Customer Acquisition Cost (CAC) With AI: Revised Formula

Customer acquisition cost is the marketing metric finance actually reads, and AI has quietly broken the way most teams calculate it. The formula did not change - total acquisition spend divided by new customers acquired - but the spend side did. AI subscriptions sit in a software budget nobody allocates to campaigns, AI-assisted labor blurs what an hour of acquisition work costs, and AI-shaped buying journeys blur which channel acquired anyone. I have watched teams report a falling CAC while their true acquisition economics got worse, purely because the new costs never entered the division.

This post is the revised calculation I use: what goes in the numerator now, the allocation rules that stop AI spend from hiding, where the benchmarks sit, and the mistakes that make the number flattering and useless. It is the metric-level companion to my AI ROI method - that post prices a workflow, this one prices a customer.

The formula, and where AI spend hides from it

The base formula is unchanged:

CAC = total sales and marketing acquisition cost / new customers acquired

The problem is the first term. AI is no longer a rounding error inside it: Gartner’s 2026 CMO Spend Survey has CMOs allocating 15.3% of the marketing budget to AI. A cost line that size, left out of the numerator because it lives in a “tools” budget, moves the CAC result materially - in the flattering direction, which is why nobody rushes to fix it.

In my experience AI spend hides from CAC in three places. Subscriptions and API usage get classed as general software overhead rather than acquisition cost. The human hours spent prompting, reviewing and correcting AI output get logged as nothing at all, even though review time is routinely the biggest AI cost a workflow carries. And agency or freelancer line items shrink because AI replaced them - which teams happily count - while the internal costs that did the replacing go uncounted. Count the savings but not the spending and you have built a one-way valve for good news.

The revised numerator

Here is the cost side I use, per period:

Acquisition cost =
    paid media spend
  + acquisition team salaries (loaded), by time share
  + agency / freelance / content production for acquisition
  + AI tools: subscriptions + usage, by acquisition share
  + AI labor: setup, prompting and review hours, loaded, by
    acquisition share
  + acquisition-serving software (landing pages, SEO tools, etc.)

Two allocation rules do the real work.

Rule 1: split every AI tool by use, not by intention. A ChatGPT or Claude seat used for ad variants, SEO content and also customer-support macros is partly an acquisition cost and partly not. I allocate by rough share of use - a one-week log of what the seat actually did is enough - and I write the split down so next quarter uses the same one. Precision is not the point. Not silently allocating everything to zero is the point.

Rule 2: AI-assisted labor is still labor. If a marketer spends ten hours a week producing acquisition content with AI, those ten loaded hours are acquisition cost, exactly as they were before the model arrived. The seduction of AI content is that it feels free because the draft appears in seconds. The true cost of the content line - briefing, prompting, editing, verifying - is smaller than the pre-AI cost, but it is not zero, and CAC math that treats it as zero will recommend scaling a channel on fictional economics.

The denominator has one rule worth stating: new customers only. Automated campaign types are skilled at converting your existing buyers and booking them as wins - Meta’s Advantage+ does this whenever the exclusion list lets it - and every existing customer in the denominator flatters CAC twice, once by inflating the count and once because those conversions were the cheap ones.

Blended and by channel: run both, trust each for what it can do

Blended CAC (all acquisition cost over all new customers) is the honest headline number, and the one that survives the attribution decay I keep writing about in the measurement guides - when AI answers and dark channels do the persuading, per-channel credit gets murky, but total cost and total new customers stay knowable. Channel CAC is the decision layer, and worth computing even imprecisely, because the spread is enormous: FirstPageSage’s benchmarks across roughly 120 B2B campaigns put average B2B CAC at $942 for organic channels against $1,907 for paid - with the caveat, which they state themselves, that these are three-year averages from their own agency’s client base.

AI moved this split in a way worth naming: it cut the production cost of organic-shaped acquisition - content, SEO, the work that gets you cited by AI assistants - much more than it cut paid media prices, which auctions keep setting. That tilts the optimal mix for many teams toward owned content, but only if the organic CAC is computed with rule 2 above. Organic was never free, and AI made it cheaper, not costless.

