Customer acquisition cost in an agentic world

What is your actual CAC now that a model writes the emails?

The old formula (paid spend divided by new customers) was always a simplification, but it worked because almost everything that drove a customer to your door cost real money in a predictable way. Now half your outreach is written by a model, sent by an agent, and qualified by another agent. The $300/mo in API spend that closed three deals last month doesn't show up anywhere in the SaaStr CAC formula. You're either understating CAC and feeling good or overstating it and feeling bad. Both are wrong.

This post gives you the new formula, the components that go into it, and a real sample calculation from our own running stack.

TL;DR

• The old CAC formula doesn't capture agent compute, agent build costs, or the amortization of agent infrastructure. In an agentic-AI-first business, those line items are usually 10-40% of the real CAC and almost always missing from the reported number.
• The new formula: CAC = (amortized agent build cost + agent compute in period + paid spend in period + human time in period × loaded cost) / closed-won customers in period
• LTV also changes. If fulfillment is mostly automated, gross margin per customer is structurally higher, which means a CAC that looked rich under the old model is often healthy under the new one.
• The healthy CAC-to-LTV ratio benchmark (3:1 in classic SaaS) loosens to 5:1 or higher for an agentic-AI-first business because the marginal cost of serving each customer is lower.
• Run both numbers monthly. Old formula for benchmarking against peers. New formula for actually running your business.

Why the old formula misses

In 2018, CAC was paid ads plus salaries of the sales and marketing team, divided by the new customers they brought in. The denominator was clean (a customer is a customer). The numerator was mostly labor cost plus paid spend. That worked.

In 2026, the numerator has three new line items that nobody includes by default:

• Agent build cost. You spent some time and money building the Outreach Closer, the Lead Sourcer, the Inbox Triager. That cost has to be amortized into CAC over its useful life. If you spent $20,000 building an agent fleet that you'll use for 24 months, that's $833/month of CAC that doesn't show up in any standard formula.
• Agent compute. Every API call, every model token, every embedding lookup, every webhook trigger your agents fire to acquire customers is a real cost. For most teams running serious outbound, this is $200-$2,000/month and is almost never categorized as CAC. It's hiding in "infrastructure" or "tools."
• Human review time on agent outputs. Even in a mostly-automated stack, a real human still spends some hours per week reviewing the agent's work. That human's loaded cost has to be allocated to CAC just like a salesperson's cost would have been.

Add these three and a CAC that looked like $40 turns into $110. That's the real number. Your actual unit economics depend on the real number.

The new CAC formula

Here it is, plain English first, formula after.

In plain English: the cost of acquiring a customer in an agentic-AI-first business is the per-period spending that drove customers to close, plus the amortized cost of the system you built to do that spending. The system has four components: the build cost, the ongoing compute, the paid amplification, and the residual human time.

In formula form:

CAC = (amortized_build + compute_in_period + paid_spend_in_period + human_hours_in_period × loaded_rate)
          / closed_won_customers_in_period
    

Where:

• amortized_build = total cost to build the acquisition agents, divided by their expected useful months, scaled to the period
• compute_in_period = all API/model/embedding/storage costs attributable to acquisition activity in the period
• paid_spend_in_period = ad spend, sponsored placements, paid syndication, paid lists
• human_hours_in_period = real human review, escalation handling, judgment calls
• loaded_rate = fully loaded cost per hour of the human (salary + benefits + overhead / available hours)
• closed_won_customers_in_period = the actual customers who paid in the period

A real sample calculation

Here are real numbers from one month of our outbound lane at VentureIO. Real on the structure, lightly rounded.

Build cost (amortized). We spent roughly $14,000 of equivalent time and tooling building the Lead Sourcer + Outreach Closer + Inbox Triager + Distribution agents. Expected useful life: 24 months. Monthly amortization: $583.

Compute in period. Model API costs ($240), embedding for lead enrichment ($85), inbox-watching webhooks ($30), Distribution syndication API costs ($45), state storage and verification logs ($25). Total compute: $425.

Paid spend. Zero this month. We don't run paid amplification on the outbound lane.

Human hours. I spent 6 hours this month on review, escalation, and approval. Loaded cost is about $180/hour for me. Total: $1,080.

Closed-won customers in period. 7 new paying customers, all attributed to the outbound lane.

Calculation. ($583 + $425 + $0 + $1,080) / 7 = $2,088 / 7 = $298 per customer

That's the real CAC. The naive calculation (paid spend / new customers) would have said $0 (since we ran no ads). The slightly-less-naive calculation (compute / new customers) would have said $61. Both materially understate the actual cost.

What the LTV side looks like

CAC doesn't matter on its own. It matters relative to lifetime value. The LTV side also changes in an agentic-AI-first business.

