AI marketing ROI: measure the cost of work you can actually use
Quick answer
AI marketing ROI tells you whether the work you get back is worth what you spend. Count the time spent preparing, checking and fixing it, as well as the software bill. Producing drafts faster means little if your team spends the afternoon putting them right.
The saving often disappears after the first draft
A marketer can produce ten campaign briefs in the time previously needed for one. That sounds like a productivity gain until an editor has to repair ten sets of unsupported claims. The useful unit of measurement is an accepted brief: something another person can safely act on without repeating the work.
This matters when teams move from individual AI subscriptions to connected production. A pilot can borrow an experienced employee’s judgement without recording the time it consumes. When the same process expands, that hidden dependency becomes a queue. Before approving more seats, ask where the work waits and who has to rescue it.
Gartner’s 2026 CMO Spend Survey found that 70% of respondents considered their internal marketing processes insufficiently mature to implement and scale AI effectively. The survey covered 401 marketing leaders, predominantly at businesses exceeding $1 billion in annual revenue. It is evidence of organisational friction in that sample, not a forecast for every small team.
Build the comparison around one job
Choose a repeatable deliverable and write its acceptance rules before timing it. For an email campaign, these might include accurate product details, approved claims, correct links, appropriate audience exclusions and a usable handoff to production. Keep the audience and difficulty comparable between the old and new process.
Record elapsed time and hands-on time separately. A workflow may reduce labour while still taking two days because a reviewer is unavailable. Both measures matter, but they answer different questions. Also record failures rather than silently deleting abandoned attempts from the denominator.
Our guide to running a marketing AI pilot explains how to preserve difficult cases for evaluation. Use the same approach here: the comparison should include the awkward briefs, not just the cleanest examples.
| Cost or outcome | Record | Why it matters |
|---|---|---|
| Preparation | Source collection and prompt setup | Often hidden in demonstrations |
| Review | Checking and correction minutes | Can erase generation savings |
| Accepted output | Deliverables that pass the quality gate | Provides a meaningful denominator |
| Operating cost | Seats, usage, connectors and maintenance | Makes the baseline comparable |
What does that save you each month?
If your team produces 40 finished briefs a month, each takes 90 minutes and labour costs £40 an hour including employment overheads, you’re spending £2,400 a month on the work.
Now bring AI into the process. Say preparation takes 15 minutes, generating and handling the draft takes five, and checking and fixing it takes another 35. That’s 55 minutes per finished brief, or roughly £1,467 in labour each month. Add £300 for software and £200 a month to cover setup and maintenance, and the total comes to about £1,967. You’ve saved £433, around 18%.
Useful, certainly. But watch what happens if checking takes 55 minutes instead of 35: the monthly bill rises to about £2,500. You’re now spending more than you were before. The number that changes the decision is review time, not how many drafts the tool can produce.
Separate capacity, cash and revenue
Hours saved create capacity. They become cash savings only if spending actually falls. They become growth only if the released capacity produces additional useful work and that work affects the business. Report each step separately.
For revenue claims, compare business outcomes through a suitable experiment or a clearly qualified observational analysis. A rise in pipeline after an AI rollout may also reflect a new offer, stronger demand or more media spend. Do not credit the tool for every change that happened after it arrived.
Subscription fees can also conceal usage charges. Compare seat costs with API and connector costs, and account for minimum commitments. The ChatGPT budgeting guide and Claude buying guide provide the product-level questions to ask; verify current contract terms before purchasing.
Our take: pay for work you can use
Before buying more seats, we’d want three things: work that meets the team’s quality standard, a lower cost per finished job, and someone who can sort things out when they go wrong. A faster first draft isn’t enough.
Keep a monthly ledger of accepted outputs, rejected outputs, review minutes, total operating cost and what the released capacity was used for. Recheck it after a model, prompt or source-system change. If the savings rely on one person quietly correcting everything, the process has not yet earned a larger budget.
The practical next step is small: time five ordinary jobs and five difficult ones from beginning to acceptance. That first ledger will tell you more about the investment than a vendor’s best demonstration.
Related reading
AI agent governance for marketing: permissions before autonomy · Marketing budget allocation: fund the next useful decision · Martech stack audit: decide what to keep, fix or retire