The Marketing Data Stack for AI Agents: Six Jobs and the Read-Wide Rule

What is a marketing data stack?

Most teams start the same way. They hand an AI assistant the keys to the ad accounts and ask which campaigns made money last quarter.

It answers in ten seconds. It sounds sure. Nobody can check it.

Then someone asks it to pause the losers. Now it spends real money on a guess.

A marketing data stack is the set of tools that lets an AI read your numbers and act on them without guessing. We describe it by job, not by vendor. Vendors change every year, and the jobs do not.

The whole stack hangs on one rule. We call it the Read-Wide Rule: the AI reads everything from a copy you own, and changes things only through one door, one approved step at a time.

This is the plumbing under the job in what a marketing engineer is. It is not a warehouse course for data engineers.

Name the stack by its six jobs, not its vendors

Every marketing data stack does six jobs, whatever the logos say. Name the jobs first and the tool choice gets easy.

Collect means copying data out of the ad platforms, analytics, the CRM and payments, a step data people call ETL. Store means keeping one copy you own in a warehouse, a database built for questions across many sources.

Run on a schedule means the work happens while you sleep. Version means every script, prompt and setting keeps its history.

Connect is the one door every outside call goes through. Check means watching the other five and telling you when one breaks.

Why jobs and not tools? Tools get bought, renamed and shut down. A job outlives all three.

Your first stack does not even need software. The Team Topologies authors say the simplest platform can be a list on a wiki page.

So write the six jobs down, with a tool and an owner next to each. That page is version one. No code yet.

Why should the AI read from a copy you own?

The read side has one rule: the AI asks your copy, never the live ad platforms. Three reasons.

First, platforms ration their APIs, the lines programs use to talk to them. Pull ninety days of rows for every question and you burn the whole daily allowance just reading.

Second, an AI holds only so much in its working memory. Paste in thousands of ad rows, join them to analytics by hand, and you get a confident answer built on a bad join.

Third, two people asking the same question should get the same number. That only happens when both read the same copy.

The copy needs shared keys. Data modelers call them conforming dimensions: fields such as date and customer ID that mean the same thing in every table.

With shared keys, you can follow one buyer from ad click to closed deal. Without them, you guess.

Storage is the easy part now. The hard part is asking a question worth answering.

Run it on a schedule, and let the system report its own failures

A stack earns its keep while you are at lunch. That takes a clock: a scheduled job that collects, stores and checks at the same time every day.

Order matters. Each step runs only after the step before it finished cleanly, so a failed import never feeds a report.

Workflow tools call that chain a directed acyclic graph. In plain words, work only moves forward, and only on good input.

Then make the stack report its own breaks. Who usually finds a broken lead form? Sales, on Monday, when the pipeline looks thin.

The Accelerate research team studied how software teams learn about failures. Some heard it from customers, others from their own alerts on a threshold or a rate of change. The alerts predicted better software delivery.

We apply the same logic to marketing. For example, a threshold alert fires when fewer than ten leads arrive in a day. A rate-of-change alert fires when leads drop by half against the same weekday last week.

A dashboard you have to remember to open is not a check. It is a hope.

Version the settings, not just the scripts

Version means every change has a date, an author and a way back. Most marketing teams skip this job, then miss it the week cost per lead jumps.

The same research had a surprise. Keeping configuration in version control was more strongly tied to delivery performance than keeping the code there.

Marketing is mostly configuration. Tracking tags, UTM rules, automation triggers, audience definitions and AI prompts are all settings, and settings drift.

So the stack keeps a written history of each one. When results move, you ask what changed that week. You get an answer in a minute, not a meeting.

The AI's own work gets logged too. Chip Huyen's rule in AI Engineering is to log everything: the model, its settings, the prompt, the output and every tool it called.

You never know which line you need until something goes wrong. So read a few real outputs by hand. Do it daily.

Logs nobody reads are just a storage bill. The growth repo shows one folder layout for all of it.