In a traditional SaaS, gross margin per customer is 70-85% after subtracting hosting, support, and a small allocation of fulfillment cost. LTV is roughly average customer revenue × gross margin × expected retention months.

In an agentic-AI-first business, gross margin per customer is structurally higher because fulfillment is mostly automated. Support is handled by an agent. Onboarding is handled by an agent. The marginal customer adds compute cost in pennies, not headcount in thousands. Gross margin is often 85-95%.

Sample LTV calculation. Average customer pays $4,200 over their lifetime (a mix of $1,997 blueprints and $250/month subscriptions for 8 months on average). Gross margin 90%. LTV is $4,200 × 0.90 = $3,780.

CAC of $298 against LTV of $3,780 is a ratio of 12.7:1. That's higher than the classic 3:1 benchmark by a large margin, and the reason is that the marginal cost of serving each customer in an agentic-AI-first business is so low that the LTV side carries the ratio.

Want to model this for your own business? Our blueprint catalog includes a CAC and unit economics blueprint that ships a fully populated spreadsheet against your actual numbers. Single email, single payment. Or email christine@operatoriq.io with last month's outbound spend and we'll do the math. Email only, no calls.

The benchmarks to use

Don't compare your new CAC number to peers' old CAC numbers. They're not the same animal. Use the right comparisons.

• For benchmarking against external peers (investors, public companies, market reports), report your CAC using the classic formula (paid + sales/marketing salaries / new customers). It's the apples-to-apples number.
• For running your own business, use the new formula. It tells you what you actually spent to land a customer.
• For CAC-to-LTV ratio, use the new formula on both sides. Classic 3:1 is a floor; 5:1 is healthy; 10:1 means you should be spending more to acquire because the unit economics are way above hurdle.
• For payback period, use the new formula divided by monthly gross profit. Payback under 12 months is the SaaS benchmark; in an agentic-AI-first business, sub-6-month payback is normal because the LTV side is so generous.

How to instrument it

Three pieces have to be tracked. None of them are exotic.

1. Tag every agent run with the lane. When the Outreach Closer runs, the log row includes lane = "acquisition". When the Support Agent runs, the lane is "fulfillment". This lets you partition compute costs cleanly. Without the tag, you can't separate acquisition compute from operational compute and your numerator is wrong.

2. Categorize every API invoice line item. Once a month, walk through your model API bill (Anthropic, OpenAI, etc.) and your other compute bills (Vercel, AWS, Stripe). Use the lane tag from the logs to allocate. If you don't tag at the source, you're guessing at the end of the month and you'll always be off by a meaningful percentage.

3. Time-track the human hours. This is the line nobody wants to instrument because it feels like reverting to billable hours. Do it anyway. Even a rough "Toggl on while I review the Outreach Closer queue" suffices. You need the number to be in the same order of magnitude as truth.

The common mistake

The most common mistake we see in founders calculating this is double-counting agent build cost. They include the full build cost in the month they shipped the agent. Their CAC for that month looks horrible. They include nothing in subsequent months. CAC for those months looks great. Neither number is real.

Amortize. Pick a reasonable useful life for the agent (we use 24 months as a default; pick longer for stable infrastructure and shorter for experimental builds). Divide the build cost by the useful life. Carry that monthly amortization line into your CAC formula every month until the agent retires or gets a major rewrite.

If you rewrite the agent, treat it as a new build cost on a fresh amortization schedule. Don't keep amortizing the old version into perpetuity.

What this means for growth strategy

When CAC is structurally lower and LTV is structurally higher, the growth math changes. A few practical implications.

• You can outspend traditional competitors at the channel level. If your real CAC is $300 against an LTV of $3,800, you can win an auction against a SaaS competitor whose CAC ceiling is $1,200 against an LTV of $2,400. They tap out before you do.
• You can pursue lower-ACV segments. Customers a traditional SaaS can't afford to acquire become viable when your CAC is one-third of what theirs would be. The bottom of the market opens up.
• You can scale outbound aggressively without burning capital. Adding 10x the email volume in our model costs roughly 10x the compute, which is still a small number. Headcount scaling didn't allow this; agent scaling does.
• You should still spend the saved money. Lower CAC is not a reason to spend less; it's a reason to spend more on the channels and offers where the unit economics hold. The discipline is reinvesting margin, not pocketing it.

What's coming next

Tomorrow's post is on the CEO of an agentic-AI-first company: what the role actually looks like, what fills the calendar, and what to do with the time you didn't have before. Together with this post on CAC and the cornerstone definition of an agentic-AI-first business, it closes the loop on the operating-and-financial picture: org chart, dashboard, failure modes, pricing, CAC, and the executive role at the top.

Want this CAC model built against your real numbers? The blueprint catalog includes a unit-economics blueprint that ships a fully populated CAC and LTV model in days. Or email christine@operatoriq.io with last month's spend and we'll do the math together. Email only, no calls.