Why the honest number matters more than it used to

The backdrop is a long squeeze: SimplicityDX’s analysis puts the rise in customer acquisition costs at 222% over the last decade, with brands on average losing $29 on each newly acquired customer’s first sale, up from $19 ten years earlier. First-purchase losses make the payback question sharp: a CAC understated by hidden AI costs overstates how quickly a customer turns profitable, and that error compounds into every growth decision downstream.

The ratio investors read is the same one I would run internally: a16z’s rule of thumb is that a lifetime value of three or more times CAC within five years signals efficient sales and marketing spend. The ratio is only as honest as its denominator. Feed it a CAC missing 15% of the real cost and the healthy-looking 3:1 might be a genuine 2.5:1 - a difference that changes what a rational team would spend next quarter.

A worked example (hypothetical)

Imagine a small B2B SaaS team acquiring 40 new customers in a quarter. Paid spend is $30,000. Two marketers spend roughly half their time on acquisition - call it $28,000 loaded. AI tools cost $900 for the quarter, and a usage log says about two-thirds of that use is acquisition work, so $600 enters. The same log shows around 120 hours of AI-assisted production and review time inside those salary hours, already counted. Freelance content that used to cost $6,000 a quarter now costs $1,500.

Old math - paid spend plus freelancers only - reports CAC at about $788 and celebrates the freelancer savings. The revised numerator lands near $60,100, or roughly $1,503 per customer. Both numbers are “correct” arithmetic. Only the second one tells the team what a customer costs to win, and only the second one will predict what happens to profit if they double acquisition volume. All figures in this scenario are invented - the shape, with true CAC roughly double the convenient version, is what I regularly see when the hidden lines get counted.

The mistakes that flatter the number

Leaving AI in overhead. The default accounting treatment, and the biggest distortion now that AI is a double-digit share of budgets. If it produces acquisition work, it belongs in acquisition cost.

Counting saved hours as negative cost. Hours AI saved are not a deduction from CAC. They only lower CAC if the acquisition cost base actually fell - a freelancer not hired, a role not backfilled - or if the same team acquired more customers. The redeployment rule applies here unchanged.

Retention traffic in the denominator. New customers only, verified against the customer list, not against what the ad platform labels a conversion.

Chasing per-channel precision the data no longer supports. When attribution cannot see the AI touchpoints, decorating channel CACs with false precision misleads. I report blended CAC with confidence, channel CAC with stated caveats, and resist averaging the two moods.

Calculating it once. AI costs are moving - subscriptions get renegotiated, usage grows, review time falls as prompts improve. A CAC computed in January describes January. Quarterly recalculation with the same allocation rules is what makes the trend, which is the actually useful output, visible.

FAQ

What is a good CAC? There is no cross-company benchmark worth aiming at - the FirstPageSage figures above vary by an order of magnitude across industries and channels. The number that matters is yours against your own margin and lifetime value: the 3:1 LTV-to-CAC rule of thumb is the most portable yardstick, applied to an honestly loaded CAC.

Do salaries really belong in CAC? For the share of time spent on acquisition, yes - and AI makes this more important, not less, because AI-assisted production shifts cost from visible invoices (agencies, freelancers) into invisible internal hours. A CAC without salaries measures your media buying, not your acquisition economics.

Does AI lower CAC? It lowers the production cost of acquisition work, which lowers CAC only if the work translates into customers at the same or better rate. In practice I see real but modest net reductions once review time and subscriptions are counted - and occasionally increases, where cheap production tempted a team into channels that never paid back. The formula exists to tell those cases apart.

Blended or paid CAC - which should I report? Both, labeled. Blended is the health metric and survives attribution loss. Paid CAC (paid spend over customers attributable to paid) is the dial you can actually turn week to week. Reporting only the flattering one, whichever it is, is how CAC becomes advocacy instead of measurement.

This post belongs to the analytics and ROI hub. The workflow-level costing underneath it is the AI marketing ROI method, the program-level leak detection is the hidden-costs audit, and the attribution decay that pushes toward blended math is in zero-click attribution. The wider practice starts at the complete guide to AI marketing.