← Data Into Labor · All essays

Data Into Labor · Essay 02 · V2

The Next Dollar

July 2026 · AJ Orbach

Suppose you had to build a media buyer from scratch, human or machine, and you could hand them anything on day one. What would they actually need? Not what software exists. What the job itself demands.

I bought media for a living before we built the machine that buys it, so the question isn't hypothetical for me. The job, stripped of its jargon, is one repeated decision: where does the next dollar go, and will it come back with friends? It turns out that decision decomposes cleanly, requirement by requirement, into a stack of instruments. This essay is the inventory. It also happens to be a map of Triple Whale, because we built the company in the order the job demands, whether we knew it at the time or not.

The reason to write the inventory down now is who's about to sit in the seat. We're building Moby to do this job, not to summarize it, and an agent is exactly as good as the instruments you give it. A brilliant model with untrusted data is a confident intern with a broken compass.

Trust

Start with the buyer's first morning. They open five ad managers, and each one is delighted with itself. Meta claims the order. Google claims the same order. TikTok, feeling left out, claims it too. Add up what the platforms report and you'll routinely find more conversions than the store actually had, because every platform grades its own homework and none of them can see the others' work.^[1] A buyer who allocates against self-reported numbers isn't allocating. They're transcribing each platform's opinion of itself.

So the first instrument is multi-touch attribution: an independent record of the customer's actual path, owned by the merchant instead of the auction. The Pixel issues first-party identifiers, stitches sessions and devices into one journey per human being, and lets you read that journey through whichever model fits the question.^[2] No single model is sacred. What matters is that the journey underneath the models is yours, verified, and consistent, so that when two channels argue over an order, somebody neutral was in the room. Every later decision inherits its quality from this layer, which is why it's where we started as a company and where anyone building a media buyer has to start.

The feed

The next requirement is broader: everything, from all the platforms, along both directions of time. Real time, because the job is intraday: a launch going sideways at 9 a.m., a budget cap about to bind before lunch, a creative that fatigued overnight. And historical, because context is the difference between a number and a signal: this CPM against last month's, this launch against last year's Q4, this cohort against the one before the price change.

This is the data layer from the previous essay doing its job: sixty-plus native connectors landing in one schema, normalized so "what did I spend" has exactly one answer, at a scale where a three-year question comes back in seconds.^[3] The consequence for the seat is that the buyer stops being a librarian of tabs. And the form matters as much as the coverage: a screenshot of yesterday is trivia. A governed table that's current as of this morning is an instrument an agent can act on.

Weather

Here's a question I asked nearly every morning when I bought media: is it me, or is it Meta? ROAS dipped, CPMs jumped, acquisition got expensive. Two explanations, opposite prescriptions. If it's me — my creative fatigued, my landing page broke, my offer went stale — the fix is work. If it's Meta — the auction got expensive for everyone this week, a holiday is inflating demand for attention — the fix is patience, and the worst possible move is tearing apart a healthy account because the sea got rough.

Without an outside reference you cannot answer that question. You'll kill winning campaigns in a market-wide squall and scale losers in a market-wide tailwind, and both mistakes will look disciplined from inside your own ad account. What the buyer needs is weather instruments, and the reference has to come from a dataset: thousands of brands measured the same way at the same time. Peer benchmarks compare you to your ten nearest neighbors in vertical, scale, and motion, which is the honest comparison; a $50M apparel brand and a $2M supplement brand do not sail the same sea. Industry benchmarks zoom out to the whole category.^[4] If your CPM rose 30% and your ten neighbors' rose 28%, that's weather; hold. If yours rose 30% and theirs is flat, that's you; get to work.

For an agent this instrument isn't optional. An AI media buyer without benchmarks does what every junior buyer does: over-attributes the market to itself, panics at weather, takes credit for tailwinds. The benchmark layer is how the machine learns humility.

The shelf

Now a requirement so obvious it's almost embarrassing that the industry ignores it: the buyer should know when you're running out of inventory. The ad platforms have no idea what's in your warehouse. They will cheerfully spend $40,000 tomorrow scaling demand for a SKU with nine units left, producing the most expensive kind of marketing there is: marketing that works, for orders you can't ship.

The join that fixes this already exists in the schema. The journey knows which ad produced which order, the order knows its SKUs, and the warehouse knows its units. Connect them and the creative itself becomes inventory-aware: this video is selling the black colorway, the black colorway has eleven days of stock at the current run rate, throttle accordingly.^[5] It works in the happy direction too: deep stock on a hero SKU is a green light a buyer can lean on with confidence.

