Volume Is Free. Correctness Is Scarce.
The shift AI actually made
For most of computing history, the bottleneck was output. More content, more analysis, more code required more people. AI changed that equation. Volume is now cheap. What it didn't change is what "right" looks like.
The core argument here draws on research and framing from Nate B Jones. The application to marketing infrastructure is ours — but the thinking originates with him. Watch the original video or read the Substack.
What AI actually made cheap
The way most people talk about AI — "it helps you do more" — is accurate but incomplete. Volume generation is now accessible to any team regardless of size. Content, code, analysis, reports — all of these can be produced at a rate no previous team structure could match. The bottleneck is no longer how much you can produce.
The bottleneck is whether what you're producing is right.
Why correctness didn't get cheaper
Correctness is a judgment problem. It requires holding a model of what the output is supposed to do — and comparing actual output against that model. That comparison is still a human task, and it's a harder one than it's ever been: the volume of output to verify has increased, the surface area of potential error has grown, and the confidence of AI-generated output often exceeds its accuracy.
A language model that writes a strategy brief sounds authoritative regardless of whether the strategy is correct. A coding agent that builds a tracking implementation looks complete regardless of whether the events fire correctly. An AI-generated attribution report presents clean numbers regardless of whether the underlying data is sound.
The outputs feel finished. The errors don't announce themselves.
The shared mental model problem
Correctness at scale requires something that volume doesn't: a shared mental model of what "correct" looks like. In a small team where everyone holds the same context, verification is fast and calibrated — each person knows what the work is supposed to do and can catch meaningful deviations.
In a larger team with distributed context, the same volume of output produces much higher verification overhead, because no single person holds enough context to efficiently check any given piece. The team optimizes for volume. The small team optimizes for correctness.
A 2025 Harvard Business School field experiment confirmed the pattern from a quality angle: teams using AI produced higher-quality ideas — not just more ideas — and the effect was strongest when shared context was tightest.1 Volume scales with team size. Correctness scales with shared context.
What this means for marketing infrastructure
In marketing, the volume/correctness divide is especially visible. The volume side is obvious — more content, more ad variations, more reports than ever before. The correctness side is where most teams struggle.
Attribution
Is the attribution model actually reflecting how customers convert — or is it crediting the last touchpoint because that's the default? Duplicate conversion events can inflate counts by 30–50%. The numbers look fine. The data isn't.
Strategy
Is the positioning sound — or does it feel coherent because the language model made it sound that way? AI-generated strategy briefs are confident regardless of whether the strategy is correct. The output feels finished. The errors don't announce themselves.
Brand
Is the content actually on-voice — or has drift accumulated across AI-assisted drafts? Voice drift is gradual. Each piece passes review individually. The pattern only becomes visible when you look across the whole body of content.
Trust & Security
Does the AI-assisted content and data handling meet the actual regulatory standard — not just the interpreted one? Consent mode that looks configured. Pixels that load regardless. The system presents compliance. The infrastructure doesn't deliver it.
The pattern: Volume doesn't answer any of these questions. Correctness requires an independent check — a set of eyes with enough context to compare the output against what it was supposed to do. That check is still a human task, and it's harder to do at higher output volume with weaker shared context.
2 The 30–50% attribution inflation figure is drawn from our own audit findings across client engagements. See also: The Perception Gap Is Real .
Correctness by design
This is what the spec-driven approach is designed to protect. A Blueprint defines correctness before the build begins — it specifies what the system is supposed to do, what the data flows are, what the success criteria look like. The build is evaluated against the spec, not just against whether it produces output.
Monitoring is the ongoing enforcement of the correctness standard over time. Alert protocols, quarterly reviews, and performance benchmarks aren't management overhead — they're the mechanism that catches the difference between a system that looks correct and one that is.
The forensic diagnostic is a correctness check applied retroactively: what is this infrastructure actually doing, against what it should be doing? The perception gap — the difference between what teams believe about their infrastructure and what it's doing — is the accumulated cost of correctness debt. It compounds quietly until it doesn't.
Notes & sources
- 1. As described in Nate B Jones's analysis (linked above). The study is a Harvard Business School field experiment published in 2025 examining AI's effect on idea quality in cross-functional teams. Key finding: AI-assisted teams were significantly more likely to produce high-quality ideas — not just higher output — and the benefit was most pronounced when team members shared working context.
- 2. Drawn from our own audit findings across client engagements. The 30–50% range reflects the gap between platform-reported conversions and independently verified conversions — a pattern we see consistently, not a figure from a single study.
What's your correctness gap?
The forensic diagnostic surfaces your tracking accuracy, attribution reliability, and compliance posture across ten marketing pillars. Five minutes, at no cost.