There's a number that follows AI in marketing around like a stray dog: 71%. It shows up in pitch decks, LinkedIn posts and vendor webinars — usually as proof that AI "works" for marketing. It's real, it's sourced, and almost everyone quoting it has it backwards.

Here's what the number actually is. Stanford HAI's 2025 AI Index, summarising McKinsey's survey data, found that among organisations using AI in marketing and sales, 71% reported revenue gains — the highest share of any business function. Marketing AI's victory lap. Now the fine print: the most common size of that gain was under 5% of revenue, and it is self-reported. Nobody measured it. They believed it.

Meanwhile the 2026 measurement surveys tell the opposite story. Jasper's State of AI in Marketing (~1,400 marketers) found only 41% can confidently prove AI ROI — down from 49% the year before. Basis's August 2026 survey found 6.5% measure AI ROI against defined benchmarks, while 65% don't measure it at all or aren't sure anyone does. Comviva's Global CMO Survey found 90% of organisations increased AI marketing investment — and 12% can prove real impact.

That's the 71% problem. Not that AI doesn't pay. That most teams have no idea whether it pays — and are quoting someone else's number instead of finding out. With 2027 budget season open and C-suite pressure to prove ROI rising (Validity's November 2026 research puts it at 61% of marketers), vibes are about to get expensive.

Here's my read on why measurement fails, and the minimum reporting rigour I'd demand before signing off any AI budget.

Where the 71% Actually Comes From (and What It Doesn't Say)

Three things the celebratory version leaves out:

  1. It's adoption self-reporting, not accounting. "Reported revenue gains" means a human ticked a box on a survey. No finance sign-off, no baseline, no ledger.
  2. The gains are small. The modal gain is under 5% of revenue — which can be real money, or can be rounding error next to what the same org spent on seats, tokens, agents and rework. Without the cost side, "gain" is a half-written sentence.
  3. The direction of travel is negative. Jasper's year-on-year drop (49% → 41% able to prove ROI) came while adoption and spend rose. More AI, less proof. That's not a tooling problem; it's a governance problem.

Why Measurement Fails (Three Structural Reasons)

1. Costs are scattered by design. AI spend hides in SaaS seats, API tokens, agent run-times, tool sprawl and — the one nobody books — human review time. Comviva found 67% of organisations can't determine their total AI costs. You cannot compute return on a number you can't see. The hidden-cost problem is old news in automation — our honest marketing automation ROI method exists because vendor calculators pretend these costs don't exist. AI multiplied them.

2. Benefits get booked as activity. Hours saved. Content volume. "Engagement." Jasper found the most common AI success metric is hours saved (57%) — a metric no CFO has ever accepted as evidence of revenue. Activity metrics feel like progress and prove nothing. If the benefit isn't expressed in money, or in a pre-agreed proxy a CFO signed off, it doesn't go in the ROI case.

3. Attribution wasn't built for AI-era journeys. Buyers now move through dark funnels, AI-assisted research and agent-mediated shortlists before anyone fills in a form. Most teams' reporting stacks were designed for a click-path world — which is why the same measurement gap eats non-AI spend too. The fix isn't a better last-click report; it's the discipline in attribution ROI: proving which 30% of budget actually works — conservative, defensible, built for messy journeys.

The Minimum Reporting Rigour (What I'd Ask For)

Not enterprise. Minimum viable. Four artefacts, one page:

ArtefactWhat goes in itCadence
Cost ledgerFully loaded: seats, tokens, agent run-time, integrations, human review timeMonthly
Benefit ledgerOnly outcomes a CFO recognises: revenue, avoided cost, defensible time-value proxiesMonthly
Baseline vs actualPre-agreed metrics frozen before each use case launches — no retrofittingPer use case
Decision logKill / scale / hold decision per use case, with the number that drove itQuarterly

If a use case can't produce all four, it isn't a failure — it's a pilot. Pilots are fine; permanent unmeasured production spend is not. The structure borrows from how we'd run a 90-day AI pilot with go/no-go criteria: agree the thresholds before launch, freeze the baseline, and make "no-go" a legitimate outcome. The teams that do this at pilot stage don't have a 71% problem — they have receipts.

The Opinion Part: Budget Season Punishes Vibes

2027 budget planning is happening right now, with flat headcount and AI asking for a bigger share. Validity's November 2026 finding — 61% of marketers under C-suite pressure to prove ROI — isn't pressure on AI specifically. It's pressure on who gets funded. When the CFO asks "what did AI return?", the team with a ledger gets budget; the team with a McKinsey citation gets a pilot cap and an audit.

There's also a stack prerequisite: you cannot report what your systems can't measure. Disconnected tools mean the cost ledger is guesswork and the benefit ledger is double-counted. Score yourself against the connected-system maturity model before you promise the board a number — if you're below "measured" on the data dimension, fix that first or under-promise.

The uncomfortable summary: the 71% headline is true, and useless. Marketing is indeed the function reporting the biggest AI gains — in a world where most teams can't prove a penny. The competitive gap in 2027 won't be AI adoption. Everyone has adopted. It'll be proof.

Get Your Receipts in Order

If your AI line in the 2027 budget rests on "everyone says it works", that's the problem this article is named after. We build the boring version: a fully loaded cost ledger, a benefits ledger your finance team signs off, baselines frozen per use case, and a quarterly kill/scale discipline. Book a call and we'll pressure-test your AI ROI reporting in one session — before your CFO does it for you.