Clocks

Not every instrument on the desk is a daily instrument. Some are chronometers, consulted when you're deciding what kind of company the budget is building. LTV and AOV per channel are the first two. A channel that looks expensive on first order and cheap on the twelve-month view isn't expensive; it's front-loaded. TikTok might bring $38 first orders that never return while Google brings $52 orders that reorder four times by month twelve, and a buyer staring only at first-order ROAS will methodically starve the better channel.

The second pair: which channels bring new customers, and which are efficient at it. Blended numbers hide this completely, because a channel can post a gorgeous ROAS by harvesting people who already loved you. New-customer CPA and new-customer ROAS split the ledger honestly, and they're only computable with first-party history, because knowing a customer is new means knowing every order they've ever placed, across years and across sales platforms. The ad platforms can't see that history. The warehouse can. None of this changes what you do at 9 a.m. tomorrow. It changes what you do this quarter: which channels get patience, which get pressure, which get retired. The daily instruments keep the boat off the rocks. The clocks decide where the boat is going.

Hands

Everything to this point is reading, and the job is not reading. The job is moving money. A buyer who can see everything and touch nothing is a critic, and brands don't pay for critics. So the stack needs hands: budget changes, pausing, scaling, creative rotation, across all the channels, from the same place the evidence lives.

The moment you say "hands," especially agentic hands, you need a governor. This took us the longest to take seriously, and we now consider it non-negotiable. Every action carries a permission model for what this agent may touch, an approval gate for anything consequential, and a log you can audit on Tuesday: what changed, by whom, on what evidence, reversible how.^[6] The approval queue is to an AI media buyer what the co-signature is to a young trader: the thing that makes real capital allocation possible at all.

The seat

Every buyer worth hiring arrives with opinions: read windows before touching a budget, maximum daily change, the promo calendar, what "expensive" means in this category. In a human, that lives in the head and walks out the door with it. In Moby it lives in skills: written playbooks the agent runs the same way every time, reviewable, editable, versioned. Codify the operating policy once — never move more than 20% a day, never touch brand campaigns without asking, always check the shelf before scaling — and tribal knowledge becomes infrastructure.

The Media Buyer Specialist is where all of it gets a seat: the same harness, focused, with the attribution layer, the feed, the weather instruments, the shelf, the clocks, the governed hands, and your skills managed as one operator. One seat is not a convenience feature. Fragmentation is where allocation errors breed: the buyer who saw the benchmark but not the inventory, the agency that saw the ROAS but not the cohort. One decision-maker, seeing all the instruments at once, is the only condition under which the next dollar gets placed well.

And when the brand gets genuinely complex — retail alongside DTC, international, channels feeding each other — you don't replace attribution; you add layers above it. MMM reads the budget's shape top-down. Incrementality experiments interrogate single decisions causally. That's Compass, and it belongs in the buyer's stack because of triangulation: three methods that fail differently, so where they agree you can bet heavily, and where they disagree is precisely where your next test belongs.^[7]

The job

Notice what the list is: a job description, written as infrastructure. When people ask whether an AI can "do media buying," they're usually asking about the model, and the model is the least scarce part. The scarce parts are the trusted journey, the unified feed, the reference dataset, the inventory join, the first-party memory, the governor, and the seat. No frontier lab is going to build a peer benchmark from thousands of brands, or a Pixel-stitched join from an ad to a shelf. That's not what labs build. It's what we build.

So call the assembled desk what it is: an automated media buyer with a human sitting on top. The machine runs the instruments; the person verifies the work and sets the strategy. That's not a consolation prize for either side. It's the correct org chart for the job. And on the instruments, the machine wins for the least mysterious reason imaginable: it's always there. No human buyer ever operated this whole desk at once. Nobody re-verifies attribution before every budget move, checks the weather hourly against ten peers, cross-references the shelf before every scale-up, and files the log entry every time, at 2 a.m., in the middle of Black Friday, on vacation. That was the quiet tragedy of the job: we built instruments a person could only sample, then graded the person on the samples. An agent samples nothing. It reads everything, every time, and it never has a morning where it skips the checklist.

What comes out the other side is the only bottom line that matters: more money, and more peace of mind. More money, because dollars stop leaking through the gaps between instruments — the winner killed during a squall, the loser scaled into a stockout, the acquisition channel starved because nobody re-ran the LTV math. Peace of mind, because every move arrives verified, logged, and explained, and the person upstairs approves the work instead of assembling it. One shows up on the P&L. The other shows up at eleven o'clock at night.