What may an AI change? Climb the ladder one rung at a time

An AI may change only what a written list allows, and each item sits on one of four rungs. A read tool lets the AI look. A write tool lets it act: pause, upload, send, spend.

Where does each action go? The textbook that coined marketing engineering has the rule. Automate a decision when a wrong call is cheap, like a keyword bid, and keep a person in it when a wrong call is expensive.

Rung one is read: reports and queries only. Rung two is draft: the AI may write, but the work lands somewhere harmless, like a paused ad or a proposed change you confirm.

Rung three is capped: live changes inside limits enforced in code, such as a budget move under a set amount. Rung four is live, and a named person approves each change that spends money or reaches a customer.

Every connection starts on rung one, even the ones a vendor calls safe. Climb a rung only after a clean log on the rung below. No exceptions.

Make stopping easier than starting. Good experiment platforms let only the owner start a test, but anyone can stop one. Give every AI connection the same off switch.

One mix is never allowed: private data, text written by strangers, and a way to send data out, all in one AI job. Simon Willison calls that mix the lethal trifecta.

With that mix, one planted line in a form field can send your CRM to a stranger. Keep at least one of the three out of every job, and keep the most private records on a local model.

Send every outside call through one door

Connect is a door: one small service between the AI and everything outside. It holds the keys, checks the list, enforces the caps and writes the log.

Why one door? A rule in a prompt is a suggestion. A rule in the door's code is a limit.

Keys never sit in a prompt or a chat. They live in one secrets file or the platform's secret store, and only the door reads them.

The door also counts cost per call, which is how you keep AI token spend from surprising you.

Add a kill switch: one setting in the door that turns every write off at once. Test it before you need it. Fire drills beat fires.

That door, plus the list and the caps, is the core of an agent harness. The model thinks. The door protects.

Buy the read side, build the write side

Our default is short: buy the read side, build the write side. Moving data is a solved problem, and a homemade version will be worse.

Writing is where your rules live. Which ads may pause, which cap holds, what needs a person: no vendor knows that, so you write it.

In Traction, Gino Wickman tells a story about building too early. He paid a firm $45,000 over months for custom software, scrapped it, and switched to a $500 product off the shelf. The mistake, by his own account: he never studied the process first.

So do the job by hand before you automate it. We run a new report by hand for three weeks, then automate the version that survived. Boring, and a lot cheaper than $45,000.

How to spend six months on a stack nobody uses, in four easy steps. Step one: buy every tool on a top-fifty list.

Step two: build a custom dashboard before anyone asks a question. Step three: give the AI admin keys so it can help.

Step four: discover the warehouse holds three definitions of a lead. Congratulations, you built a very expensive way to argue.

Before you sign with any bundled vendor, check one thing. Can you export your data and your scripts the day you leave?

Grow the stack only when a system asks for it

Start with collect and store, and nothing else. Point the AI at the copy and ask the questions you already ask every Monday.

Add the schedule and the checks when a question repeats every week. Add the door and the ladder when an action repeats.

The first thing worth building on the copy is the growth cockpit, a weekly memo that replaces the status meeting. The most valuable early write is usually sending real outcomes back to the ad platforms.

Data in a dozen tools today? Our piece on fragmented data explains why you search it where it lives before you move anything.

Each job should arrive because a system asked for it. A vendor demo is not a request.

A stack built ahead of its questions is a museum.

Vendor names and prices change faster than this page. The box below holds today's examples, with the date we checked them.

Map your own stack with AI

The prompt below maps your tools to the six jobs and puts every AI connection on a rung. Paste it into your preferred AI and answer its five questions.

This week: run it, then turn on collect and store for your two biggest sources. Put every AI connection on rung one.

Next week, ask your Monday questions of the copy. Climb a rung only when the log is clean.

Want the first system built next to you? That is our marketing engineering work.

Rather build it yourself? Our training runs on live work.

Remember the AI that answered in ten seconds? With a copy to read and a door to write through, it can show its work.

Give it a library card long before you give it a credit card.