Guide · 5 min read
How to Calculate the ROI of AI in Marketing
By Fredrika Frenkiel, Head of Studio & AI Creative Operations at Lunar, founder of Code of Alfred. About my work
Short answer
You calculate the ROI of AI in marketing per workflow, not per tool. Baseline one workflow before you change it — hours spent, revision loops, cost per asset, time from brief to delivery — then measure the same workflow after redesign and subtract the full cost of the change, including licences, build time and training. Anything measured at the level of 'the department uses AI now' is unprovable and will not survive a budget review.
1. Baseline before you change anything
The most common reason AI investment cannot be defended is that nobody recorded what the work cost before. Spend one or two weeks measuring the workflow you intend to change: hours by role, number of revision rounds, elapsed days from brief to delivery, error and rework rate, and external spend. If that data does not exist, a rough estimate agreed by the team is far better than nothing.
2. Count all four kinds of return
Time saved is the obvious one, but the honest picture has four parts: hours released, cost avoided (external production, agency adaptation fees, licence consolidation), quality gains (fewer errors, fewer revision loops, more on-time launches), and capacity gains (work that becomes possible at all, more markets, more variants, faster response). Capacity is usually where the largest commercial value sits and the part most often left out.
3. Include the real cost of the change
Licences are the small part. Count the build or configuration time, the internal hours spent designing the workflow, training and onboarding, the ongoing maintenance owner, and the productivity dip during the transition. An ROI number that only subtracts subscription cost is not credible to a CFO.
4. Convert hours into something the business recognises
Hours saved are only value if they are redeployed. State explicitly what the released time is used for: more output, faster response, work brought in-house, or headcount avoided. In an agency the same hours may translate into margin or into billable capacity. Make the destination explicit, otherwise the saving is theoretical.
5. Report it as a before and after on one workflow
A believable ROI case is narrow and specific: this workflow, this period, this measurement, this cost. One proven workflow with real numbers wins more budget than an ambitious department-wide estimate, and it gives you the pattern to extend to the next bottleneck.
How to do it, step by step
- 01
Baseline the workflow first
Record hours, revision loops, elapsed time and external spend before changing anything.
- 02
Count four kinds of return
Measure hours released, cost avoided, quality gains and new capacity, not time alone.
- 03
Subtract the full cost of change
Include build time, design hours, training, maintenance and the transition dip, not just licences.
- 04
Say where the saved time goes
Name the redeployment: more output, faster response, in-housing or avoided headcount.
- 05
Report one workflow, before and after
Present a narrow, specific case with real numbers rather than a department-wide estimate.
Measurement is not an afterthought to an AI project, it is the part that lets the next one get funded. Build the loop into the design from day one. See how I approach this in practice.
Common questions
- How do you measure ROI on AI in marketing?
- Per workflow. Baseline hours, revision loops, elapsed time and cost before the change, measure the same afterwards, and subtract the full cost of the change including build and training time.
- What is a realistic time saving from AI in marketing?
- It varies by workflow and should always be measured locally. Repetitive adaptation, versioning and reporting typically show the largest gains; strategy and creative judgment show almost none, by design.
- Why do AI business cases fail in budget reviews?
- Usually because there is no baseline, the claim covers the whole department instead of one workflow, or the cost side only counts subscriptions and ignores build, training and maintenance.
Last updated: