A quarter's spend gets committed on a plan, and the numbers that would have changed the plan arrive at the end of it. By then the money is already placed.
Marketing AI automation means the numbers that decide a campaign arrive before the campaign does. Past performance is used to forecast what a new variant will do, and the forecast comes with a range rather than a single figure. A person still signs off the spend.
Marketing decisions get made on last quarter's numbers because this quarter's are not in yet. The work that would tell you which half of the spend is working is the work nobody has time for. So the same guess gets repeated at a larger budget.
Numbers get pulled out of four platforms into one sheet, and somebody does it every Monday.
A thousand search terms get sorted into groups by a person reading them one at a time.
Assets are produced to a schedule rather than to a gap, so some of them never run at all.
A model trained on your own campaign history estimates what a new variant will return, and it reports a range rather than a number. Where the history is thin it says so.
A few thousand keywords get grouped by what the searcher actually wants rather than by the words they share. The groups map onto pages, so the clustering produces a plan instead of a spreadsheet.
Numbers come out of the ad platforms and the CMS on a schedule and land in one place. What changed and why is drafted for a person to check rather than written from scratch.
Four steps. A marketer sits at the last one. The budget is theirs to commit.
Spend and results come out of the platforms.
Terms and creatives get grouped by intent.
A model estimates what each group returns.
A marketer reads the range and decides.
Human checkpointOnce a stack can forecast a result and group a thousand terms, several other jobs turn out to be the same two jobs.
Three answers a marketing team needs before anything else. A keyword strategy is the one asset an agency cannot get back once it leaks.
Keyword sets and spend figures never train a model. An open weight model on your own hardware means the numbers never leave your account at all.
Every estimate records what it was trained on and how far it has been wrong before. A number with no error attached is what loses money.
Consent is recorded at the point it is given, with what was agreed to and when. That is what a regulator asks for and it is also what makes a list worth having.
These are the pieces a marketing build tends to need. Your own ad accounts are the starting point, and nothing on this list is a requirement.
Good enough to rank options, not to promise a number. A model trained on your own history will separate a strong variant from a weak one long before it can tell you the exact return. Anybody quoting precision on an ad forecast is selling you the part that does not exist.
No, and it changes what you pay them for. The sorting of ten thousand terms is machine work and the judgement about which clusters are worth a page is not. An agency that only did the sorting was already exposed.
Yes, and we will argue about how much you publish. Google has a name for pages generated at volume, and the name is scaled content abuse. The version that works is one draft per cluster, finished by a person who knows the subject.
No. The models sit beside the platforms and read them through their APIs, so the accounts you run stay the accounts you run. What changes is where the decision gets made, which moves from the console to a forecast you can argue with.
Enough that the model has seen a bad quarter as well as a good one. A year of spend across a few channels is workable and six weeks of one campaign is not. A model fitted to a lucky run forecasts luck. Where it is short we say so early.
It is a contract question rather than a technical one, and it should be settled before anything is trained. A forecasting model fitted to your spend history is worth more than the code around it, and a supplier who keeps it holds your strategy. Ask it of everybody you talk to, including us.
Data entry, answering the same tickets, chasing numbers between systems. We automate the parts that repeat. Your team keeps the parts that need judgement.
