“We use AI” is not a result. Measure AI projects the way you’d measure a new hire: record what the task cost before, track the hours saved and where they went, check quality against the old baseline, and count only the output that reaches customers. Then subtract the full cost, including the human time spent reviewing and fixing what the AI produces.
Why AI ROI is hard to see
Most marketing teams adopt AI the way they adopt a browser extension. Someone finds a tool, it spreads, and six months later the only evidence is a line on the software bill. Nobody wrote down how long the task took before, so there’s nothing to compare against.
Three things hide the return:
- The gains are spread thin. Saving 20 minutes on each of 40 small tasks never shows up as a line item. It disappears into the week.
- The costs are hidden. Seat licenses are visible. The hours spent writing prompts, checking output and fixing mistakes are not.
- Output is not outcome. Producing three times as many blog drafts or ad variants is only a return if those drafts publish, rank or win tests.
A new hire faces a clear test: what did they cost, what did they produce, and did the work reach customers? Hold every AI project to the same test. When I review a marketing team’s AI stack, the usual gap isn’t the tools. It’s that no workflow has an owner, a baseline or a number attached to it.
Baselining tasks before you automate
You can’t measure time saved without knowing the time spent. Before switching on any AI workflow, baseline the task for two to four weeks, or reconstruct it from tickets and project history.
For each task, record:
- Volume: units per week or month (briefs, ad variants, reports, enriched leads)
- Hands-on time per unit: actual working time, not elapsed calendar time
- Cycle time: request to delivery, including time waiting for review
- Cost per unit: hours times loaded hourly cost, plus any freelancer or agency fees
- Error or rework rate: share of units sent back, corrected after publishing or rejected
- Downstream result: what the output feeds, such as conversion rate, reply rate or publish rate
Loaded hourly cost is salary plus benefits, payroll taxes and overhead, divided by working hours. Use one blended figure per role and keep it fixed so the math stays comparable month to month.
Pick tasks that are frequent, repetitive and measurable. Weekly reporting, ad copy variants, lead enrichment and content briefs are good candidates. “Strategy” is not. If you’re still choosing where to start, AI agents for marketing: 12 use cases lists the ones where the time savings tend to be real.
Time saved vs capacity redeployed
This is where most AI ROI claims fall apart. Time saved is not money saved. It becomes a return in only three ways:
- Cost removed. You cut contractor hours, reduce an agency scope or skip a planned hire. This is the cleanest return and the easiest to prove.
- Capacity redeployed. The freed hours go to work with a measurable result: more tests launched, more accounts worked, the content refresh program that was always “next quarter.”
- Speed that changes an outcome. A faster cycle matters when timing drives results, such as responding to inbound leads in minutes instead of hours, or shipping a campaign before a seasonal window closes.
If none of those happened, the hours were absorbed. That may help a stretched team, but it isn’t ROI.
A hypothetical example with round numbers: an AI workflow cuts weekly performance reporting from 6 hours of hands-on time to 1.5 hours, including review. That frees 4.5 hours a week, about 18 hours a month. At a loaded cost of $75 an hour, that’s $1,350 a month of capacity. It counts as a return only if you can name where those 18 hours went, for example two extra landing page tests a month, logged in your experiment backlog.
Ask the workflow owner to fill in one extra field for the first few months: “freed hours went to…” If it stays blank, the project is a convenience, not an investment.
Quality and error rates
Speed with worse quality is a false saving. Measure quality against the same baseline you set before automation, with the same reviewer where possible.
| Task | Quality measure | Warning sign |
|---|---|---|
| Ad copy and creative variants | Share of AI variants that beat or match control in tests | Volume up, win rate down |
| SEO content drafts | Edit time per draft; share published without a major rewrite | Drafts that take longer to fix than to write |
| Reporting and dashboards | Errors caught before and after sending | Wrong numbers reaching leadership |
| Lead enrichment and routing | Match rate; share of leads routed to the wrong owner | Silent failures on edge cases |
| Customer-facing email | Brand and factual errors per send; complaint and unsubscribe trends | Off-brand tone, invented claims |
Two rules keep this honest. First, count review and rework time against the project. If a draft takes 10 minutes to generate and 50 minutes to fix, the task takes an hour. Second, keep sampling output after it looks stable. AI workflows fail quietly when inputs change: a new product line, a renamed CRM field, a tracking change upstream. A monthly spot check of a fixed number of outputs catches drift before customers do.
