.svg){kind=icon}
.svg){kind=icon}
Key takeaways
- Per-token prices are falling, but agentic marketing is pushing total AI consumption, and the bill, higher.
- Agentic AI can hit cost-per-acquisition targets efficiently. On its own, it can’t tell whether those customers stay.
- More automation in growth marketing calls for stronger guardrails, because thousands of decisions made in milliseconds can’t be reviewed one by one.
- The Investment Framework scores every channel, product and market combination on lifetime value per incremental dollar, then funds the highest scorers first.
- In agentic marketing, the framework sets the goal, defines limits that can’t be crossed, and checks results later with holdout tests and cohort tracking.
- Email, push and in-app messages are no longer close to free. When agents personalize every send, each one carries a compute cost.
- AI costs belong in three buckets (acquisition, retention and team overhead), and each is judged differently
AI agents were supposed to make growth cheaper, but for many companies, they are making it more expensive in ways that don’t show up until the bill arrives. This has made the cost of running agents at scale, meaning every model call, every generated message and every decision in real time, a boardroom topic.
In June 2026, TechCrunch reported about falling per-token prices, but agentic use has pushed consumption so high that some companies were three times over their full-year token budget by April.
At the same time, the cost of acquiring a customer keeps climbing. A study by Adobe, Publicis Sapient and Incisiv,found that in travel and hospitality, acquisition costs rose about 35% from 2022 to 2025 while customer lifetime value grew just 4.5%. The obvious response is to point agents at acquisition and let them drive that cost down. Agents are very good at this. They can hit a cost-per-signup target with remarkable efficiency. What they can’t tell you on their own is whether those customers stay.
An agent can grow revenue quarter after quarter while quietly filling the business with customers who leave in 60 days. So the real question for marketers isn’t whether to use agents. It’s what to tell them to optimize for.
More automation calls for better guardrails
The common assumption is that as agents take over execution, measurement matters less, because the algorithm optimizes itself. In reality, the opposite is true. You can sit beside a human media buyer, read their bid sheets and question their reasoning. While you cannot do that for thousands of automated decisions made in milliseconds, building a strong governance framework helps in placing guardrails.
The lapse happens when enterprises adopt Agentic AI without foundational systems. In a recent EY survey of 202 senior AI executives at large US companies, nearly half said their governance framework had not yet been updated for agents, and 72% worried about being unable to trace or audit the data behind critical AI decisions.
The same marketers are putting more money into proof. In a July 2025 EMARKETER and TransUnion survey of US brand and agency marketers, 46.9% planned to increase investment in marketing mix modeling, and 36.2% wanted to invest more in incrementality testing.
These are not opposing trends. The faster the engine runs, the better the brakes and steering have to be.
Why agentic AI needs a framework
In such a situation, clarity on fundamentals is critical. While working with enterprises in the agentic era, our work over the past decade has given us a strong foundational understanding of where to chase growth.
For about a decade we at LatentView worked for a global consumer technology platform enabling them to achieve 10x growth, to more than 125 million paid subscribers, including trials, by March 2025. Our team worked alongside Shikha Agarwal, who led subscription growth for the platform, to design the Investment Framework and build the measurement floor underneath.
To begin with, the problem was never a lack of data. Every channel reported its own cost per acquisition. No one could answer the basic economic question: was a customer won through one channel worth more or less than a customer won through another at the same price?
A low-cost subscriber who left after 60 days and a premium subscriber who stayed three years were judged against the same efficiency target. The metrics had no way to tell them apart. Shikha, who conceptualised the Investment Framework has written about it extensively in her book ‘Scaling Subscriptions’, explaining what bridges this gap.
How the framework works
Most marketing teams judge a channel by one number: what it costs to win a customer. The Investment Framework adds a second question that has to be answered at the same time: what is that customer worth over the whole time they stay?
It puts two numbers side by side. The first is what a customer is worth: how many months they keep paying, multiplied by what the business earns each month. The second is what it truly cost to win them, counting only customers who would not have arrived anyway. Dividing the first by the second gives one score: lifetime value per dollar spent.
Every combination of channel, product and market gets that score. The budget goes to the highest-scoring combination first and keeps flowing there until extra spend stops paying back.
Then it moves to the next one. A small, fixed slice is set aside for deliberate strategic bets, so judgment calls are visible and capped rather than hidden in the numbers.
To illustrate how this works, think of two channels each costing $50 to win a subscriber. Customers from the first stay four months on average, and customers from the second stay 18. A cost-per-acquisition dashboard calls that a tie. The framework shows the second channel is worth more than four times as much, and moves money accordingly. (Illustrative figures, not client data.)
The first step of this framework begins with measuring what a customer truly costs, using lift studies, or media mix modeling where controlled tests aren’t possible. Behind that line sit test design, model building and calibration, retention curves that hold up when you slice them finely enough to act on, and margin data joined across systems never built to connect. All of it has to meet a standard finance will sign. That first step took our team the better part of 18 months, and it is why most companies admire this kind of framework without ever running one.
That has always been the shape of our collaboration. Shikha’s framework describes what a good decision looks like. The measurement floor decides whether you are capable of making one.
What happens when budgets follow lifetime value
With more data, the original subscription platform also keeps improving. The earlier lifetime-value model projected retention from just the first two months of data. The new version added a third month and discounted future margins for the time value of money, then fed directly into the investment workflow that evaluates every offer. It contributed to 210,000 paid starts in a single quarter of 2025, worth an estimated $6.4M in lifetime value, and now underpins offer planning for a $29M budget in 2026.
Over the years, the principles behind the framework work for growth problems across sectors. It helps judge spend by lifetime value, count only the results marketing actually caused, and prove it before scaling. And we have seen those principles pay off well beyond subscriptions.
At a global creative-software company, product budgets were set on recurring revenue alone. That ignored cancellations, plan switches and resubscriptions, so products that kept customers longer were underfunded. We built a lifetime-value model that followed customers past their first subscription, counting those who came back after cancelling and those who added a second product.
Budgets were then reallocated on the new numbers.
The lesson is not that spending more works, but goes deeper to show where more spending would pay back, and give it the confidence to commit.
What to tell an AI agent
Here is what happens when an AI system is given the wrong goal. Our team worked with a financial software company set out to cut 3 million customer contacts, with 2 million to be handled by AI and self-help tools.
On its own terms, that was a sensible cost target. The analysis found a problem. Across 4.8 million contacts, nearly half came from the customers most likely to buy more. They were mostly new customers in their first 30 days with no add-on product yet. Those conversations carried an estimated $52.7M in revenue. Deflecting them all to AI would have saved support cost while quietly giving that revenue up.
The fix was a guardrail: keep high-value customers out of AI containment until the model has earned trust. That decision is credited with avoiding about $4.3M in lost revenue. The issue arose when AI was pointed at the wrong number. In my view, the Investment Framework prevents that in three ways:
- Sets the goal – Without a framework, an agent optimizes for whatever it can see: clicks, signups, deflected tickets, conversions an ad platform can claim credit for. The framework gives it the business’s real objective, which is long-term value per dollar, not first-day results.
- Identifies the limits – Agents need lines they cannot cross, however good a shortcut looks. Examples include a minimum value-per-dollar score below which no more money is spent, a cap on discounting to win or keep a customer, and a ceiling on the strategic-bet budget so experiments can’t quietly become the strategy.
- Checks results afterward – Agents act in seconds, but customer value shows up over months. The framework requires going back to test whether the customers an agent won actually stayed and paid, using holdout tests and cohort tracking rather than the platform’s own reporting.
While agents are excellent at growing revenue, the framework is what tells you whether that growth is profitable.
Why agents end era of “free channels”
In the race to deploy AI, enterprises often do not account for the cost of agents running emails, push notifications and in-app messages, all of which have for long been treated as nearly free.
Traditionally, once a system is built, one more message costs almost nothing, so few teams hold these channels to the same return standards as paid media.
Agents change that. When an AI writes each message, picks the moment to send it and decides which customer gets which offer, every send carries a compute cost. Personalizing one weekly message for every customer in a base of several million means paying for millions of AI decisions every week.
That cost grows with the audience and the number of messages, not as a one-time software license.
The implication is that owned channels now have to earn their budget the same way paid media does, on the same value-per-dollar score. Deeper personalization is no longer free polish. It is an investment with its own point of diminishing returns. I should be honest that this is a hypothesis my team is actively testing, not a settled finding.
Early evidence points the same way. A large software company replaced several simultaneous campaigns with one ranked recommendation per user and kept the same growth with fewer messages.
Another company shifted email volume toward its highest-propensity customers and lifted revenue per email by 43%. In both cases the gain came from measuring the return on each send.
Where AI cost actually lands
AI spending doesn’t sit on one tidy line, and putting it in the wrong place leads to the wrong decisions. In my experience, it belongs in three different places:
- Winning customers
If an agent helps acquire customers, count only the extra cost it adds beyond what the ad platform already charges. Compare that with the extra customers it actually brings in. Ad platforms already build their own automation into the media price, so counting it again inflates the cost. - Keeping customers
This is where the bill hides. Picture a retention agent that sends personalized offers and well-timed discounts. Customers stay longer, and the retention dashboard turns green. But each extra month now earns less, because of the discounts and the cost of running the agent. The business can keep more customers while making less from each of them, and nobody notices unless retention cost is counted against customer value. - Running the team
AI tools that help the growth team itself, such as analysis assistants and experiment-design tools, are overhead like any other team tool. Spreading that cost across channels makes every channel look worse and tells you nothing you can act on. Judge it on its own terms. A third-party optimizer adds one more risk. Ad-platform bidding systems are built to maximize the conversions they can see and claim credit for, not your customers’ value after churn and margin.
The more of the loop you hand to external tools, the more your budget is steered by goals that aren’t yours. The framework is where your real goal still lives.
A decade ago, a company could pause rigorous measurement for a quarter and reconcile the numbers later. With agents making thousands of spending decisions a day, that margin for error is gone. So the first question for any company moving to agentic marketing is not which tools to buy, it is to identify which tools drive actual growth.
Acknowledgment: The Investment Framework was designed by Shikha Agarwal and is set out in her book Scaling Subscriptions. My thanks to her for a decade of partnership on this work.
FAQs
1. Why are AI agents making marketing more expensive?
Per-token prices are falling, but agents make so many model calls, messages and real-time decisions that total usage is rising faster. Some companies were three times over their full-year token budget by April 2026.
2. What is the Investment Framework?
A budgeting method that sets what a customer is worth over their lifetime against what it truly cost to win them, counting only customers who wouldn’t have arrived anyway. Budget goes to the highest lifetime-value-per-dollar combinations first.
3. Why isn’t cost per acquisition enough to guide AI agents?
CPA treats a subscriber who leaves after 60 days the same as one who stays three years. An agent optimizing for CPA can grow revenue while filling the business with customers who churn.
4. What guardrails should AI marketing agents have?
A minimum value-per-dollar score below which spending stops, a cap on discounting, and a ceiling on strategic-bet budgets. Results should be checked afterward with holdout tests, not the platform’s own reporting.
5. Are owned channels like email still close to free?
No. When an agent writes, times and targets each message, every send carries a compute cost that grows with audience size and message volume. The author frames this as a hypothesis his team is still testing.
