Marketing software thinks now. It still won’t make the call.
Every marketer runs the same questions on a loop:
Should more budget go into Google, and how much before it stops paying off? Is that creative fatigued, or is the audience saturated? Two different problems, two different fixes, and the top-line numbers look the same for both. What drove last quarter's number? What was ROI across everything, in one figure you can stand behind? And what do I tell my boss on Monday?
At Meta, marketing leaders had a team of data scientists for questions like these. You have a login, a dashboard, and lately a chatbot.
I know because I was on that team. I spent years as a data scientist on Meta's marketing side doing incrementality, attribution, and diagnosis. The job took years of statistics, a lot of programming, and a master's in analytics from Carnegie Mellon.
All of that was the price of admission. The job itself was recommendations: telling executives what to do, and why. And that's the lesson that stuck with me. The data, the pipelines, the analysis, the interpretation, none of it was worth anything until it became a decision. Everything upstream was just a cost. Including me.
Showing, then thinking
For a long time the problem was just tabs. Your data was spread across platforms, dashboards, and spreadsheets. That never fully goes away, but the problem is understood, and there has been a generation of tools built to fix this, to show you your data.
Then AI showed up and the tools learned to think. You can ask a question now and get actual analysis back, not just a chart. The excitement around this is grounded because having tools that can think about your data is extremely powerful and is changing the way marketing teams operate. There is a lot more to say about that shift, enough for its own letter. This one is about what comes after.
I saw this enthusiasm first hand last week at Advertising Week NYC where sessions with "AI-Native Marketing Organization" or "Agentic Agency" in the title were standing room only. The least attended session I went to, with maybe a dozen people, was called "Can Your Data Infrastructure Defend the Decisions Agents Make". That's where the industry's attention is: optimizing workflows, automating analysis, deploying agents.
There is little discussion about whether the decisions coming out the other end hold up. And as I learned at Meta, the value is in the decisions. So let me ask you: are you making better decisions, or are you getting more answers, faster?
Because those are not the same thing. The first generation of tools got you better access to your data and this generation is getting you better access to insights. Both are useful, but neither one makes the call, and nobody has built the thing that does.
That's the gap, and it's not by accident. A decision has a bar, and meeting it forces requirements that the current tools weren't built for. That's where the value is, and that's what I left Meta to build.
What a decision takes
At Meta, a recommendation handed to an exec had five parts, each feeding the next. Anything less was an observation, not a call. Each part forces a requirement on the tooling:
Observation. It arrives before you ask. Most weeks the problem isn't that you can't get an answer, it's that you don't know which question matters most right now. So the software has to watch every campaign on every platform, every day, including the ones you'd never think to check. An observation you had to request is a search result. Requirement: proactive and autonomous.
Diagnosis. Grounded in statistics, every run. Same answer every time you run it; if it changes between runs, it's a guess with confidence. And the number has to pass a test before it earns a sentence; if it doesn't, it's noise with a narrative. Fatigued creative and a saturated audience show the same falling CTR and rising CPA. One wants new creative; the other wants a bigger audience, and new creative just fatigues faster. Telling them apart takes computed statistics, applied the same way twice. Requirement: deterministic and tested.
Recommendation. Made at the level you operate, and loyal to one thing: your objective. You manage one budget across several platforms, so a call has to weigh all of them at once. "Move spend to Google" will never come from Meta; no platform recommends spending away from itself, and no platform can see the others. And the call has to answer to your KPI, not to the platform's revenue or the tool's own metric. Every call checked against every other, so the guidance never contradicts itself. Requirement: independent and holistic.
Rationale. It shows its work, with evidence, in plain language. An argument you can check, not a conclusion you have to trust. Every sentence traces to a number, and the words exist to make the numbers readable, not to stand in for them. Requirement: transparent and legible.
Headroom. The move, priced. What each call is worth is what lets ten calls be ranked, and ranking is the job. Then the number gets checked against what the KPI actually did, because an unchecked projection is decoration. A checked one is a grade, and every grade sharpens the next call. Requirement: accountable and measurable.
Any tool missing one of these can inform you; it cannot decide with you. Score your current stack against these requirements. Then ask where its last ten decisions are written down.
Meet Aviio
Aviio's home screen is not a dashboard. It's a Decision Queue: the moves that matter most across your entire spend, ranked by impact, each carrying the same five parts a Meta recommendation had.
A card looks like this:
Recommendation: Shift $1,800/mo from Meta prospecting into Google Search non-brand. Observation: non-brand Search CPA down 34% over 14 days while budget and impression share sat flat. Diagnosis: auction pressure eased, not a tracking artifact and not brand-term cannibalization: spare efficiency in a contested auction. Rationale: the efficiency curve holds through +40% spend at high confidence, and the finding is identical on every run. Headroom: up to +11 conversions/mo at current blended CAC.
You approve it, or mark that you already did it, and Aviio tracks what happened either way: the KPI in question, before and after, checked against what the card projected. That grade feeds back into calibration, so the engine gets sharper with every decision it watches land. Every card carries a confidence level and your exposure, and when the engine isn't sure, the card says so.
Under the hood, the diagnosis is computed, not written. Baselines, significance, effect sizes, maintained daily against your own history, the way a data science team would keep them. The explanation on the card is written after the numbers exist, from the numbers. Run it twice, same finding. Ask why it fired, and it shows you the math.
See your own queue
Aviio is live. $5.37M of real ad spend has already run through the engine, across accounts synced from Google, Meta, LinkedIn, and TikTok. If you spend across platforms, book a walkthrough. Connect your accounts and I'll show you the top recommendations Aviio finds in your own spend, with the diagnosis and headroom attached.