Marketing teams adopted AI for drafting almost immediately, and remain reluctant to trust it with decisions about what to build or who to sell to.
That gap is the subject of this series. Generation is cheap now. Anyone can produce a competent post, a plausible positioning line, a passable campaign brief in seconds. The thinking stayed hard: understanding demand, choosing where to play, telling a good answer from a plausible wrong one in a domain with no clean scoreboard. The tools have flooded the easy, visible slice of the job and left the valuable work almost untouched.
This series walks that ground in four pieces. The short version of the argument comes first, and each section notes where the full piece takes it further.
Part one: the marketing value chain
Ask most people what marketing is and they say advertising, or social, or content. That is the loud, visible slice of a much bigger discipline. Marketing is the craft of understanding demand and organising a business to meet it profitably. Under the noise, the whole job runs on a spine: who is the customer and what do they need, what proposition do you build for them, and how do you deliver it and get paid.
Break that job into five pieces and you can see where the leverage sits. The five pieces are understanding demand, strategising where to play and how to win, creating the product and proposition, communicating through content and channels, and measuring what drove revenue. Almost every "AI for marketing" tool you have seen lives in the communicate box, generating the post or spinning the variant. Those tools are useful, and the space is crowded. The hard, high-stakes work sits in understand, strategise and measure, and that is where the room is empty.
Part one lays out the full value chain, the three growth levers every plan reduces to, and why automating only the output box scales activity without improving the decisions that make the output worth anything. Read the full piece: The Marketing Value Chain.
Part two: how AI changed each box
AI compressed cost and time in every box of that chain. A buyer study that took months, a positioning draft that took a strategy team weeks, and campaign variants that took agency rounds all collapse to minutes. That is real, and it is worth having. The failure mode is identical everywhere, though: a plausible output with no ground truth to check it against.
Walk the chain and each box has its own version. In understand, it is a clean-looking ideal customer profile that is confidently wrong: a buyer who looks real and is not yours. In strategise, it is a positioning draft that reads well and differentiates nothing. In create, it is polished, interchangeable copy. In communicate, it is spraying the wrong accounts faster than ever. In measure, it is a precise answer to the wrong question, optimising lead counts while revenue drifts. As generation gets cheaper, the bottleneck moves to one skill: telling good from plausible-wrong.
One conclusion follows from that: AI applied to a bad strategy scales the bad strategy. Part two gives you the before-and-after map and two questions to run against any AI workflow you build. Read the full piece: Before and After AI.
Part three: judging B2B positioning
Positioning is where the plausible-wrong problem bites hardest, because there is no clean answer to check against. Ask any model to position a B2B product and it returns a line like "the innovative, trusted partner for enterprise X". The line is fluent, and nothing on the page tells you whether it is any good.
There is a model for this, from a global professional-services brand study. B2B buyers move along a preference loop - aware of you, considering you, preferring you for one service, preferring you for many - and positioning's job is to move them rightward. The drivers that predict becoming a client's preferred partner are weighted and known. Strong relationships are the single biggest, at roughly a quarter of the whole decision. Favourability comes next, built from how it feels to work with you and from genuine distinctiveness. Awareness and a clever tagline sit near zero. The test is concrete: whether the positioning builds the high-weight drivers, and whether a competitor could truthfully claim the same line. If a rival could say it too, it is not positioning yet.
Part three covers the preferred-partner model in full, why B2B behaves differently from B2C - multi-party decisions, long lags, a shortlist of about three vendors - and how to use the test as a constraint every time you ask a model for a positioning line. Read the full piece: B2B Preferred-Partner Positioning.
Part four: the evaluation function
Pull the first three parts together and one idea sits underneath all of them. The hard problem of applied AI in marketing is knowing whether an output is any good, in a domain with fuzzy ground truth. The name for encoded domain judgement is an evaluation function, and a model does not have yours unless you give it to it.
The practical step is going from prompt to skill. A prompt is judgement you retype and re-trust every time; it is invisible, unversioned and inconsistent. A skill is that judgement written down once, version-controlled, with a gate baked in: a point where the tool stops and says "I cannot proceed, I do not yet have the ICP, the competitor I beat, or the demand I capture." Most tools would bluff a confident answer at that point. That refusal is the evaluation function made visible. The weighted driver model from part three is the pre-AI version of the same thing - explicit drivers, explicit weights, explicit rules - and it is a clean thing to hand a machine.
Part four is the recipe: name the failure mode, write the test that catches it, add a gate, keep the reasoning, version it. When generation is free, that discipline is the durable advantage. Read the full piece: The Evaluation Function.
Practical steps for marketing leaders
Three practical steps come out of the four parts. Point AI at the hard boxes - understand, strategise, measure - and treat the content box as the commodity it has become. Before you ship any AI output, name its failure mode and write the test that would catch it; if you cannot, you are shipping plausible-wrong at scale. Encode the judgement that matters to you once, as a skill with a gate, so it compounds with every run and stops living only in your head.
The generating is solved and everyone has it. The taste, the vision and the call on what is good stay with you, and they are worth more now than they were, because judgement has become the bottleneck. Use AI to improve your thinking as well as your speed.
If you want the frameworks from this series as working tools you can run yourself, they are in the marketing skills pack.