Klaviyo AI: What Composer and Customer Agent Cost, and Which One You Cannot Cap

Most vendors in this category publish AI features and withhold AI terms. Klaviyo publishes both, and that is the main reason it takes 4.2 on our AI axis, the highest in this category, rather than anything about the models themselves.

This page covers what the AI layer actually does, what each piece costs, which ones you can cap, and the one thing the vendor still does not publish.

What K:AI actually refers to

K:AI is the umbrella rather than a product you buy. Underneath it sit two things you pay for separately and several that are included in the plan.

The two priced products are Composer, described by the vendor as an agent you can tell "what you're trying to achieve, and it'll build the audience, draft the content, and map out the send strategy" ^[14], and Customer Agent, which handles shopper support conversations end to end.

The included pieces are the optimisations: send-time selection, channel affinity, and anomaly-style monitoring. The vendor's framing of channel affinity is worth quoting because it describes an ability the single-profile architecture makes possible and most rivals cannot attempt: "Email, text, push, or WhatsApp—K:AI automatically delivers each message on the channel where it's most likely to perform" ^[14].

That distinction matters commercially. The optimisations come with the plan. The agents are line items.

## What Composer costs, precisely

Twenty dollars a month. Five dollars of usage included even on the free tier. Additional usage billed at ten dollars per thousand credits through Flex ^[2].

The unit is the problem, and it is worth being honest about. A credit is not a thing you can count in advance the way you count contacts or sends. Composer bills on AI credits consumed within a billing cycle, which means your first forecast is a guess and your second is a measurement.

The vendor does give you the instrument: the account usage page shows Composer credits consumed over the last three, six or twelve cycles ^[5]. Treat the first two cycles as calibration and set the plan from what you observe rather than from what you hope.

Composer can be left to its allowance, which makes it the safer of the two agents to switch on while you learn what it costs.

What Customer Agent costs, and why it is different

Fifty dollars a month for seventy-five resolved shopper conversations ^[5].

The vendor's description is unusually concrete for an AI pitch: "When a customer asks where their order is, Customer Agent can find the order, check the status, and send an update automatically" ^[14]. It also sells the commercial angle directly, recommending "the right products based on each shopper's behavior and purchase history" so a service interaction becomes a sale, and writing preferences and conversation history back to the profile ^[14].

The denominator here is further outside your control than Composer's. Conversations resolved is a function of how many shoppers write in and how often the agent closes the thread rather than escalating. A product launch, a shipping delay or a sale all move it, and none of them is a decision you make about Klaviyo.

That matters more than usual because of the next section.

## The capping asymmetry, which is the thing to know

Email can be capped. The None option stops sending at your plan limit instead of upgrading you, and the vendor documents it plainly ^[11].

Customer Agent cannot. "You must be enrolled in either Flexible Overages or Auto-Upgrade in order to continue using Helpdesk and Customer Agent" ^[5], and on purchase you are "automatically enrolled into Flexible Overages as your default billing preference" ^[5].

So there is no configuration in which Customer Agent runs and your spend has a ceiling. The published way to stop accruing is to turn the product off or cancel its plan.

For a product whose volume is driven by your customers rather than by you, that combination deserves a decision rather than a default. Model a bad month, such as a delayed shipment during a sale, before you enable it, not after.

What the included optimisations actually do

These get less attention than the agents because nobody bills you for them, which is exactly backwards: they are the part of the AI layer most accounts get value from.

Send-time selection decides when an individual profile is most likely to open rather than when your whole list is. Channel affinity decides where: "Email, text, push, or WhatsApp—K:AI automatically delivers each message on the channel where it's most likely to perform" ^[14]. That second one is only possible because the channels share a profile, and it is the clearest practical argument for the single-record architecture.

The vendor also applies AI to form optimisation and to campaign variant testing, both running against the same profile data.

The reason to care is forecast quality rather than novelty. These run on the plan you already have, they have no meter, and they improve the numbers your paid agents will later be judged against. Any evaluation of Composer that has not first let the free optimisations run for a month is measuring the wrong thing, because the baseline is still moving.

## Where the AI sits relative to your data

Worth understanding structurally, because it explains both the capability and the lock-in.

Klaviyo's agents run against the Klaviyo Data Platform rather than against your mailing list. That platform carries purchase history, browsing behaviour, predicted value, memberships, appointments and subscriptions on the profile, and resolves returning anonymous visitors to known profiles ^[12]. The vendor says it holds 7.3 billion consumer profiles across its customers and processes 2.5 billion events a day ^[12].

This is why Customer Agent can answer "where is my order" without an integration: the order is already on the profile it is reading. It is also why the same question is hard for a vendor whose AI sits on top of a list.

The corollary is the lock-in. The more of your customer record lives there, the more the agents are worth and the more an exit costs. That is not a criticism: it is the trade every data platform offers: but it belongs in the decision rather than being discovered during it.

