Calculating AI Marketing Returns

Why the inputs matter more than the formula

Every AI marketing ROI model multiplies the same few things: hours, an hourly cost, extra leads, a close rate and a profit figure. The formula is trivial. The work is in the inputs, and most published figures fail there, usually by pricing an hour at salary alone, counting every freed hour as money, and treating every new lead as extra.

This guide takes each input in turn. When you have them, the arithmetic lives in the AI Marketing ROI Calculator.

Step 1: Price an hour properly

An hour of a salaried person costs more than salary divided by hours in the year. Add the employer's on-costs, then divide by the hours actually available for work.

For the UK tax year 2026/27 (6 April 2026 to 5 April 2027), employer Class 1 National Insurance is 15% on earnings above the secondary threshold of £5,000 a year, according to HMRC's published rates and thresholds for employers. Workplace pension automatic enrolment requires an employer contribution of at least 3% of qualifying earnings, and the lower level of qualifying earnings for 2026/27 is £6,240, per The Pensions Regulator. Both are taken from published summaries of those pages; check GOV.UK and The Pensions Regulator before relying on them.

Illustrative example: a marketing executive on £35,000.

• • Employer NI: (£35,000 minus £5,000) × 15% = £4,500
• • Employer pension: to keep the sum simple we apply 3% to full salary, £1,050. On the statutory qualifying earnings basis it would be (£35,000 minus £6,240) × 3% = £862.80, so the simplification slightly overstates cost.
• • Total employment cost: £35,000 + £4,500 + £1,050 = £40,550

Now the hours. Assume 37.5 hours a week and 46 working weeks once holiday and bank holidays are removed (an assumption; use your own contract). That is 37.5 × 46 = 1,725 hours. Assume a further 20% goes on meetings, admin and interruptions, leaving 1,725 × 0.8 = 1,380 productive hours.

Loaded cost per productive hour: £40,550 ÷ 1,380 = £29.38.

The naive figure, £35,000 ÷ 2,080 hours, is £16.83. Using it would undervalue every hour saved by more than 40%. Leave out desk, equipment and management overhead unless you can defend a figure for them; the result is conservative, which is the right direction to err.

Step 2: Count hours that are freed and reused

Time saved is only worth money if something valuable fills it, or if a cost goes away (fewer freelancer hours, a hire you no longer make).

1. • Time the task today. Take a sample of real instances, say ten reports or twenty briefs, and record how long each took end to end, including the back and forth.
2. • Time the new process. Include the human review, correction and prompting, not just the machine's run time.
3. • Multiply by monthly volume. The difference per instance × instances per month gives gross hours freed.
4. • Apply a redeployment share. Ask honestly what fraction of those hours will go to work you would otherwise pay for or that earns revenue. If you cannot name the work, the share is low.

Illustrative example: monthly performance reporting takes 12 hours per client report today and 4 hours with an AI-assisted process that still includes a person checking every figure. With 6 reports a month, gross hours freed are (12 minus 4) × 6 = 48. The team plans to use the time on a backlog of landing page tests, but realistically only about 30 hours will go there. Use 30, not 48.

Step 3: Estimate leads that would not have happened anyway

This is the input that decides most cases and the one most often overstated. A lead counts only if it is incremental: it arrived because of the new work and would not have arrived without it.

Three ways to estimate it, from strongest to weakest:

• • A controlled test. Run the new approach on part of the audience, region or product line and hold the rest back. The difference is your estimate. Platform lift tests work on this principle.
• • A before and after comparison with a benchmark. Compare the months after launch with the months before, and adjust using something unaffected by the change, such as a region or channel you did not touch, to strip out seasonality.
• • A cautious share. If you have neither, take a small share of current qualified lead volume, state it as an assumption, and plan to replace it with measured data within a quarter.

Whatever the method, use qualified leads, defined the way your sales team defines them, and the close rate from your CRM over at least two quarters. A single strong month is not a rate.

Step 4: Replace one answer with three

A single ROI figure hides the uncertainty. Build three scenarios instead and decide on the low one.

Illustrative example, using the loaded rate from Step 1 and the hours from Step 2, a monthly cost of £2,500 and gross profit per new customer of £2,000.

Three-scenario estimate (illustrative figures)

Input or result	Low	Expected	High
Hours reused per month	20	30	48
Value of hours at £29.38	£587.60	£881.40	£1,410.24
Incremental qualified leads per month	2	4	8
Close rate	15%	20%	25%
New customers per month	0.3	0.8	2.0
Value of customers at £2,000	£600	£1,600	£4,000
Monthly value	£1,187.60	£2,481.40	£5,410.24
Net after £2,500 monthly cost	minus £1,312.40	minus £18.60	£2,910.24

This example is instructive precisely because the expected case only just fails to break even, and the decision swings on the lead estimate. The right response is not to push the inputs up until the answer turns positive. It is to run a small, cheap test that replaces the lead assumption with a measurement, then decide.

Step 5: Record the assumptions with the answer

Keep a short note next to the figure: the source of each input, the date, and who agreed it. When the real numbers arrive, compare them with the note. That comparison is how your next estimate gets better, and it is the basis of measuring the investment after launch, covered in How to Measure ROI of AI Marketing Automation.

If your case is a single workflow rather than a whole marketing function, AI Marketing Automation ROI shows the break-even volume calculation. If you need the full UK cost of hiring rather than an hourly rate, see AI Marketing Automation Cost UK.

FAQ

Should I include overheads like office space in the loaded hourly cost?

Only if you can defend the figure and the cost would actually change. Leaving overheads out keeps the rate conservative, so any return you find is less likely to be overstated.

What if the freed hours are spread thinly across several people?

Scattered half-hours rarely turn into valuable work. Count them at a low redeployment share, or not at all, unless the change lets you avoid a hire or cut paid freelance time.

Can I use the vendor's case study figures as my incremental lead estimate?

Treat them as a hypothesis, not an input. Their baseline, audience and offer differ from yours. Use your own test or a cautious share of your current volume instead.

Why decide on the low scenario rather than the expected one?

Because the low case tells you what you lose if the optimistic assumptions fail. If the low case is survivable and the expected case is attractive, the decision is sound.