Weight errors by where they land. A typo in an internal summary is cheap. A wrong price in a campaign email or a misrouted enterprise lead is expensive.
Revenue-linked outcomes
The strongest AI projects connect to a number the business already tracks. Not every project will; know which kind you’re running:
- Revenue-linked: the output directly touches pipeline or sales. Faster lead routing, more tested ad creative, better product copy on pages that convert. Measure against the business metric: speed-to-lead and meeting rate, creative win rate and cost per acquisition, product page conversion.
- Capacity-linked: the output frees time for revenue work. Reporting, research summaries, call notes. Measure hours freed and where they went.
- Risk-linked: the output prevents mistakes. Data validation, brand checks, alerts for broken tracking. Measure errors caught and what they would have cost.
For revenue-linked projects, isolate the effect where you can. Run the AI workflow on part of the volume and keep the old process on the rest, or roll it out by segment, region or product line and compare. A simple before-and-after works only if nothing else changed, and in marketing something else almost always did, so treat it as directional.
Count only output that reaches customers. Fifty AI-generated ad variants of which eight launched is an eight-variant project. Twenty drafts of which twelve published is a twelve-post project.
Full costs: tools, setup, maintenance and review time
Most of the cost of an AI project sits outside the software bill. List all of it:
- Tools: seat licenses, API usage, automation platform fees, enrichment or data services
- Setup: designing the workflow, writing prompts, connecting systems and testing, whether internal hours or a one-off build fee
- Maintenance: fixing broken integrations, updating prompts when products or messaging change, retesting when the underlying model changes. Start with an assumption of a few hours a month per workflow, then replace it with logged time
- Review time: every minute a human spends checking, editing or approving output
- Adoption: training and the time it takes to get the whole team using the workflow the same way
Amortize setup over a fixed period, such as 12 months, so a project isn’t judged a failure in month one or a runaway success because setup was ignored.
Back to the reporting example. Review time is already netted out of the 18 freed hours, so it doesn’t appear twice; just make sure it isn’t left out. Monthly costs: $200 in tools and API usage, $300 of maintenance (4 hours at $75) and $500 of amortized setup (a $6,000 build over 12 months). That’s $1,000 against $1,350 of freed capacity, a thin margin that becomes a real return only if those hours produced something.
Building versus buying changes this cost mix; build vs buy AI tools covers that decision.
A scorecard for AI projects
Review every AI project monthly on one page: one block per workflow, the same rows every time. Here is the reporting example filled in.
| Metric | Baseline | This month | Notes |
|---|---|---|---|
| Volume | 4 reports | 4 reports | Same scope |
| Hands-on hours | 24 | 6 | Includes review |
| Hours freed | — | 18 | Valued at $1,350 |
| Freed hours went to | — | 2 extra landing page tests | From experiment log |
| Errors | 1 correction | 0 | 10-output spot check |
| Business outcome | — | 1 test winner shipped | Tied to conversion rate |
| Full monthly cost | — | $1,000 | Tools, maintenance, amortized setup |
| Net result | — | +$350 capacity, plus test output | Decision: keep |
Every workflow’s block ends the month in one of four decisions:
- Keep and expand: net value positive, quality at or above baseline, freed time has a named destination
- Keep and fix: time saved but quality slipping or review time creeping up. Tighten inputs, prompts or guardrails before scaling
- Reassign: time saved but nowhere useful to put it. Either remove the cost it replaced or point the capacity at a specific backlog
- Shut down: full cost above value for three consecutive months, or errors reaching customers
Report the whole portfolio, not just the winners. A quarterly summary for leadership should list every AI project, its net value and its decision, including the ones you switched off.
This is how I run AI automation work: every agent or n8n workflow ships with a baseline, an owner and a scorecard row, so keeping it is a decision based on numbers rather than enthusiasm.
Get it built
If your team pays for AI tools and can’t say what they return, I can baseline your workflows, build the ones worth automating and set up the scorecard. The AI automation add-on starts at $2,500 a month, or start with a Growth Audit, $1,500 fixed and credited if we continue. See pricing or get in touch.