The provenance question, which Klaviyo answers

This platform asks every vendor the same question about AI: whose model is it, and what are the terms. Most do not answer. Klaviyo answers partly, and further than anyone else here.

It publishes a named Claude connector, described as bringing "your real Klaviyo data into Claude and get work back: reports, audits, and campaign drafts", with examples including pointing Claude at a segment for subject lines and asking it to "audit your active flows, flag the underperformers, and hand back a prioritized, ready-to-share doc" ^[15].

It publishes a named ChatGPT connector, where you "Tag @Klaviyo and ask which campaigns performed best last week" and receive cards and tables in the conversation ^[16].

And it publishes an MCP server in its developer documentation: "The Klaviyo Model Context Protocol (MCP) server seamlessly integrates with Klaviyo's APIs, enabling AI clients to interact with your Klaviyo data" ^[17].

Three named routes, publicly documented. A buyer can see which model sits behind which door, and can choose to use their own assistant against their own data rather than the vendor's.

## What is still not published

The model behind K:AI itself. Composer and Customer Agent are described by what they do, and the underlying model is not named. That is the ordinary state of this category, but it is worth stating rather than assuming, because the connectors create an impression of disclosure that the first-party agents do not quite meet.

Also unpublished: any country or language restriction on the agents. Mobile messaging rates vary "the list (standard) price per message for a given channel, country, and number type" and Klaviyo serves eight regional pricing pages, so it is reasonable to expect the agents have geographic limits too. No page states them.

And the credit itself is undefined in public. Ten dollars per thousand credits is a rate without a unit anyone outside the product can estimate.

What this layer is genuinely good for

Two things, on the evidence.

The first is reporting and auditing through your own assistant. The Claude and ChatGPT connectors plus the MCP server turn a reporting chore into a question, and the examples the vendor publishes, weekly performance summaries, flow audits ranked by underperformance, are exactly the work that otherwise does not get done because nobody has an hour for it.

The second is support deflection on a specific class of question. Order status, returns, where is my parcel. Those are high-volume, low-judgement, and the answer already lives in the data Klaviyo holds. That is a reasonable thing to automate.

What it is not good for is replacing judgement about what to say to whom. Composer builds an audience and drafts content; the decision about whether that audience should hear from you this week is still yours, and it is the decision that actually determines whether the programme works.

How it compares with the rest of the category

The gap is in disclosure rather than in capability, and the comparison makes that clear.

Mailchimp scores 3.0 on this axis. It ships generative content tools, predicted attributes and send-time optimisation, and marks AI features as included or not included by tier in its comparison table. What it does not publish is a usage allowance or an overage rate for any of them, which means a buyer can see whether they have the feature but not what using it heavily will cost.

ActiveCampaign's AI is metered from zero included units on its lower tiers: a rate without an allowance, which is the opposite shape of risk from Mailchimp's allowance without a rate.

Klaviyo publishes both halves for Composer: twenty dollars, five included, ten per thousand beyond ^[2]. That is why it scores where it does. The underlying technology across all three is broadly comparable and none of them is doing something the others cannot; what differs is whether you can put a number in a budget before you commit.

For a buyer, the practical reading is that AI should be compared on terms rather than on demonstrations. Every vendor's demonstration works.

A practical way to adopt it

Four steps, in an order that keeps the cost knowable.

Start with the included optimisations, which cost nothing extra: send-time selection and channel affinity run against the plan you already have. Give them a month and look at whether the reporting moves. If it does not, no paid agent will fix that.

Second, connect your own assistant rather than buying one. If your team already uses Claude or ChatGPT, the published connectors and the MCP server let you ask questions of live Klaviyo data without adding a line to the invoice ^{[15] [16] [17]}. The flow-audit example the vendor publishes is genuinely the highest-value use here, because flow audits are the work that never gets scheduled.

Third, add Composer when you have identified a specific recurring production task it should do. Its allowance survives on the free tier, so you can watch the credit consumption before you commit to the plan.

Fourth, treat Customer Agent as a separate decision with its own business case, because it is the only piece here with no ceiling and the only one whose volume your customers set.

## How this scores, and why

Klaviyo takes 4.2 on the AI axis, against Mailchimp's 3.0. The axis measures real capabilities and their terms, what is included, what is metered, at what rate, on which tiers, rather than how impressive the features sound.

Klaviyo publishes a rate for both agents, an allowance for one of them that extends even to the free tier, usage history per cycle, and three named external model routes. What costs it the last eight tenths is the unnamed first-party model, the undefined credit, the absent geographic terms, and one of the two agents that cannot be capped at all.

That is a good score for a layer that is still, by the vendor's own framing, early. It is also a reminder of what the axis is for: an AI you cannot forecast is a risk, however good it is.