Measuring AI Marketing ROI: A Framework That Survives the CFO

"What's the ROI of our AI spend?" is now a standing budget-season question, and most marketing answers to it are either vibes ("the team loves it") or theater (multiplying saved minutes by salary and calling it millions). Both get demolished in the room. Here's a framework that doesn't.

Measure in three tiers, honestly labeled

Tier 1 — Time. The easiest to measure, the easiest to inflate. The honest version: measure task time before and after on specific recurring workflows (the weekly report, the campaign brief, the content draft), and count only time that was genuinely redeployed to other work. "The team saved 400 hours" is only a benefit if you can say what those hours produced instead. Time savings that became slack are real but modest; claim them modestly.

Tier 2 — Throughput and quality. More assets shipped, faster cycle times, fewer errors, more variants tested. These are measurable and closer to business value: creative testing velocity has a known relationship to ad performance; error rates in sends have a known cost. This tier is where most durable AI value actually lives, and it's chronically under-instrumented. Before-and-after baselines are mandatory — decide the metric before the rollout, or the after will be retrofitted.

Tier 3 — Outcomes. Pipeline, revenue, conversion, retention. The tier everyone wants and the hardest to attribute honestly, because AI is rarely the only thing that changed. The credible approaches: holdouts (one region, segment, or team runs without the AI workflow), stagger the rollout and compare cohorts, or claim contribution rather than causation with the mechanism spelled out ("refreshed content via the AI pipeline; those pages' conversions rose X% versus the untouched set").

Count the full cost side

The line item that discredits AI ROI math is the one that's missing. Total cost includes:

• Subscriptions and usage — the visible spend, including API/token costs that scale with volume.
• People time — building, maintaining, and reviewing. Human-in-the-loop review is a real recurring cost; leaving it out overstates ROI on every workflow that has it.
• Drift and maintenance — re-testing after model updates, fixing broken integrations. Budget it; it happens regardless.
• The failed experiments — pilots that went nowhere are part of the program's cost. Mature teams report them as R&D, not as embarrassments to hide.

The presentation format that works

One page per major workflow: what it does, cost (full), Tier 1–3 benefits (each labeled by confidence: measured / estimated / directional), and the decision it supports (expand / hold / kill). A portfolio of honest one-pagers — including a killed workflow or two — builds more budget credibility than any aggregate "AI saved us $2M" slide, because it demonstrates you're managing the program rather than marketing it internally.

The uncomfortable rules

1. If you didn't baseline it, you can't claim it. Retroactive baselines are fiction with a spreadsheet.
2. Adoption is not a benefit. "80% of the team uses AI weekly" measures behavior, not value. It belongs on the slide only as context.
3. Kill metrics belong up front. Define what result would make you stop each workflow. Programs with kill criteria get trusted; programs where everything "works" get audited.

Start by instrumenting your two biggest AI workflows this quarter — real baselines, full costs, one page each. Two honest pages beat twenty enthusiastic ones every budget cycle that will ever occur.