dictionary and looks for mistakes.
- Employee C, the writer — turns the checked result into a short summary for the manager.
This is why agentic AI works better than one very long prompt. The agents catch each other's mistakes before anything reaches you. A single model answering in one pass has nobody checking it.
The AI platform — the workplace
One supplier selling you several layers bundled together: the building, the filing cabinets, the locks and keys, and the person. Snowflake, Microsoft Fabric, Google Cloud, Gemini Enterprise.
The layers are horizontal. Agents and platforms are vertical.
An agent does not bring your dictionary — it reads yours. Nobody can sell you the definition of your own ROAS. A vendor gives you the place to write it down; writing it is your job.
So when somebody says "our AI agent will answer your business questions", ask one thing: whose dictionary does it read? If the answer is "it works that out automatically", you now know which layer will fail first.
The five layers in one minute each
Part 2 goes through the tools on each layer. Here is the short version, so the map makes sense on its own.
Layer 1 — Data. One place where everything lives, in tables an agent can read. Your GA4 events, Google Ads costs, CRM (Customer Relationship Management) deals and offline orders together, instead of five exports that never match. If you already use BigQuery, Snowflake or Microsoft Fabric, this layer is done.
Layer 2 — Meaning. The layer that explains your business to the agent. What "active customer" means. Which of the six revenue columns is real. Which table is old. It has three parts: a semantic model for your numbers, vector search and RAG (Retrieval-Augmented Generation) for your documents, and a catalog for your rules. This is the layer that decides whether the project works, and it is the one nobody can sell you.
Layer 3 — Access. The key. Without it, the agent can only talk. With it, the agent can look things up in your systems and show you where the number came from. In 2026 this mostly means MCP (Model Context Protocol), the open standard almost every vendor now supports.
Layer 4 — Thinking. Where the model lives — and only here. It splits a big question into small steps, checks each one, and fixes mistakes. Two parts: the model (the brain) and the framework that manages several agents (the manager). People confuse these two constantly, and I will untangle them properly in Part 2.
Layer 5 — Action. How the answer reaches a person. Nobody opens a dashboard; at 08:00 you get a message saying paid search cost per order rose 22%, which campaign it was, and what changed on Monday. Then you decide what to do about it.
Where to start
Three steps, in order. This is the part people get wrong most often.
Step 1. Write down ten definitions.
Not a hundred. Ten. The ones your leadership actually asks about. Put them in a semantic model — LookML, dbt, Cube, Snowflake Semantic Views, a Power BI semantic model. Whatever your stack already uses.
This is unglamorous and it is the highest-value week you will spend on this whole subject.
Step 2. Connect one agent to one question.
Pick a question you get every week. "What did we spend on paid media yesterday, by channel?" Give the agent read-only access. Test it against answers you already know are correct.
Do not go live until it gets twenty known questions right in a row.
Step 3. Add one alert.
Only after step 2 works. One metric, one threshold, one channel — and check who is in that channel before you switch it on.
Most teams try to start at step 3. That is how you end up with an expensive chatbot nobody trusts.
Glossary
Every word in this series, in plain English.
| Word | What it means | Layer |
|---|---|---|
| LLM — the brain | The model that reasons and writes text (Claude, GPT, Gemini, Llama). General knowledge, but it cannot see your company's data by itself. | 4 |
| AI agent — the employee | The LLM given a system login, a company dictionary and a specific job. It reads your Layers 1 and 2 — it does not supply them. | 3–5 |
| AI platform — the workplace | One supplier selling you several layers bundled together (Snowflake, Microsoft Fabric, Google Cloud). | 1–5 |
| Semantic model | Your business rules written once: what revenue is, how ROAS is calculated, who may see what. | 2 |
| Vector database | A store that lets you search text by meaning, not exact words. | 2 |
| RAG | The agent searches your documents first, then answers from what it found. | 2 |
| MCP | A free open standard for plugging an agent into a data source. Like USB, but for AI tools. | 3 |
| Multi-agent | Several narrow agents doing one step each, instead of one agent doing everything. | 4 |
| Open source | Free to self-host. Often has a paid cloud or enterprise tier alongside — that is normal, not a trick. | any |
| Open weights | You can download and run the model, under the vendor's own licence. Not the same as open source. | 4 |
| Human in the loop | A person approves before anything changes for real. | all |
Coming next
Part 2 — The tools on each layer. What Google Cloud, Snowflake, Microsoft, OpenAI and Anthropic actually sell on each layer, which free open-source tool does the same job, and whether open source is really free and really safe.
Part 3 — Privacy, security and the human in the loop. The risks that are specific to each layer, including the one almost nobody plans for: how a vector database quietly becomes a data leak.
Disclaimer
Informational only; not advice. Believed accurate as of August 2026, without warranty. Vendors change names, features and licensing — check official sources before deciding. Trademarks belong to their owners.
This article reflects my personal views, not those of any employer or client.
I write about data, analytics and AI at makskulish.com. You can also find me on LinkedIn.